Add files using upload-large-folder tool

MANIFEST.json

@@ -0,0 +1,271 @@
+{
+  "schemaVersion": "blobfish.huggingface-domainbench.v1",
+  "datasetId": "SamuelChien821/blobfish-domainbench-24",
+  "generatedFrom": "blobfish.benchmark-evidence.v1",
+  "benchmarks": 6,
+  "tasks": 24,
+  "trajectories": 12,
+  "assets": 42,
+  "files": [
+    {
+      "path": "README.md",
+      "bytes": 4835,
+      "sha256": "f04be73829c8041d062e565fe92bc9530cba79b1a5450999b47d29d766e1087e"
+    },
+    {
+      "path": "data/assets.jsonl",
+      "bytes": 16121,
+      "sha256": "01457ba12872d85e0cd5cfa3712c71f9103817bdc152313cc75e7ed6754a8f5a"
+    },
+    {
+      "path": "data/tasks.jsonl",
+      "bytes": 32893,
+      "sha256": "0b2e9018e66f33671f142f124969efb7d41e0a5ad8d78688b1c588f9bf552cf3"
+    },
+    {
+      "path": "data/trajectories.jsonl",
+      "bytes": 30209,
+      "sha256": "c5cc8c14549769e07c78e71884900b54e9419d3076b92079a997a415d67eb818"
+    },
+    {
+      "path": "evidence/defensebench-100.json",
+      "bytes": 19646,
+      "sha256": "1c9628cd2c2f1752c9d49365e9520c80c18c82a10fac1b425c6a746bca1a16cd"
+    },
+    {
+      "path": "evidence/discoverybench-102.json",
+      "bytes": 14871,
+      "sha256": "2e96508fc666e2a2c536ed3472b3c2f5324985621ebb951e6d0859c6402555b6"
+    },
+    {
+      "path": "evidence/filingbench-50.json",
+      "bytes": 15584,
+      "sha256": "49d9ce8a0708151d5dd02e8004e439fb8261a2215c222668e5a8a4c3761d9e58"
+    },
+    {
+      "path": "evidence/opsbench-14.json",
+      "bytes": 15005,
+      "sha256": "10939d5c973ec1ba685f748f62651d3a00f0d12ffd08d1c4510996815d481793"
+    },
+    {
+      "path": "evidence/repobench-500.json",
+      "bytes": 15155,
+      "sha256": "8926d0f4e8f8cb48656fa98fb4187ee5441b59bd92b8f11411e31622333eaae6"
+    },
+    {
+      "path": "evidence/servicebench-375.json",
+      "bytes": 14728,
+      "sha256": "e48bacdce6cce781eab4b1239b78f6ed599c8f7cf342576fe1ca622c17d6709a"
+    },
+    {
+      "path": "worlds/defensebench-100/environment.db",
+      "bytes": 32768,
+      "sha256": "4940300240cbe15d35d1fd45016398abe150ca2ee0361d04b2a17432f8a26233"
+    },
+    {
+      "path": "worlds/defensebench-100/mcp/manifest.json",
+      "bytes": 1969,
+      "sha256": "3c809b60c7ab96026e4ee0f4ac5ba0d04b011005465adff97185e0f98d1893ac"
+    },
+    {
+      "path": "worlds/defensebench-100/schema.sql",
+      "bytes": 1251,
+      "sha256": "08c2e8abf194ee2764b390f01b9e8185c8653950c9582b3e2210e0f5f4a7b70b"
+    },
+    {
+      "path": "worlds/defensebench-100/seed.sql",
+      "bytes": 14152,
+      "sha256": "9c2b62927c80c430b0808605fbdefa36c200ef146cd539f6d02d6b1acb66573c"
+    },
+    {
+      "path": "worlds/defensebench-100/tasks.json",
+      "bytes": 106978,
+      "sha256": "44ae2da2ce9ef5faccb06335c0a242eabcdd8e41626925131bb01665d1165aa5"
+    },
+    {
+      "path": "worlds/defensebench-100/tools.json",
+      "bytes": 90244,
+      "sha256": "46021a577b0d1e7e3d8c5befed8ac5f6ea9d3d84ce61fc8e34a446fe59b35c43"
+    },
+    {
+      "path": "worlds/defensebench-100/world.json",
+      "bytes": 2540,
+      "sha256": "c8a86f80f071188223d082706306fb84984c1fbb6cce9341f4ab0f8e4a48ef8a"
+    },
+    {
+      "path": "worlds/discoverybench-102/environment.db",
+      "bytes": 32768,
+      "sha256": "44b907f65127934c49e2569a0226486c470231e635d324cec1aa09ebf4a6df05"
+    },
+    {
+      "path": "worlds/discoverybench-102/mcp/manifest.json",
+      "bytes": 1862,
+      "sha256": "6dfe7eff3b010b684dadaac82edc606243c2582461bb691c512da2f3d7c074a4"
+    },
+    {
+      "path": "worlds/discoverybench-102/schema.sql",
+      "bytes": 1497,
+      "sha256": "4dfc04a1e3d89b3839449644c07633a4bc2b86d65bc8d6d459ac0c3bb4989c83"
+    },
+    {
+      "path": "worlds/discoverybench-102/seed.sql",
+      "bytes": 11964,
+      "sha256": "e3e3c20b66558ba03fff8da8292f92b28b031e1f2dc7a2310e5e702cff3398bc"
+    },
+    {
+      "path": "worlds/discoverybench-102/tasks.json",
+      "bytes": 103570,
+      "sha256": "73b32ea4a6834a886e55ae6cf26412bec6eadff9e06959912dbf905ab98881aa"
+    },
+    {
+      "path": "worlds/discoverybench-102/tools.json",
+      "bytes": 97373,
+      "sha256": "511ab9dc9a947171a0a70b0d75d1201d40298c285f217b904ba69ef631ae5e5a"
+    },
+    {
+      "path": "worlds/discoverybench-102/world.json",
+      "bytes": 2492,
+      "sha256": "9788c5cbc28caa5c426e82ab59d220ad7d7aaccf913b4a77dc0e7a8e4f4c6f14"
+    },
+    {
+      "path": "worlds/filingbench-50/environment.db",
+      "bytes": 32768,
+      "sha256": "f70a5b84565d5aa98bbcb1aeff9c7a0742213f85c3ad876a46b0d59265875e51"
+    },
+    {
+      "path": "worlds/filingbench-50/mcp/manifest.json",
+      "bytes": 2039,
+      "sha256": "e92914c7b33fd3e5a50cf7c908177962734d3e7e775da258ca4e62f1437fd301"
+    },
+    {
+      "path": "worlds/filingbench-50/schema.sql",
+      "bytes": 1574,
+      "sha256": "510bf7e2943495cb8a329c4ccce54ffc7376226dedd2b9c11288d8fa78dc40dd"
+    },
+    {
+      "path": "worlds/filingbench-50/seed.sql",
+      "bytes": 14116,
+      "sha256": "c888acecc74c67054a759370315e8573f3315579c287ac540b21b57d37a3dc44"
+    },
+    {
+      "path": "worlds/filingbench-50/tasks.json",
+      "bytes": 106953,
+      "sha256": "96388b0f39e51ed14f718001371783a506e5654586235e0ae1efd4decfbaf80b"
+    },
+    {
+      "path": "worlds/filingbench-50/tools.json",
+      "bytes": 104507,
+      "sha256": "6653f9d2a6e133047d7bb9f30b8ef81a74debb4efc94b94785664894036d4414"
+    },
+    {
+      "path": "worlds/filingbench-50/world.json",
+      "bytes": 2527,
+      "sha256": "8475c86667f578b550dcc0bf5c449981036ea1a8c15dc9613906bb30165f972e"
+    },
+    {
+      "path": "worlds/opsbench-14/environment.db",
+      "bytes": 32768,
+      "sha256": "d6ca2df3416d0c441903d697c21418639839a44e6f43dd92cc0451baaead43fb"
+    },
+    {
+      "path": "worlds/opsbench-14/mcp/manifest.json",
+      "bytes": 1989,
+      "sha256": "d431118d99a969b59a6f86bad4a178ddf673da1501091a11ba744a35df285dd5"
+    },
+    {
+      "path": "worlds/opsbench-14/schema.sql",
+      "bytes": 1442,
+      "sha256": "0e2ed2c463b00b1896d51e02a8abd3b9e743bfcdc73817ed8a09624f5fb8109d"
+    },
+    {
+      "path": "worlds/opsbench-14/seed.sql",
+      "bytes": 13353,
+      "sha256": "9457c0e86a30ba262563ebe9bbac22edeb38d07702018689154ec6e783373934"
+    },
+    {
+      "path": "worlds/opsbench-14/tasks.json",
+      "bytes": 106649,
+      "sha256": "0994054741ca6f7517a42584643e4570f434f7347dcf4fd59587f5525f97b1f8"
+    },
+    {
+      "path": "worlds/opsbench-14/tools.json",
+      "bytes": 99191,
+      "sha256": "c8fbd5a4df7d08a3bdd0ad292f2cbc362f118e03f3c4b4c2da5f2ed7fbe0a315"
+    },
+    {
+      "path": "worlds/opsbench-14/world.json",
+      "bytes": 2539,
+      "sha256": "42447057cbe5a476888045d6601ad7c428d05a7aa3c2bc144cebb176bd85dd1b"
+    },
+    {
+      "path": "worlds/repobench-500/environment.db",
+      "bytes": 32768,
+      "sha256": "12007f3262431078bc3f6476bec0de28d01e0b93acfcc2d86268d1ba36210303"
+    },
+    {
+      "path": "worlds/repobench-500/mcp/manifest.json",
+      "bytes": 1954,
+      "sha256": "1b85dad89ae04a265b46ef7b53d52a8660c7973ca1158442637d84cf25de5faf"
+    },
+    {
+      "path": "worlds/repobench-500/schema.sql",
+      "bytes": 1690,
+      "sha256": "76b797e52cbe538d5907b14387a8da0e6fae3c883e546d5af1c1504a3973255f"
+    },
+    {
+      "path": "worlds/repobench-500/seed.sql",
+      "bytes": 15051,
+      "sha256": "d24ba1b70830c020f4abfa3a702a5e6a1bcb5378202a1a3c605af09fbcdecf47"
+    },
+    {
+      "path": "worlds/repobench-500/tasks.json",
+      "bytes": 106569,
+      "sha256": "c05032fc9010fed55dd7dd853f10622fa4657c7fe1844160270e8644bd26fda0"
+    },
+    {
+      "path": "worlds/repobench-500/tools.json",
+      "bytes": 103816,
+      "sha256": "75077c064c2c026100e6e30008bac57e490f60918e72b48a0b8b18acfa1d61ea"
+    },
+    {
+      "path": "worlds/repobench-500/world.json",
+      "bytes": 2517,
+      "sha256": "c38c5e25cf4ba718e434d8da896a9c90ee02b5f8d25a6c9357c4ee10b2f3ff7d"
+    },
+    {
+      "path": "worlds/servicebench-375/environment.db",
+      "bytes": 32768,
+      "sha256": "f1c39d517436dd11eff05bf2da9440635d53b397ca3966e386e01f7f4a596537"
+    },
+    {
+      "path": "worlds/servicebench-375/mcp/manifest.json",
+      "bytes": 1959,
+      "sha256": "d7fc28450ed8676dec434717efb0d3e3b1b802ab58c4b904504d7b84192e83f5"
+    },
+    {
+      "path": "worlds/servicebench-375/schema.sql",
+      "bytes": 1521,
+      "sha256": "aae113321a7e8003690d02fdf2e6540f74f9735211eaaefacbb8da160f481a2d"
+    },
+    {
+      "path": "worlds/servicebench-375/seed.sql",
+      "bytes": 15471,
+      "sha256": "9b5b097b96bcd595d8f3edc80d437e9881739f350848b4414840effa81dd7839"
+    },
+    {
+      "path": "worlds/servicebench-375/tasks.json",
+      "bytes": 105899,
+      "sha256": "1a7d4d40cff99e0ae7e7fa21dc6fd5122f4f5a730fc939fec2c499e8046a613f"
+    },
+    {
+      "path": "worlds/servicebench-375/tools.json",
+      "bytes": 101198,
+      "sha256": "bbfffdac46d3bd61a8eea7e17c6c74effca6ceb2e2342daaf03cbb65f70da2a5"
+    },
+    {
+      "path": "worlds/servicebench-375/world.json",
+      "bytes": 2495,
+      "sha256": "79dadcb06d047317b285f5f996c9e7568a386854d5e4f997b5667ce39bfda67c"
+    }
+  ]
+}

README.md

@@ -0,0 +1,83 @@
+---
+license: other
+language:
+- en
+pretty_name: Blobfish DomainBench-24
+task_categories:
+- question-answering
+- reinforcement-learning
+tags:
+- agent-evaluation
+- mcp
+- tool-use
+- trajectories
+- sqlite
+size_categories:
+- n<1K
+configs:
+- config_name: tasks
+  data_files:
+  - split: train
+    path: data/tasks.jsonl
+- config_name: trajectories
+  data_files:
+  - split: train
+    path: data/trajectories.jsonl
+- config_name: assets
+  data_files:
+  - split: train
+    path: data/assets.jsonl
+---
+
+# Blobfish DomainBench-24
+
+DomainBench-24 is the public release record for 24 stateful agent tasks across six professional domains. Every task runs in a checked-in SQLite-backed MCP world with typed read/write tools and deterministic verification.
+
+The release includes:
+
+- 24 task records across six independent companion worlds;
+- a 24/24 deterministic reference ceiling and a 0/24 seeded-random floor;
+- 12 full sample trajectories: one successful reference replay and one failing floor episode per world;
+- 42 hashed world assets, including the runnable database, schema, seed, tools, task manifest, MCP manifest, and world manifest.
+
+No reference or random control is a model result. There are no frontier-model rankings in this release.
+
+## Domains and provenance
+
+The Harbor datasets below are upstream professional-domain evaluation suites. The Blobfish worlds are independent diagnostic environments in the same domains; their 24 tasks are not converted or claimed as subsets of the upstream suites.
+
+| Domain | Blobfish dataset page | Upstream Harbor source | Companion world |
+|---|---|---|---|
+| Financial research | [FilingBench-50](https://blobfish.ai/benchmarks/filingbench-50) | [vals/financeagent](https://hub.harborframework.com/datasets/vals/financeagent) | [world](https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/filingbench-50) |
+| Customer operations | [ServiceBench-375](https://blobfish.ai/benchmarks/servicebench-375) | [sierra-research/tau3-bench](https://hub.harborframework.com/datasets/sierra-research/tau3-bench) | [world](https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/servicebench-375) |
+| Scientific discovery | [DiscoveryBench-102](https://blobfish.ai/benchmarks/discoverybench-102) | [scienceagentbench/scienceagentbench](https://hub.harborframework.com/datasets/scienceagentbench/scienceagentbench) | [world](https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/discoverybench-102) |
+| Defensive security | [DefenseBench-100](https://blobfish.ai/benchmarks/defensebench-100) | [polyvorlabs/cyberdefense-bench](https://hub.harborframework.com/datasets/polyvorlabs/cyberdefense-bench) | [world](https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/defensebench-100) |
+| Enterprise operations | [OpsBench-14](https://blobfish.ai/benchmarks/opsbench-14) | [Enterprise-Bench/l1-l2-bench](https://hub.harborframework.com/datasets/Enterprise-Bench/l1-l2-bench) | [world](https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/opsbench-14) |
+| Software engineering | [RepoBench-500](https://blobfish.ai/benchmarks/repobench-500) | [swe-bench/swe-bench-verified](https://hub.harborframework.com/datasets/swe-bench/swe-bench-verified) | [world](https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/repobench-500) |
+
+## Files
+
+- `data/tasks.jsonl` — prompts, split labels, required tools, verifier basis, and control outcomes.
+- `data/trajectories.jsonl` — stable trajectory metadata plus serialized tool steps and verifier results.
+- `data/assets.jsonl` — byte size, SHA-256, and repository URL for each released asset.
+- `evidence/*.json` — complete public evidence bundle used by blobfish.ai.
+- `worlds/<benchmark>/` — the actual released world files.
+- `MANIFEST.json` — SHA-256 and byte length for every uploaded file.
+
+## Evaluation contract
+
+Each episode starts from a fresh copy of the released SQLite state. The harness records tool calls, arguments, outputs, and errors, then grades the final state with deterministic VCode or an exact-answer check. Reference replay proves solvability; the seeded-random failure floor proves basic discrimination. Neither establishes empirical model difficulty.
+
+## License and upstream terms
+
+Blobfish-authored release metadata and scripts follow the terms of the linked Blobfish repository. Upstream benchmark names, task sets, and source projects retain their own licenses. This release does not redistribute upstream Harbor tasks; consult each linked source page before using those separate datasets.
+
+## Reproduce
+
+```bash
+git clone https://github.com/blobfishai/blobfishai.git
+cd blobfishai
+node benchmark/validate.mjs
+```
+
+See the [portfolio page](https://blobfish.ai/benchmarks) for interactive task, asset, environment, trajectory, and methodology views.

data/assets.jsonl

@@ -0,0 +1,42 @@
+{"benchmark_slug": "filingbench-50", "bytes": 32768, "description": "Pristine state copied for every episode.", "kind": "SQLite", "label": "Seeded SQLite snapshot", "path": "environment.db", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/filingbench-50/environment.db", "sha256": "f70a5b84565d5aa98bbcb1aeff9c7a0742213f85c3ad876a46b0d59265875e51"}
+{"benchmark_slug": "filingbench-50", "bytes": 1574, "description": "Human-readable table and constraint definitions.", "kind": "SQL", "label": "Relational schema", "path": "schema.sql", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/filingbench-50/schema.sql", "sha256": "510bf7e2943495cb8a329c4ccce54ffc7376226dedd2b9c11288d8fa78dc40dd"}
+{"benchmark_slug": "filingbench-50", "bytes": 14116, "description": "Human-readable source for the released world state.", "kind": "SQL", "label": "Seed data", "path": "seed.sql", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/filingbench-50/seed.sql", "sha256": "c888acecc74c67054a759370315e8573f3315579c287ac540b21b57d37a3dc44"}
+{"benchmark_slug": "filingbench-50", "bytes": 104507, "description": "Public tool names, schemas, and implementations.", "kind": "JSON", "label": "MCP tool catalog", "path": "tools.json", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/filingbench-50/tools.json", "sha256": "6653f9d2a6e133047d7bb9f30b8ef81a74debb4efc94b94785664894036d4414"}
+{"benchmark_slug": "filingbench-50", "bytes": 106953, "description": "Instructions, splits, reference traces, and isolated verifier programs.", "kind": "JSON", "label": "Task manifest", "path": "tasks.json", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/filingbench-50/tasks.json", "sha256": "96388b0f39e51ed14f718001371783a506e5654586235e0ae1efd4decfbaf80b"}
+{"benchmark_slug": "filingbench-50", "bytes": 2039, "description": "Stdio server declaration for the world.", "kind": "JSON", "label": "MCP manifest", "path": "mcp/manifest.json", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/filingbench-50/mcp/manifest.json", "sha256": "e92914c7b33fd3e5a50cf7c908177962734d3e7e775da258ca4e62f1437fd301"}
+{"benchmark_slug": "filingbench-50", "bytes": 2527, "description": "Content identity, counts, split, and admission evidence.", "kind": "JSON", "label": "World manifest", "path": "world.json", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/filingbench-50/world.json", "sha256": "8475c86667f578b550dcc0bf5c449981036ea1a8c15dc9613906bb30165f972e"}
+{"benchmark_slug": "servicebench-375", "bytes": 32768, "description": "Pristine state copied for every episode.", "kind": "SQLite", "label": "Seeded SQLite snapshot", "path": "environment.db", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/servicebench-375/environment.db", "sha256": "f1c39d517436dd11eff05bf2da9440635d53b397ca3966e386e01f7f4a596537"}
+{"benchmark_slug": "servicebench-375", "bytes": 1521, "description": "Human-readable table and constraint definitions.", "kind": "SQL", "label": "Relational schema", "path": "schema.sql", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/servicebench-375/schema.sql", "sha256": "aae113321a7e8003690d02fdf2e6540f74f9735211eaaefacbb8da160f481a2d"}
+{"benchmark_slug": "servicebench-375", "bytes": 15471, "description": "Human-readable source for the released world state.", "kind": "SQL", "label": "Seed data", "path": "seed.sql", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/servicebench-375/seed.sql", "sha256": "9b5b097b96bcd595d8f3edc80d437e9881739f350848b4414840effa81dd7839"}
+{"benchmark_slug": "servicebench-375", "bytes": 101198, "description": "Public tool names, schemas, and implementations.", "kind": "JSON", "label": "MCP tool catalog", "path": "tools.json", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/servicebench-375/tools.json", "sha256": "bbfffdac46d3bd61a8eea7e17c6c74effca6ceb2e2342daaf03cbb65f70da2a5"}
+{"benchmark_slug": "servicebench-375", "bytes": 105899, "description": "Instructions, splits, reference traces, and isolated verifier programs.", "kind": "JSON", "label": "Task manifest", "path": "tasks.json", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/servicebench-375/tasks.json", "sha256": "1a7d4d40cff99e0ae7e7fa21dc6fd5122f4f5a730fc939fec2c499e8046a613f"}
+{"benchmark_slug": "servicebench-375", "bytes": 1959, "description": "Stdio server declaration for the world.", "kind": "JSON", "label": "MCP manifest", "path": "mcp/manifest.json", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/servicebench-375/mcp/manifest.json", "sha256": "d7fc28450ed8676dec434717efb0d3e3b1b802ab58c4b904504d7b84192e83f5"}
+{"benchmark_slug": "servicebench-375", "bytes": 2495, "description": "Content identity, counts, split, and admission evidence.", "kind": "JSON", "label": "World manifest", "path": "world.json", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/servicebench-375/world.json", "sha256": "79dadcb06d047317b285f5f996c9e7568a386854d5e4f997b5667ce39bfda67c"}
+{"benchmark_slug": "discoverybench-102", "bytes": 32768, "description": "Pristine state copied for every episode.", "kind": "SQLite", "label": "Seeded SQLite snapshot", "path": "environment.db", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/discoverybench-102/environment.db", "sha256": "44b907f65127934c49e2569a0226486c470231e635d324cec1aa09ebf4a6df05"}
+{"benchmark_slug": "discoverybench-102", "bytes": 1497, "description": "Human-readable table and constraint definitions.", "kind": "SQL", "label": "Relational schema", "path": "schema.sql", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/discoverybench-102/schema.sql", "sha256": "4dfc04a1e3d89b3839449644c07633a4bc2b86d65bc8d6d459ac0c3bb4989c83"}
+{"benchmark_slug": "discoverybench-102", "bytes": 11964, "description": "Human-readable source for the released world state.", "kind": "SQL", "label": "Seed data", "path": "seed.sql", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/discoverybench-102/seed.sql", "sha256": "e3e3c20b66558ba03fff8da8292f92b28b031e1f2dc7a2310e5e702cff3398bc"}
+{"benchmark_slug": "discoverybench-102", "bytes": 97373, "description": "Public tool names, schemas, and implementations.", "kind": "JSON", "label": "MCP tool catalog", "path": "tools.json", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/discoverybench-102/tools.json", "sha256": "511ab9dc9a947171a0a70b0d75d1201d40298c285f217b904ba69ef631ae5e5a"}
+{"benchmark_slug": "discoverybench-102", "bytes": 103570, "description": "Instructions, splits, reference traces, and isolated verifier programs.", "kind": "JSON", "label": "Task manifest", "path": "tasks.json", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/discoverybench-102/tasks.json", "sha256": "73b32ea4a6834a886e55ae6cf26412bec6eadff9e06959912dbf905ab98881aa"}
+{"benchmark_slug": "discoverybench-102", "bytes": 1862, "description": "Stdio server declaration for the world.", "kind": "JSON", "label": "MCP manifest", "path": "mcp/manifest.json", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/discoverybench-102/mcp/manifest.json", "sha256": "6dfe7eff3b010b684dadaac82edc606243c2582461bb691c512da2f3d7c074a4"}
+{"benchmark_slug": "discoverybench-102", "bytes": 2492, "description": "Content identity, counts, split, and admission evidence.", "kind": "JSON", "label": "World manifest", "path": "world.json", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/discoverybench-102/world.json", "sha256": "9788c5cbc28caa5c426e82ab59d220ad7d7aaccf913b4a77dc0e7a8e4f4c6f14"}
+{"benchmark_slug": "defensebench-100", "bytes": 32768, "description": "Pristine state copied for every episode.", "kind": "SQLite", "label": "Seeded SQLite snapshot", "path": "environment.db", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/defensebench-100/environment.db", "sha256": "4940300240cbe15d35d1fd45016398abe150ca2ee0361d04b2a17432f8a26233"}
+{"benchmark_slug": "defensebench-100", "bytes": 1251, "description": "Human-readable table and constraint definitions.", "kind": "SQL", "label": "Relational schema", "path": "schema.sql", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/defensebench-100/schema.sql", "sha256": "08c2e8abf194ee2764b390f01b9e8185c8653950c9582b3e2210e0f5f4a7b70b"}
+{"benchmark_slug": "defensebench-100", "bytes": 14152, "description": "Human-readable source for the released world state.", "kind": "SQL", "label": "Seed data", "path": "seed.sql", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/defensebench-100/seed.sql", "sha256": "9c2b62927c80c430b0808605fbdefa36c200ef146cd539f6d02d6b1acb66573c"}
+{"benchmark_slug": "defensebench-100", "bytes": 90244, "description": "Public tool names, schemas, and implementations.", "kind": "JSON", "label": "MCP tool catalog", "path": "tools.json", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/defensebench-100/tools.json", "sha256": "46021a577b0d1e7e3d8c5befed8ac5f6ea9d3d84ce61fc8e34a446fe59b35c43"}
+{"benchmark_slug": "defensebench-100", "bytes": 106978, "description": "Instructions, splits, reference traces, and isolated verifier programs.", "kind": "JSON", "label": "Task manifest", "path": "tasks.json", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/defensebench-100/tasks.json", "sha256": "44ae2da2ce9ef5faccb06335c0a242eabcdd8e41626925131bb01665d1165aa5"}
+{"benchmark_slug": "defensebench-100", "bytes": 1969, "description": "Stdio server declaration for the world.", "kind": "JSON", "label": "MCP manifest", "path": "mcp/manifest.json", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/defensebench-100/mcp/manifest.json", "sha256": "3c809b60c7ab96026e4ee0f4ac5ba0d04b011005465adff97185e0f98d1893ac"}
+{"benchmark_slug": "defensebench-100", "bytes": 2540, "description": "Content identity, counts, split, and admission evidence.", "kind": "JSON", "label": "World manifest", "path": "world.json", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/defensebench-100/world.json", "sha256": "c8a86f80f071188223d082706306fb84984c1fbb6cce9341f4ab0f8e4a48ef8a"}
+{"benchmark_slug": "opsbench-14", "bytes": 32768, "description": "Pristine state copied for every episode.", "kind": "SQLite", "label": "Seeded SQLite snapshot", "path": "environment.db", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/opsbench-14/environment.db", "sha256": "d6ca2df3416d0c441903d697c21418639839a44e6f43dd92cc0451baaead43fb"}
+{"benchmark_slug": "opsbench-14", "bytes": 1442, "description": "Human-readable table and constraint definitions.", "kind": "SQL", "label": "Relational schema", "path": "schema.sql", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/opsbench-14/schema.sql", "sha256": "0e2ed2c463b00b1896d51e02a8abd3b9e743bfcdc73817ed8a09624f5fb8109d"}
+{"benchmark_slug": "opsbench-14", "bytes": 13353, "description": "Human-readable source for the released world state.", "kind": "SQL", "label": "Seed data", "path": "seed.sql", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/opsbench-14/seed.sql", "sha256": "9457c0e86a30ba262563ebe9bbac22edeb38d07702018689154ec6e783373934"}
+{"benchmark_slug": "opsbench-14", "bytes": 99191, "description": "Public tool names, schemas, and implementations.", "kind": "JSON", "label": "MCP tool catalog", "path": "tools.json", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/opsbench-14/tools.json", "sha256": "c8fbd5a4df7d08a3bdd0ad292f2cbc362f118e03f3c4b4c2da5f2ed7fbe0a315"}
+{"benchmark_slug": "opsbench-14", "bytes": 106649, "description": "Instructions, splits, reference traces, and isolated verifier programs.", "kind": "JSON", "label": "Task manifest", "path": "tasks.json", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/opsbench-14/tasks.json", "sha256": "0994054741ca6f7517a42584643e4570f434f7347dcf4fd59587f5525f97b1f8"}
+{"benchmark_slug": "opsbench-14", "bytes": 1989, "description": "Stdio server declaration for the world.", "kind": "JSON", "label": "MCP manifest", "path": "mcp/manifest.json", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/opsbench-14/mcp/manifest.json", "sha256": "d431118d99a969b59a6f86bad4a178ddf673da1501091a11ba744a35df285dd5"}
+{"benchmark_slug": "opsbench-14", "bytes": 2539, "description": "Content identity, counts, split, and admission evidence.", "kind": "JSON", "label": "World manifest", "path": "world.json", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/opsbench-14/world.json", "sha256": "42447057cbe5a476888045d6601ad7c428d05a7aa3c2bc144cebb176bd85dd1b"}
+{"benchmark_slug": "repobench-500", "bytes": 32768, "description": "Pristine state copied for every episode.", "kind": "SQLite", "label": "Seeded SQLite snapshot", "path": "environment.db", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/repobench-500/environment.db", "sha256": "12007f3262431078bc3f6476bec0de28d01e0b93acfcc2d86268d1ba36210303"}
+{"benchmark_slug": "repobench-500", "bytes": 1690, "description": "Human-readable table and constraint definitions.", "kind": "SQL", "label": "Relational schema", "path": "schema.sql", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/repobench-500/schema.sql", "sha256": "76b797e52cbe538d5907b14387a8da0e6fae3c883e546d5af1c1504a3973255f"}
+{"benchmark_slug": "repobench-500", "bytes": 15051, "description": "Human-readable source for the released world state.", "kind": "SQL", "label": "Seed data", "path": "seed.sql", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/repobench-500/seed.sql", "sha256": "d24ba1b70830c020f4abfa3a702a5e6a1bcb5378202a1a3c605af09fbcdecf47"}
+{"benchmark_slug": "repobench-500", "bytes": 103816, "description": "Public tool names, schemas, and implementations.", "kind": "JSON", "label": "MCP tool catalog", "path": "tools.json", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/repobench-500/tools.json", "sha256": "75077c064c2c026100e6e30008bac57e490f60918e72b48a0b8b18acfa1d61ea"}
+{"benchmark_slug": "repobench-500", "bytes": 106569, "description": "Instructions, splits, reference traces, and isolated verifier programs.", "kind": "JSON", "label": "Task manifest", "path": "tasks.json", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/repobench-500/tasks.json", "sha256": "c05032fc9010fed55dd7dd853f10622fa4657c7fe1844160270e8644bd26fda0"}
+{"benchmark_slug": "repobench-500", "bytes": 1954, "description": "Stdio server declaration for the world.", "kind": "JSON", "label": "MCP manifest", "path": "mcp/manifest.json", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/repobench-500/mcp/manifest.json", "sha256": "1b85dad89ae04a265b46ef7b53d52a8660c7973ca1158442637d84cf25de5faf"}
+{"benchmark_slug": "repobench-500", "bytes": 2517, "description": "Content identity, counts, split, and admission evidence.", "kind": "JSON", "label": "World manifest", "path": "world.json", "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/repobench-500/world.json", "sha256": "c38c5e25cf4ba718e434d8da896a9c90ee02b5f8d25a6c9357c4ee10b2f3ff7d"}

data/tasks.jsonl

@@ -0,0 +1,24 @@
+{"benchmark_name": "FilingBench-50", "benchmark_slug": "filingbench-50", "companion_relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks.", "difficulty_disclosure": "Generator label only; no empirical model difficulty claim is published.", "domain": "Financial research", "generator_difficulty_label": "hard", "instruction": "Finalise the Solstice base-case DCF (model_id 9201, company 9001). Total the free_cash_flow across every financial period belonging to company 9001 — other companies' periods do not count. Set model 9201's enterprise_value to that total and its status to 'final', mark its WACC assumption (assumption_id 9301) as status 'verified', and file the review: create a review note on model 9201 from reviewer 'Head of Research' with status 'open' and a body containing the summed cash flow.", "reference_passed": true, "reference_steps": 4, "required_tools": ["search_financial_periods", "update_dcf_model", "update_dcf_assumption", "create_review_note"], "seeded_random_passed": false, "source_dataset_id": "vals/financeagent", "source_dataset_url": "https://hub.harborframework.com/datasets/vals/financeagent", "split": "train", "task_id": "fa_finalize_solstice_dcf", "verifier_basis": "vcode", "world_digest": "cd3085c1e21b8793bd4ba01fe48bd2a711e793f8233d931b4944fa1bafadaf41", "world_id": "env_blobfish_benchmarks_filingbench_financia_599d3443", "world_repository_url": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/filingbench-50"}
+{"benchmark_name": "FilingBench-50", "benchmark_slug": "filingbench-50", "companion_relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks.", "difficulty_disclosure": "Generator label only; no empirical model difficulty claim is published.", "domain": "Financial research", "generator_difficulty_label": "medium", "instruction": "Model governance: a DCF may not carry status 'final' with a discount_rate of 0 — there is nothing to discount by. Find that model and set its status back to 'draft'. Models that are already draft are left alone.", "reference_passed": true, "reference_steps": 2, "required_tools": ["search_dcf_models", "update_dcf_model"], "seeded_random_passed": false, "source_dataset_id": "vals/financeagent", "source_dataset_url": "https://hub.harborframework.com/datasets/vals/financeagent", "split": "train", "task_id": "fa_revert_model_without_discount_rate", "verifier_basis": "vcode", "world_digest": "cd3085c1e21b8793bd4ba01fe48bd2a711e793f8233d931b4944fa1bafadaf41", "world_id": "env_blobfish_benchmarks_filingbench_financia_599d3443", "world_repository_url": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/filingbench-50"}
+{"benchmark_name": "FilingBench-50", "benchmark_slug": "filingbench-50", "companion_relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks.", "difficulty_disclosure": "Generator label only; no empirical model difficulty claim is published.", "domain": "Financial research", "generator_difficulty_label": "easy", "instruction": "The same issuer appears twice under two listings. How many companies are named exactly 'Solstice Energy'? Answer with the number.", "reference_passed": true, "reference_steps": 1, "required_tools": ["count_fa_companies"], "seeded_random_passed": false, "source_dataset_id": "vals/financeagent", "source_dataset_url": "https://hub.harborframework.com/datasets/vals/financeagent", "split": "train", "task_id": "fa_duplicate_company_listing", "verifier_basis": "answer_match (vcode does not discriminate)", "world_digest": "cd3085c1e21b8793bd4ba01fe48bd2a711e793f8233d931b4944fa1bafadaf41", "world_id": "env_blobfish_benchmarks_filingbench_financia_599d3443", "world_repository_url": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/filingbench-50"}
+{"benchmark_name": "FilingBench-50", "benchmark_slug": "filingbench-50", "companion_relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks.", "difficulty_disclosure": "Generator label only; no empirical model difficulty claim is published.", "domain": "Financial research", "generator_difficulty_label": "medium", "instruction": "Model 9202 carries the 'Terminal growth' assumption twice (assumption_id 9311 and 9312). Delete the duplicate 9312 and keep 9311.", "reference_passed": true, "reference_steps": 2, "required_tools": ["search_dcf_assumptions", "delete_dcf_assumption"], "seeded_random_passed": false, "source_dataset_id": "vals/financeagent", "source_dataset_url": "https://hub.harborframework.com/datasets/vals/financeagent", "split": "heldout", "task_id": "fa_dedupe_dcf_assumption", "verifier_basis": "vcode", "world_digest": "cd3085c1e21b8793bd4ba01fe48bd2a711e793f8233d931b4944fa1bafadaf41", "world_id": "env_blobfish_benchmarks_filingbench_financia_599d3443", "world_repository_url": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/filingbench-50"}
+{"benchmark_name": "ServiceBench-375", "benchmark_slug": "servicebench-375", "companion_relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks.", "difficulty_disclosure": "Generator label only; no empirical model difficulty claim is published.", "domain": "Customer operations", "generator_difficulty_label": "hard", "instruction": "It is now minute 180. Ticket 9201 belongs to an enterprise-plan customer and still has no first response. Look up the first_response_minutes of the SLA policy for that customer's plan and work out how many minutes past the SLA deadline the response now is (deadline = opened_minute + first_response_minutes). Then set ticket 9201's priority to 'urgent' and status to 'escalated', create an escalation on ticket 9201 with reason 'sla_first_response_breach', tier 'tier2' and status 'open', and post a holding reply: a ticket message on 9201 from author 'agent' at sent_minute 180 whose body states the overdue amount as 'N minutes'.", "reference_passed": true, "reference_steps": 4, "required_tools": ["search_sla_policies", "update_ticket", "create_escalation", "create_ticket_message"], "seeded_random_passed": false, "source_dataset_id": "sierra-research/tau3-bench", "source_dataset_url": "https://hub.harborframework.com/datasets/sierra-research/tau3-bench", "split": "train", "task_id": "cs_escalate_sla_breach", "verifier_basis": "vcode", "world_digest": "ab0d887316ac287e98532ff96cd1ec24f979922a054a8e02b58e4d2e75f37260", "world_id": "env_blobfish_benchmarks_servicebench_custome_79e1b30f", "world_repository_url": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/servicebench-375"}
+{"benchmark_name": "ServiceBench-375", "benchmark_slug": "servicebench-375", "companion_relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks.", "difficulty_disclosure": "Generator label only; no empirical model difficulty claim is published.", "domain": "Customer operations", "generator_difficulty_label": "medium", "instruction": "Tidy the queue: any ticket that already has a resolved_minute recorded must not still be sitting in status 'open' — set it to 'resolved'. Tickets without a resolved_minute stay open.", "reference_passed": true, "reference_steps": 2, "required_tools": ["search_tickets", "update_ticket"], "seeded_random_passed": false, "source_dataset_id": "sierra-research/tau3-bench", "source_dataset_url": "https://hub.harborframework.com/datasets/sierra-research/tau3-bench", "split": "train", "task_id": "cs_close_resolved_tickets", "verifier_basis": "vcode", "world_digest": "ab0d887316ac287e98532ff96cd1ec24f979922a054a8e02b58e4d2e75f37260", "world_id": "env_blobfish_benchmarks_servicebench_custome_79e1b30f", "world_repository_url": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/servicebench-375"}
+{"benchmark_name": "ServiceBench-375", "benchmark_slug": "servicebench-375", "companion_relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks.", "difficulty_disclosure": "Generator label only; no empirical model difficulty claim is published.", "domain": "Customer operations", "generator_difficulty_label": "easy", "instruction": "Two subsidiaries were onboarded under one trading name, so routing is ambiguous. How many support customers are named exactly 'Bluepeak Robotics'? Answer with the number.", "reference_passed": true, "reference_steps": 1, "required_tools": ["count_support_customers"], "seeded_random_passed": false, "source_dataset_id": "sierra-research/tau3-bench", "source_dataset_url": "https://hub.harborframework.com/datasets/sierra-research/tau3-bench", "split": "train", "task_id": "cs_duplicate_customer_name", "verifier_basis": "answer_match (vcode does not discriminate)", "world_digest": "ab0d887316ac287e98532ff96cd1ec24f979922a054a8e02b58e4d2e75f37260", "world_id": "env_blobfish_benchmarks_servicebench_custome_79e1b30f", "world_repository_url": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/servicebench-375"}
+{"benchmark_name": "ServiceBench-375", "benchmark_slug": "servicebench-375", "companion_relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks.", "difficulty_disclosure": "Generator label only; no empirical model difficulty claim is published.", "domain": "Customer operations", "generator_difficulty_label": "medium", "instruction": "Two knowledge-base articles cover the same fix under the identical title 'Recovering an offline fleet controller' (article_id 9501 is published, 9502 is a duplicate draft). Delete the duplicate draft 9502 and keep the published article 9501.", "reference_passed": true, "reference_steps": 2, "required_tools": ["search_kb_articles", "delete_kb_article"], "seeded_random_passed": false, "source_dataset_id": "sierra-research/tau3-bench", "source_dataset_url": "https://hub.harborframework.com/datasets/sierra-research/tau3-bench", "split": "heldout", "task_id": "cs_dedupe_kb_article", "verifier_basis": "vcode", "world_digest": "ab0d887316ac287e98532ff96cd1ec24f979922a054a8e02b58e4d2e75f37260", "world_id": "env_blobfish_benchmarks_servicebench_custome_79e1b30f", "world_repository_url": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/servicebench-375"}
+{"benchmark_name": "DiscoveryBench-102", "benchmark_slug": "discoverybench-102", "companion_relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks.", "difficulty_disclosure": "Generator label only; no empirical model difficulty claim is published.", "domain": "Scientific discovery", "generator_difficulty_label": "hard", "instruction": "Call the potency of assay run 9301. Compute the mean ic50_nm across its replicates that PASSED QC (qc_flag 'pass') — replicates flagged 'fail' are excluded. A mean below 100 nM makes the compound a hit, so: set compound 9101's status to 'hit' and record the mean in its screening_note, set assay run 9301's status to 'complete', and queue confirmation by creating an assay run on assay 9201 for compound 9101 with run_on '2026-08-01', plate_label 'CONF-001' and status 'pending'.", "reference_passed": true, "reference_steps": 4, "required_tools": ["search_assay_results", "update_compound", "update_assay_run", "create_assay_run"], "seeded_random_passed": false, "source_dataset_id": "scienceagentbench/scienceagentbench", "source_dataset_url": "https://hub.harborframework.com/datasets/scienceagentbench/scienceagentbench", "split": "train", "task_id": "bio_promote_hit_compound", "verifier_basis": "vcode", "world_digest": "475b17df4a7efe8265400a4d47889288764829150a4dc63e9b1188a06948cd29", "world_id": "env_blobfish_benchmarks_discoverybench_scien_f7e1dbce", "world_repository_url": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/discoverybench-102"}
+{"benchmark_name": "DiscoveryBench-102", "benchmark_slug": "discoverybench-102", "companion_relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks.", "difficulty_disclosure": "Generator label only; no empirical model difficulty claim is published.", "domain": "Scientific discovery", "generator_difficulty_label": "medium", "instruction": "Quality control: an assay run where EVERY replicate carries qc_flag 'fail' produced no usable data and must be set to status 'invalid'. Runs with at least one passing replicate are left alone.", "reference_passed": true, "reference_steps": 2, "required_tools": ["search_assay_results", "update_assay_run"], "seeded_random_passed": false, "source_dataset_id": "scienceagentbench/scienceagentbench", "source_dataset_url": "https://hub.harborframework.com/datasets/scienceagentbench/scienceagentbench", "split": "train", "task_id": "bio_invalidate_failed_run", "verifier_basis": "vcode", "world_digest": "475b17df4a7efe8265400a4d47889288764829150a4dc63e9b1188a06948cd29", "world_id": "env_blobfish_benchmarks_discoverybench_scien_f7e1dbce", "world_repository_url": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/discoverybench-102"}
+{"benchmark_name": "DiscoveryBench-102", "benchmark_slug": "discoverybench-102", "companion_relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks.", "difficulty_disclosure": "Generator label only; no empirical model difficulty claim is published.", "domain": "Scientific discovery", "generator_difficulty_label": "easy", "instruction": "Two structurally distinct compounds were filed under the same common name. How many compounds are named exactly 'Velatinib'? Answer with the number.", "reference_passed": true, "reference_steps": 1, "required_tools": ["count_compounds"], "seeded_random_passed": false, "source_dataset_id": "scienceagentbench/scienceagentbench", "source_dataset_url": "https://hub.harborframework.com/datasets/scienceagentbench/scienceagentbench", "split": "train", "task_id": "bio_duplicate_compound_name", "verifier_basis": "answer_match (vcode does not discriminate)", "world_digest": "475b17df4a7efe8265400a4d47889288764829150a4dc63e9b1188a06948cd29", "world_id": "env_blobfish_benchmarks_discoverybench_scien_f7e1dbce", "world_repository_url": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/discoverybench-102"}
+{"benchmark_name": "DiscoveryBench-102", "benchmark_slug": "discoverybench-102", "companion_relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks.", "difficulty_disclosure": "Generator label only; no empirical model difficulty claim is published.", "domain": "Scientific discovery", "generator_difficulty_label": "medium", "instruction": "Run 9302 has replicate 1 recorded twice (result_id 9411 and 9412) because the plate reader exported twice. Delete the duplicate 9412 and keep 9411.", "reference_passed": true, "reference_steps": 2, "required_tools": ["search_assay_results", "delete_assay_result"], "seeded_random_passed": false, "source_dataset_id": "scienceagentbench/scienceagentbench", "source_dataset_url": "https://hub.harborframework.com/datasets/scienceagentbench/scienceagentbench", "split": "heldout", "task_id": "bio_dedupe_assay_result", "verifier_basis": "vcode", "world_digest": "475b17df4a7efe8265400a4d47889288764829150a4dc63e9b1188a06948cd29", "world_id": "env_blobfish_benchmarks_discoverybench_scien_f7e1dbce", "world_repository_url": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/discoverybench-102"}
+{"benchmark_name": "DefenseBench-100", "benchmark_slug": "defensebench-100", "companion_relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks.", "difficulty_disclosure": "Generator label only; no empirical model difficulty claim is published.", "domain": "Defensive security", "generator_difficulty_label": "hard", "instruction": "Document 9201 ('FY27 Budget Model') is confidential and may only be readable by the Finance team. Review its access grants: revoke every still-active grant held by an employee who is not on the Finance team by setting that grant's status to 'revoked', leave Finance grants and already-revoked grants alone, do not touch grants on other documents, and finally set document 9201's status to 'restricted'.", "reference_passed": true, "reference_steps": 5, "required_tools": ["search_access_grants", "search_employees", "update_access_grant", "update_indexed_document"], "seeded_random_passed": false, "source_dataset_id": "polyvorlabs/cyberdefense-bench", "source_dataset_url": "https://hub.harborframework.com/datasets/polyvorlabs/cyberdefense-bench", "split": "train", "task_id": "es_lock_down_confidential_document", "verifier_basis": "vcode", "world_digest": "e1df820514e93081e497aa1aa2fd50423e3dbf0cd766e4850838096a1bc641f6", "world_id": "env_blobfish_benchmarks_defensebench_access__d2a0878c", "world_repository_url": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/defensebench-100"}
+{"benchmark_name": "DefenseBench-100", "benchmark_slug": "defensebench-100", "companion_relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks.", "difficulty_disclosure": "Generator label only; no empirical model difficulty claim is published.", "domain": "Defensive security", "generator_difficulty_label": "medium", "instruction": "A connector has been disconnected. Every indexed document that came from a source with connected = 0 must be set to status 'stale'; documents from connected sources stay 'indexed'.", "reference_passed": true, "reference_steps": 3, "required_tools": ["search_sources", "search_indexed_documents", "update_indexed_document"], "seeded_random_passed": false, "source_dataset_id": "polyvorlabs/cyberdefense-bench", "source_dataset_url": "https://hub.harborframework.com/datasets/polyvorlabs/cyberdefense-bench", "split": "train", "task_id": "es_mark_disconnected_source_stale", "verifier_basis": "vcode", "world_digest": "e1df820514e93081e497aa1aa2fd50423e3dbf0cd766e4850838096a1bc641f6", "world_id": "env_blobfish_benchmarks_defensebench_access__d2a0878c", "world_repository_url": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/defensebench-100"}
+{"benchmark_name": "DefenseBench-100", "benchmark_slug": "defensebench-100", "companion_relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks.", "difficulty_disclosure": "Generator label only; no empirical model difficulty claim is published.", "domain": "Defensive security", "generator_difficulty_label": "easy", "instruction": "Two different employees share a display name, which makes people search ambiguous. How many employees are named exactly 'Rowan Ferreira'? Answer with the number.", "reference_passed": true, "reference_steps": 1, "required_tools": ["count_employees"], "seeded_random_passed": false, "source_dataset_id": "polyvorlabs/cyberdefense-bench", "source_dataset_url": "https://hub.harborframework.com/datasets/polyvorlabs/cyberdefense-bench", "split": "train", "task_id": "es_duplicate_person_name", "verifier_basis": "answer_match (vcode does not discriminate)", "world_digest": "e1df820514e93081e497aa1aa2fd50423e3dbf0cd766e4850838096a1bc641f6", "world_id": "env_blobfish_benchmarks_defensebench_access__d2a0878c", "world_repository_url": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/defensebench-100"}
+{"benchmark_name": "DefenseBench-100", "benchmark_slug": "defensebench-100", "companion_relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks.", "difficulty_disclosure": "Generator label only; no empirical model difficulty claim is published.", "domain": "Defensive security", "generator_difficulty_label": "medium", "instruction": "A double sync indexed the same passage of document 9201 twice (chunk_id 9301 and 9302), so it is over-weighted in retrieval. Delete the duplicate chunk 9302 and keep 9301.", "reference_passed": true, "reference_steps": 2, "required_tools": ["search_document_chunks", "delete_document_chunk"], "seeded_random_passed": false, "source_dataset_id": "polyvorlabs/cyberdefense-bench", "source_dataset_url": "https://hub.harborframework.com/datasets/polyvorlabs/cyberdefense-bench", "split": "heldout", "task_id": "es_dedupe_document_chunk", "verifier_basis": "vcode", "world_digest": "e1df820514e93081e497aa1aa2fd50423e3dbf0cd766e4850838096a1bc641f6", "world_id": "env_blobfish_benchmarks_defensebench_access__d2a0878c", "world_repository_url": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/defensebench-100"}
+{"benchmark_name": "OpsBench-14", "benchmark_slug": "opsbench-14", "companion_relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks.", "difficulty_disclosure": "Generator label only; no empirical model difficulty claim is published.", "domain": "Enterprise operations", "generator_difficulty_label": "hard", "instruction": "Change request 9301 is a high-risk change to the Vendor Onboarding process (doc_id 9101). Policy: a high-risk change cannot be approved while any step of that process has no control_ref mapped. Count the DISTINCT steps of doc 9101 with an empty control_ref — note step 9202 is a duplicate of 9201 and counts once. Because at least one step is uncontrolled: set change request 9301's status to 'rejected', set process doc 9101's status to 'needs_controls', and open the remediation by creating a compliance control named 'Step control backfill' with framework 'SOC2' and status 'untested'.", "reference_passed": true, "reference_steps": 4, "required_tools": ["search_process_steps", "update_change_request", "update_process_doc", "create_compliance_control"], "seeded_random_passed": false, "source_dataset_id": "Enterprise-Bench/l1-l2-bench", "source_dataset_url": "https://hub.harborframework.com/datasets/Enterprise-Bench/l1-l2-bench", "split": "heldout", "task_id": "ops_reject_uncontrolled_change", "verifier_basis": "vcode", "world_digest": "39937737e9b493a2e1a4a10ba85780b15de76a222cf180538e40969cb6e3f1a8", "world_id": "env_blobfish_benchmarks_opsbench_enterprise__8377a7f8", "world_repository_url": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/opsbench-14"}
+{"benchmark_name": "OpsBench-14", "benchmark_slug": "opsbench-14", "companion_relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks.", "difficulty_disclosure": "Generator label only; no empirical model difficulty claim is published.", "domain": "Enterprise operations", "generator_difficulty_label": "medium", "instruction": "Today is 2026-07-20. Any active vendor whose renewal_on date has already passed must be set to status 'renewal_overdue'. Vendors whose renewal is still ahead stay 'active'.", "reference_passed": true, "reference_steps": 2, "required_tools": ["search_vendors", "update_vendor"], "seeded_random_passed": false, "source_dataset_id": "Enterprise-Bench/l1-l2-bench", "source_dataset_url": "https://hub.harborframework.com/datasets/Enterprise-Bench/l1-l2-bench", "split": "train", "task_id": "ops_flag_overdue_vendor_renewal", "verifier_basis": "vcode", "world_digest": "39937737e9b493a2e1a4a10ba85780b15de76a222cf180538e40969cb6e3f1a8", "world_id": "env_blobfish_benchmarks_opsbench_enterprise__8377a7f8", "world_repository_url": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/opsbench-14"}
+{"benchmark_name": "OpsBench-14", "benchmark_slug": "opsbench-14", "companion_relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks.", "difficulty_disclosure": "Generator label only; no empirical model difficulty claim is published.", "domain": "Enterprise operations", "generator_difficulty_label": "easy", "instruction": "Two legal entities are registered under one trading name, so spend is split. How many vendors are named exactly 'Cascade Facilities Ltd'? Answer with the number.", "reference_passed": true, "reference_steps": 1, "required_tools": ["count_vendors"], "seeded_random_passed": false, "source_dataset_id": "Enterprise-Bench/l1-l2-bench", "source_dataset_url": "https://hub.harborframework.com/datasets/Enterprise-Bench/l1-l2-bench", "split": "heldout", "task_id": "ops_duplicate_vendor_name", "verifier_basis": "answer_match (vcode does not discriminate)", "world_digest": "39937737e9b493a2e1a4a10ba85780b15de76a222cf180538e40969cb6e3f1a8", "world_id": "env_blobfish_benchmarks_opsbench_enterprise__8377a7f8", "world_repository_url": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/opsbench-14"}
+{"benchmark_name": "OpsBench-14", "benchmark_slug": "opsbench-14", "companion_relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks.", "difficulty_disclosure": "Generator label only; no empirical model difficulty claim is published.", "domain": "Enterprise operations", "generator_difficulty_label": "medium", "instruction": "A bad merge left step 1 of the Vendor Onboarding process recorded twice (step_id 9201 and 9202). Delete the duplicate 9202 and keep 9201.", "reference_passed": true, "reference_steps": 2, "required_tools": ["search_process_steps", "delete_process_step"], "seeded_random_passed": false, "source_dataset_id": "Enterprise-Bench/l1-l2-bench", "source_dataset_url": "https://hub.harborframework.com/datasets/Enterprise-Bench/l1-l2-bench", "split": "train", "task_id": "ops_dedupe_process_step", "verifier_basis": "vcode", "world_digest": "39937737e9b493a2e1a4a10ba85780b15de76a222cf180538e40969cb6e3f1a8", "world_id": "env_blobfish_benchmarks_opsbench_enterprise__8377a7f8", "world_repository_url": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/opsbench-14"}
+{"benchmark_name": "RepoBench-500", "benchmark_slug": "repobench-500", "companion_relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks.", "difficulty_disclosure": "Generator label only; no empirical model difficulty claim is published.", "domain": "Software engineering", "generator_difficulty_label": "hard", "instruction": "Incident 9401 on the ingest service was mitigated at minute 175. Close it out: set incident 9401's resolved_minute to 175 and status to 'resolved', work out how long it lasted (resolved_minute minus detected_minute), then record the follow-up decision by creating an ADR on service 9001 titled 'Rate-limit the ingest queue' with status 'accepted', decided_on '2026-07-21' and a decision that states the incident duration as 'N minutes'; finally post a standup note on service 9001 by author 'Rowan Novak' for day_label 'tuesday' whose body references incident 9401.", "reference_passed": true, "reference_steps": 4, "required_tools": ["get_incident", "update_incident", "create_adr_record", "create_standup_note"], "seeded_random_passed": false, "source_dataset_id": "swe-bench/swe-bench-verified", "source_dataset_url": "https://hub.harborframework.com/datasets/swe-bench/swe-bench-verified", "split": "train", "task_id": "eng_resolve_ingest_incident", "verifier_basis": "vcode", "world_digest": "482b4cabbd31f60b18d71d5d709ae725e85c1a57f5736fdee8a40b4677546a4e", "world_id": "env_blobfish_benchmarks_repobench_software_e_0af840ee", "world_repository_url": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/repobench-500"}
+{"benchmark_name": "RepoBench-500", "benchmark_slug": "repobench-500", "companion_relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks.", "difficulty_disclosure": "Generator label only; no empirical model difficulty claim is published.", "domain": "Software engineering", "generator_difficulty_label": "medium", "instruction": "Merge policy: a pull request may not sit in status 'approved' while its checks_status is 'failing'. Find the offending PR and set its status back to 'open'. PRs whose checks pass keep their approved status.", "reference_passed": true, "reference_steps": 2, "required_tools": ["search_pull_requests", "update_pull_request"], "seeded_random_passed": false, "source_dataset_id": "swe-bench/swe-bench-verified", "source_dataset_url": "https://hub.harborframework.com/datasets/swe-bench/swe-bench-verified", "split": "train", "task_id": "eng_reopen_pr_with_red_ci", "verifier_basis": "vcode", "world_digest": "482b4cabbd31f60b18d71d5d709ae725e85c1a57f5736fdee8a40b4677546a4e", "world_id": "env_blobfish_benchmarks_repobench_software_e_0af840ee", "world_repository_url": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/repobench-500"}
+{"benchmark_name": "RepoBench-500", "benchmark_slug": "repobench-500", "companion_relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks.", "difficulty_disclosure": "Generator label only; no empirical model difficulty claim is published.", "domain": "Software engineering", "generator_difficulty_label": "easy", "instruction": "After a service split, two services own a repository with the same name. How many repositories are named exactly 'ingest-worker'? Answer with the number.", "reference_passed": true, "reference_steps": 1, "required_tools": ["count_repositories"], "seeded_random_passed": false, "source_dataset_id": "swe-bench/swe-bench-verified", "source_dataset_url": "https://hub.harborframework.com/datasets/swe-bench/swe-bench-verified", "split": "train", "task_id": "eng_duplicate_repository_name", "verifier_basis": "answer_match (vcode does not discriminate)", "world_digest": "482b4cabbd31f60b18d71d5d709ae725e85c1a57f5736fdee8a40b4677546a4e", "world_id": "env_blobfish_benchmarks_repobench_software_e_0af840ee", "world_repository_url": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/repobench-500"}
+{"benchmark_name": "RepoBench-500", "benchmark_slug": "repobench-500", "companion_relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks.", "difficulty_disclosure": "Generator label only; no empirical model difficulty claim is published.", "domain": "Software engineering", "generator_difficulty_label": "medium", "instruction": "A flaky bot posted Monday's standup note twice on service 9001 (note_id 9501 and 9502). Delete the duplicate 9502 and keep 9501.", "reference_passed": true, "reference_steps": 2, "required_tools": ["search_standup_notes", "delete_standup_note"], "seeded_random_passed": false, "source_dataset_id": "swe-bench/swe-bench-verified", "source_dataset_url": "https://hub.harborframework.com/datasets/swe-bench/swe-bench-verified", "split": "heldout", "task_id": "eng_dedupe_standup_note", "verifier_basis": "vcode", "world_digest": "482b4cabbd31f60b18d71d5d709ae725e85c1a57f5736fdee8a40b4677546a4e", "world_id": "env_blobfish_benchmarks_repobench_software_e_0af840ee", "world_repository_url": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/repobench-500"}

data/trajectories.jsonl

@@ -0,0 +1,12 @@
+{"benchmark_slug": "filingbench-50", "disclosure": "A checked-in reference trace replayed against a fresh SQLite copy. This proves the task can pass; it is not a model submission.", "domain": "Financial research", "kind": "deterministic solvability ceiling", "label": "Reference solver", "model_id": null, "status": "passed", "step_count": 4, "steps_json": "[{\"arguments\": {\"company_id\": 9001, \"limit\": 50}, \"error\": null, \"index\": 1, \"ok\": true, \"output\": [{\"capex\": 60.0, \"company_id\": 9001, \"ebitda\": 200.0, \"free_cash_flow\": 100.0, \"period_id\": 9101, \"period_label\": \"2024\", \"revenue\": 800.0}, {\"capex\": 65.0, \"company_id\": 9001, \"ebitda\": 230.0, \"free_cash_flow\": 120.0, \"period_id\": 9102, \"period_label\": \"2025\", \"revenue\": 900.0}, {\"capex\": 70.0, \"company_id\": 9001, \"ebitda\": 260.0, \"free_cash_flow\": 140.0, \"period_id\": 9103, \"period_label\": \"2026\", \"revenue\": 1000.0}], \"tool\": \"search_financial_periods\", \"type\": \"tool_call\"}, {\"arguments\": {\"enterprise_value\": 360.0, \"model_id\": 9201, \"status\": \"final\"}, \"error\": null, \"index\": 2, \"ok\": true, \"output\": {\"company_id\": 9001, \"discount_rate\": 0.09, \"enterprise_value\": 360.0, \"model_id\": 9201, \"name\": \"Solstice base case\", \"status\": \"final\", \"terminal_growth\": 0.02}, \"tool\": \"update_dcf_model\", \"type\": \"tool_call\"}, {\"arguments\": {\"assumption_id\": 9301, \"status\": \"verified\"}, \"error\": null, \"index\": 3, \"ok\": true, \"output\": {\"assumption_id\": 9301, \"model_id\": 9201, \"name\": \"WACC\", \"source\": \"Capital IQ\", \"status\": \"verified\", \"value\": 0.09}, \"tool\": \"update_dcf_assumption\", \"type\": \"tool_call\"}, {\"arguments\": {\"body\": \"Unlevered free cash flow across the period set totals 360.\", \"model_id\": 9201, \"reviewer_name\": \"Head of Research\", \"status\": \"open\"}, \"error\": null, \"index\": 4, \"ok\": true, \"output\": {\"body\": \"Unlevered free cash flow across the period set totals 360.\", \"model_id\": 9201, \"note_id\": 9502, \"reviewer_name\": \"Head of Research\", \"status\": \"open\"}, \"tool\": \"create_review_note\", \"type\": \"tool_call\"}]", "task_id": "fa_finalize_solstice_dcf", "trajectory_id": "reference", "verifier_json": "{\"basis\": \"vcode\", \"gradable\": true, \"passed\": true, \"reward\": 1.0}"}
+{"benchmark_slug": "filingbench-50", "disclosure": "A seeded random tool policy against the same task and fresh world. It is a discrimination floor, not an agent or model result.", "domain": "Financial research", "kind": "non-model failure floor", "label": "Seeded random floor", "model_id": null, "status": "failed", "step_count": 6, "steps_json": "[{\"arguments\": {\"company_id\": 4, \"discount_rate\": 67.54, \"enterprise_value\": 8.06, \"name\": \"pending\", \"status\": \"test\", \"terminal_growth\": 78.06}, \"error\": \"CHECK constraint failed: status\", \"index\": 1, \"ok\": false, \"output\": null, \"tool\": \"create_dcf_model\", \"type\": \"tool_call\"}, {\"arguments\": {\"model_id\": 9, \"name\": \"active\", \"source\": \"pending\", \"status\": \"active\", \"value\": 57.89}, \"error\": null, \"index\": 2, \"ok\": true, \"output\": {\"assumption_id\": 9313, \"model_id\": 9, \"name\": \"active\", \"source\": \"pending\", \"status\": \"active\", \"value\": 57.89}, \"tool\": \"create_dcf_assumption\", \"type\": \"tool_call\"}, {\"arguments\": {\"company_id\": 19, \"name\": \"test\", \"sector\": \"test\", \"ticker\": \"42\"}, \"error\": null, \"index\": 3, \"ok\": true, \"output\": 0, \"tool\": \"count_fa_companies\", \"type\": \"tool_call\"}, {\"arguments\": {\"note_id\": 16}, \"error\": null, \"index\": 4, \"ok\": true, \"output\": {\"body\": \"{\\\"checksum\\\": \\\"042d186d\\\", \\\"received_via\\\": \\\"manual_entry\\\", \\\"record_type\\\": \\\"review_note\\\", \\\"reference\\\": \\\"RN-001004\\\", \\\"source\\\": \\\"web_portal\\\", \\\"version\\\": 1}\", \"model_id\": 10, \"note_id\": 16, \"reviewer_name\": \"Standard Internal Review Note\", \"status\": \"closed\"}, \"tool\": \"get_review_note\", \"type\": \"tool_call\"}, {\"arguments\": {\"company_id\": 7, \"discount_rate\": 39.62, \"enterprise_value\": 91.65, \"limit\": 19, \"name\": \"active\", \"query\": \"active\", \"status\": \"pending\", \"terminal_growth\": 89.02}, \"error\": null, \"index\": 5, \"ok\": true, \"output\": [], \"tool\": \"search_dcf_models\", \"type\": \"tool_call\"}, {\"arguments\": {\"limit\": 19, \"model_id\": 6}, \"error\": null, \"index\": 6, \"ok\": true, \"output\": [{\"assumption_id\": 7, \"model_id\": 6, \"name\": \"Quarterly Regional Dcf Assumption\", \"source\": \"General Source F92\", \"status\": \"unverified\", \"value\": 83.46}], \"tool\": \"list_dcf_assumptions_for_dcf_model\", \"type\": \"tool_call\"}]", "task_id": "fa_finalize_solstice_dcf", "trajectory_id": "random", "verifier_json": "{\"basis\": \"vcode\", \"gradable\": true, \"passed\": false, \"reward\": 0.0}"}
+{"benchmark_slug": "servicebench-375", "disclosure": "A checked-in reference trace replayed against a fresh SQLite copy. This proves the task can pass; it is not a model submission.", "domain": "Customer operations", "kind": "deterministic solvability ceiling", "label": "Reference solver", "model_id": null, "status": "passed", "step_count": 4, "steps_json": "[{\"arguments\": {\"limit\": 10, \"plan\": \"enterprise\"}, \"error\": null, \"index\": 1, \"ok\": true, \"output\": [{\"first_response_minutes\": 30, \"name\": \"Enterprise SLA\", \"plan\": \"enterprise\", \"policy_id\": 9101, \"resolve_minutes\": 240}], \"tool\": \"search_sla_policies\", \"type\": \"tool_call\"}, {\"arguments\": {\"priority\": \"urgent\", \"status\": \"escalated\", \"ticket_id\": 9201}, \"error\": null, \"index\": 2, \"ok\": true, \"output\": {\"customer_id\": 9001, \"first_response_minute\": null, \"opened_minute\": 100, \"priority\": \"urgent\", \"resolved_minute\": null, \"status\": \"escalated\", \"subject\": \"Fleet controller offline\", \"ticket_id\": 9201}, \"tool\": \"update_ticket\", \"type\": \"tool_call\"}, {\"arguments\": {\"reason\": \"sla_first_response_breach\", \"status\": \"open\", \"ticket_id\": 9201, \"tier\": \"tier2\"}, \"error\": null, \"index\": 3, \"ok\": true, \"output\": {\"escalation_id\": 9402, \"reason\": \"sla_first_response_breach\", \"status\": \"open\", \"ticket_id\": 9201, \"tier\": \"tier2\"}, \"tool\": \"create_escalation\", \"type\": \"tool_call\"}, {\"arguments\": {\"author\": \"agent\", \"body\": \"Apologies — our first response is 50 minutes past the enterprise SLA. An engineer is now on this.\", \"channel\": \"email\", \"sent_minute\": 180, \"ticket_id\": 9201}, \"error\": null, \"index\": 4, \"ok\": true, \"output\": {\"author\": \"agent\", \"body\": \"Apologies — our first response is 50 minutes past the enterprise SLA. An engineer is now on this.\", \"channel\": \"email\", \"message_id\": 9303, \"sent_minute\": 180, \"ticket_id\": 9201}, \"tool\": \"create_ticket_message\", \"type\": \"tool_call\"}]", "task_id": "cs_escalate_sla_breach", "trajectory_id": "reference", "verifier_json": "{\"basis\": \"vcode\", \"gradable\": true, \"passed\": true, \"reward\": 1.0}"}
+{"benchmark_slug": "servicebench-375", "disclosure": "A seeded random tool policy against the same task and fresh world. It is a discrimination floor, not an agent or model result.", "domain": "Customer operations", "kind": "non-model failure floor", "label": "Seeded random floor", "model_id": null, "status": "failed", "step_count": 6, "steps_json": "[{\"arguments\": {\"name\": \"test\", \"plan\": \"test\", \"region\": \"active\"}, \"error\": null, \"index\": 1, \"ok\": true, \"output\": {\"customer_id\": 9004, \"name\": \"test\", \"plan\": \"test\", \"region\": \"active\"}, \"tool\": \"create_support_customer\", \"type\": \"tool_call\"}, {\"arguments\": {\"first_response_minutes\": 12, \"limit\": 11, \"name\": \"pending\", \"plan\": \"42\", \"query\": \"42\", \"resolve_minutes\": 4}, \"error\": null, \"index\": 2, \"ok\": true, \"output\": [], \"tool\": \"search_sla_policies\", \"type\": \"tool_call\"}, {\"arguments\": {\"first_response_minutes\": 20, \"name\": \"test\", \"plan\": \"42\", \"policy_id\": 14, \"resolve_minutes\": 1}, \"error\": null, \"index\": 3, \"ok\": true, \"output\": 0, \"tool\": \"count_sla_policies\", \"type\": \"tool_call\"}, {\"arguments\": {\"escalation_id\": 14, \"reason\": \"test\", \"status\": \"test\", \"ticket_id\": 7, \"tier\": \"active\"}, \"error\": null, \"index\": 4, \"ok\": true, \"output\": 0, \"tool\": \"count_escalations\", \"type\": \"tool_call\"}, {\"arguments\": {\"escalation_id\": 6}, \"error\": null, \"index\": 5, \"ok\": true, \"output\": true, \"tool\": \"delete_escalation\", \"type\": \"tool_call\"}, {\"arguments\": {\"first_response_minutes\": 5, \"name\": \"pending\", \"plan\": \"42\", \"policy_id\": 6, \"resolve_minutes\": 13}, \"error\": null, \"index\": 6, \"ok\": true, \"output\": 0, \"tool\": \"count_sla_policies\", \"type\": \"tool_call\"}]", "task_id": "cs_escalate_sla_breach", "trajectory_id": "random", "verifier_json": "{\"basis\": \"vcode\", \"gradable\": true, \"passed\": false, \"reward\": 0.0}"}
+{"benchmark_slug": "discoverybench-102", "disclosure": "A checked-in reference trace replayed against a fresh SQLite copy. This proves the task can pass; it is not a model submission.", "domain": "Scientific discovery", "kind": "deterministic solvability ceiling", "label": "Reference solver", "model_id": null, "status": "passed", "step_count": 4, "steps_json": "[{\"arguments\": {\"limit\": 50, \"run_id\": 9301}, \"error\": null, \"index\": 1, \"ok\": true, \"output\": [{\"ic50_nm\": 20.0, \"percent_inhibition\": 91.0, \"qc_flag\": \"pass\", \"replicate\": 1, \"result_id\": 9401, \"run_id\": 9301}, {\"ic50_nm\": 24.0, \"percent_inhibition\": 89.0, \"qc_flag\": \"pass\", \"replicate\": 2, \"result_id\": 9402, \"run_id\": 9301}, {\"ic50_nm\": 22.0, \"percent_inhibition\": 90.0, \"qc_flag\": \"pass\", \"replicate\": 3, \"result_id\": 9403, \"run_id\": 9301}, {\"ic50_nm\": 400.0, \"percent_inhibition\": 11.0, \"qc_flag\": \"fail\", \"replicate\": 4, \"result_id\": 9404, \"run_id\": 9301}], \"tool\": \"search_assay_results\", \"type\": \"tool_call\"}, {\"arguments\": {\"compound_id\": 9101, \"screening_note\": \"Mean QC-passing IC50 22 nM on run 9301.\", \"status\": \"hit\"}, \"error\": null, \"index\": 2, \"ok\": true, \"output\": {\"code\": \"BF-1042\", \"compound_id\": 9101, \"molecular_weight\": 421.5, \"name\": \"Velatinib\", \"screening_note\": \"Mean QC-passing IC50 22 nM on run 9301.\", \"status\": \"hit\"}, \"tool\": \"update_compound\", \"type\": \"tool_call\"}, {\"arguments\": {\"run_id\": 9301, \"status\": \"complete\"}, \"error\": null, \"index\": 3, \"ok\": true, \"output\": {\"assay_id\": 9201, \"compound_id\": 9101, \"plate_label\": \"PLT-0091\", \"run_id\": 9301, \"run_on\": \"2026-07-14\", \"status\": \"complete\"}, \"tool\": \"update_assay_run\", \"type\": \"tool_call\"}, {\"arguments\": {\"assay_id\": 9201, \"compound_id\": 9101, \"plate_label\": \"CONF-001\", \"run_on\": \"2026-08-01\", \"status\": \"pending\"}, \"error\": null, \"index\": 4, \"ok\": true, \"output\": {\"assay_id\": 9201, \"compound_id\": 9101, \"plate_label\": \"CONF-001\", \"run_id\": 9303, \"run_on\": \"2026-08-01\", \"status\": \"pending\"}, \"tool\": \"create_assay_run\", \"type\": \"tool_call\"}]", "task_id": "bio_promote_hit_compound", "trajectory_id": "reference", "verifier_json": "{\"basis\": \"vcode\", \"gradable\": true, \"passed\": true, \"reward\": 1.0}"}
+{"benchmark_slug": "discoverybench-102", "disclosure": "A seeded random tool policy against the same task and fresh world. It is a discrimination floor, not an agent or model result.", "domain": "Scientific discovery", "kind": "non-model failure floor", "label": "Seeded random floor", "model_id": null, "status": "failed", "step_count": 6, "steps_json": "[{\"arguments\": {\"assay_id\": 13, \"compound_id\": 16, \"plate_label\": \"active\", \"run_id\": 6, \"run_on\": \"pending\", \"status\": \"pending\"}, \"error\": null, \"index\": 1, \"ok\": true, \"output\": 0, \"tool\": \"count_assay_runs\", \"type\": \"tool_call\"}, {\"arguments\": {\"ref_id\": 20}, \"error\": null, \"index\": 2, \"ok\": true, \"output\": null, \"tool\": \"get_literature_ref\", \"type\": \"tool_call\"}, {\"arguments\": {\"citation\": \"42\", \"doi\": \"active\", \"relevance\": \"active\", \"target_id\": 7}, \"error\": \"CHECK constraint failed: relevance\", \"index\": 3, \"ok\": false, \"output\": null, \"tool\": \"create_literature_ref\", \"type\": \"tool_call\"}, {\"arguments\": {\"assay_id\": 3}, \"error\": null, \"index\": 4, \"ok\": true, \"output\": {\"assay_id\": 3, \"assay_type\": \"premium\", \"name\": \"Priority Retail Assay\", \"readout\": \"General Readout B61\", \"status\": \"suspended\", \"target_id\": 2}, \"tool\": \"get_assay\", \"type\": \"tool_call\"}, {\"arguments\": {\"assay_id\": 2}, \"error\": null, \"index\": 5, \"ok\": true, \"output\": true, \"tool\": \"delete_assay\", \"type\": \"tool_call\"}, {\"arguments\": {\"code\": \"test\", \"compound_id\": 16, \"molecular_weight\": 77.86, \"name\": \"pending\", \"screening_note\": \"active\", \"status\": \"active\"}, \"error\": null, \"index\": 6, \"ok\": true, \"output\": 0, \"tool\": \"count_compounds\", \"type\": \"tool_call\"}]", "task_id": "bio_promote_hit_compound", "trajectory_id": "random", "verifier_json": "{\"basis\": \"vcode\", \"gradable\": true, \"passed\": false, \"reward\": 0.0}"}
+{"benchmark_slug": "defensebench-100", "disclosure": "A checked-in reference trace replayed against a fresh SQLite copy. This proves the task can pass; it is not a model submission.", "domain": "Defensive security", "kind": "deterministic solvability ceiling", "label": "Reference solver", "model_id": null, "status": "passed", "step_count": 5, "steps_json": "[{\"arguments\": {\"document_id\": 9201, \"limit\": 50}, \"error\": null, \"index\": 1, \"ok\": true, \"output\": [{\"document_id\": 9201, \"employee_id\": 9101, \"grant_id\": 9401, \"level\": \"read\", \"status\": \"active\"}, {\"document_id\": 9201, \"employee_id\": 9102, \"grant_id\": 9402, \"level\": \"read\", \"status\": \"active\"}, {\"document_id\": 9201, \"employee_id\": 9103, \"grant_id\": 9403, \"level\": \"read\", \"status\": \"active\"}, {\"document_id\": 9201, \"employee_id\": 9104, \"grant_id\": 9404, \"level\": \"read\", \"status\": \"revoked\"}], \"tool\": \"search_access_grants\", \"type\": \"tool_call\"}, {\"arguments\": {\"limit\": 50}, \"error\": null, \"index\": 2, \"ok\": true, \"output\": [{\"email\": \"elizabethlee@gmail.com\", \"employee_id\": 1, \"name\": \"Standard Internal Employee\", \"team\": \"Standard Team E80\"}, {\"email\": \"john_jackson713@work.org\", \"employee_id\": 2, \"name\": \"Recurring Internal Employee\", \"team\": \"Consolidated Team C88\"}, {\"email\": \"jessicajackson@gmail.com\", \"employee_id\": 3, \"name\": \"Consolidated Partner Employee\", \"team\": \"Consolidated Team A59\"}, {\"email\": \"elizabeth_gonzalez83@work.org\", \"employee_id\": 4, \"name\": \"Standard Domestic Employee\", \"team\": \"Preferred Team A46\"}, {\"email\": \"barbara.martinez606@work.org\", \"employee_id\": 5, \"name\": \"Quarterly Regional Employee\", \"team\": \"Backup Team A22\"}, {\"email\": \"lisawilson@company.com\", \"employee_id\": 6, \"name\": \"Standard Partner Employee\", \"team\": \"General Team E78\"}, {\"email\": \"sandra_thompson1@outlook.com\", \"employee_id\": 7, \"name\": \"Recurring Internal Employee\", \"team\": \"General Team D87\"}, {\"email\": \"kennethwalker478@work.org\", \"employee_id\": 8, \"name\": \"Quarterly Regional Employee\", \"team\": \"Backup Team A56\"}, {\"email\": \"karen_gonzalez@outlook.com\", \"employee_id\": 9, \"name\": \"Expedited Partner Employee\", \"team\": \"Standard Team C17\"}, {\"email\": \"sarah_wilson@outlook.com\", \"employee_id\": 10, \"name\": \"Consolidated Retail Employee\", \"team\": \"Backup Team B44\"}, {\"email\": \"betty.smith@outlook.com\", \"employee_id\": 11, \"name\": \"Expedited Partner Employee\", \"team\": \"Preferred Team A60\"}, {\"email\": \"sandra.martin63@work.org\", \"employee_id\": 12, \"name\": \"Consolidated Partner Employee\", \"team\": \"Standard Team D35\"}, {\"email\": \"karenperez@outlook.com\", \"employee_id\": 13, \"name\": \"Expedited Export Employee\", \"team\": \"Preferred Team B21\"}, {\"email\": \"joshua.smith@company.com\", \"employee_id\": 14, \"name\": \"Expedited Export Employee\", \"team\": \"Preferred Team C80\"}, {\"email\": \"jessicarobinson@outlook.com\", \"employee_id\": 15, \"name\": \"Annual Export Employee\", \"team\": \"Backup Team F96\"}, {\"email\": \"lisa_wilson@work.org\", \"employee_id\": 16, \"name\": \"Quarterly Partner Employee\", \"team\": \"Secondary Team F88\"}, {\"email\": \"david_brown782@yahoo.com\", \"employee_id\": 17, \"name\": \"Recurring Domestic Employee\", \"team\": \"Standard Team C36\"}, {\"email\": \"ines@acme.example.com\", \"employee_id\": 9101, \"name\": \"Ines Bakker\", \"team\": \"Finance\"}, {\"email\": \"rowan.f@acme.example.com\", \"employee_id\": 9102, \"name\": \"Rowan Ferreira\", \"team\": \"Sales\"}, {\"email\": \"r.ferreira@acme.example.com\", \"employee_id\": 9103, \"name\": \"Rowan Ferreira\", \"team\": \"Engineering\"}, {\"email\": \"tobias@acme.example.com\", \"employee_id\": 9104, \"name\": \"Tobias Lund\", \"team\": \"Finance\"}], \"tool\": \"search_employees\", \"type\": \"tool_call\"}, {\"arguments\": {\"grant_id\": 9402, \"status\": \"revoked\"}, \"error\": null, \"index\": 3, \"ok\": true, \"output\": {\"document_id\": 9201, \"employee_id\": 9102, \"grant_id\": 9402, \"level\": \"read\", \"status\": \"revoked\"}, \"tool\": \"update_access_grant\", \"type\": \"tool_call\"}, {\"arguments\": {\"grant_id\": 9403, \"status\": \"revoked\"}, \"error\": null, \"index\": 4, \"ok\": true, \"output\": {\"document_id\": 9201, \"employee_id\": 9103, \"grant_id\": 9403, \"level\": \"read\", \"status\": \"revoked\"}, \"tool\": \"update_access_grant\", \"type\": \"tool_call\"}, {\"arguments\": {\"document_id\": 9201, \"status\": \"restricted\"}, \"error\": null, \"index\": 5, \"ok\": true, \"output\": {\"document_id\": 9201, \"owner_email\": \"ines@acme.example.com\", \"sensitivity\": \"confidential\", \"source_id\": 9001, \"status\": \"restricted\", \"title\": \"FY27 Budget Model\", \"updated_on\": \"2026-07-10\"}, \"tool\": \"update_indexed_document\", \"type\": \"tool_call\"}]", "task_id": "es_lock_down_confidential_document", "trajectory_id": "reference", "verifier_json": "{\"basis\": \"vcode\", \"gradable\": true, \"passed\": true, \"reward\": 1.0}"}
+{"benchmark_slug": "defensebench-100", "disclosure": "A seeded random tool policy against the same task and fresh world. It is a discrimination floor, not an agent or model result.", "domain": "Defensive security", "kind": "non-model failure floor", "label": "Seeded random floor", "model_id": null, "status": "failed", "step_count": 6, "steps_json": "[{\"arguments\": {\"connected\": 8, \"kind\": \"test\", \"last_synced_on\": \"test\", \"name\": \"pending\", \"source_id\": 3}, \"error\": null, \"index\": 1, \"ok\": true, \"output\": {\"connected\": 8, \"kind\": \"test\", \"last_synced_on\": \"test\", \"name\": \"pending\", \"source_id\": 3}, \"tool\": \"update_source\", \"type\": \"tool_call\"}, {\"arguments\": {\"employee_id\": 15, \"limit\": 10}, \"error\": null, \"index\": 2, \"ok\": true, \"output\": [], \"tool\": \"list_access_grants_for_employee\", \"type\": \"tool_call\"}, {\"arguments\": {\"limit\": 6, \"source_id\": 15}, \"error\": null, \"index\": 3, \"ok\": true, \"output\": [{\"document_id\": 16, \"owner_email\": \"john.lewis377@gmail.com\", \"sensitivity\": \"internal\", \"source_id\": 15, \"status\": \"indexed\", \"title\": \"Standard Regional Indexed Document\", \"updated_on\": \"2025-03-07\"}], \"tool\": \"list_indexed_documents_for_source\", \"type\": \"tool_call\"}, {\"arguments\": {\"limit\": 18, \"owner_email\": \"pending\", \"query\": \"active\", \"sensitivity\": \"pending\", \"source_id\": 8, \"status\": \"active\", \"title\": \"active\", \"updated_on\": \"active\"}, \"error\": null, \"index\": 4, \"ok\": true, \"output\": [], \"tool\": \"search_indexed_documents\", \"type\": \"tool_call\"}, {\"arguments\": {\"employee_id\": 14, \"limit\": 13, \"query\": \"active\", \"query_text\": \"pending\", \"ran_on\": \"test\", \"result_count\": 17}, \"error\": null, \"index\": 5, \"ok\": true, \"output\": [], \"tool\": \"search_search_queries\", \"type\": \"tool_call\"}, {\"arguments\": {\"connected\": 5, \"kind\": \"active\", \"last_synced_on\": \"test\", \"name\": \"pending\", \"source_id\": 20}, \"error\": null, \"index\": 6, \"ok\": true, \"output\": 0, \"tool\": \"count_sources\", \"type\": \"tool_call\"}]", "task_id": "es_lock_down_confidential_document", "trajectory_id": "random", "verifier_json": "{\"basis\": \"vcode\", \"gradable\": true, \"passed\": false, \"reward\": 0.0}"}
+{"benchmark_slug": "opsbench-14", "disclosure": "A checked-in reference trace replayed against a fresh SQLite copy. This proves the task can pass; it is not a model submission.", "domain": "Enterprise operations", "kind": "deterministic solvability ceiling", "label": "Reference solver", "model_id": null, "status": "passed", "step_count": 4, "steps_json": "[{\"arguments\": {\"doc_id\": 9101, \"limit\": 50}, \"error\": null, \"index\": 1, \"ok\": true, \"output\": [{\"control_ref\": \"CTRL-014\", \"description\": \"Collect vendor tax documentation\", \"doc_id\": 9101, \"step_id\": 9201, \"step_number\": 1}, {\"control_ref\": \"CTRL-014\", \"description\": \"Collect vendor tax documentation\", \"doc_id\": 9101, \"step_id\": 9202, \"step_number\": 1}, {\"control_ref\": \"CTRL-021\", \"description\": \"Run sanctions screening\", \"doc_id\": 9101, \"step_id\": 9203, \"step_number\": 2}, {\"control_ref\": \"\", \"description\": \"Grant system access\", \"doc_id\": 9101, \"step_id\": 9204, \"step_number\": 3}], \"tool\": \"search_process_steps\", \"type\": \"tool_call\"}, {\"arguments\": {\"request_id\": 9301, \"status\": \"rejected\"}, \"error\": null, \"index\": 2, \"ok\": true, \"output\": {\"doc_id\": 9101, \"request_id\": 9301, \"requested_by\": \"Riley Garcia\", \"risk_level\": \"high\", \"status\": \"rejected\", \"title\": \"Skip sanctions screening for tier3\"}, \"tool\": \"update_change_request\", \"type\": \"tool_call\"}, {\"arguments\": {\"doc_id\": 9101, \"status\": \"needs_controls\"}, \"error\": null, \"index\": 3, \"ok\": true, \"output\": {\"doc_id\": 9101, \"name\": \"Vendor Onboarding\", \"owner_name\": \"Jordan Patel\", \"status\": \"needs_controls\", \"version\": 3}, \"tool\": \"update_process_doc\", \"type\": \"tool_call\"}, {\"arguments\": {\"framework\": \"SOC2\", \"last_tested_on\": \"\", \"name\": \"Step control backfill\", \"status\": \"untested\"}, \"error\": null, \"index\": 4, \"ok\": true, \"output\": {\"control_id\": 9402, \"framework\": \"SOC2\", \"last_tested_on\": \"\", \"name\": \"Step control backfill\", \"status\": \"untested\"}, \"tool\": \"create_compliance_control\", \"type\": \"tool_call\"}]", "task_id": "ops_reject_uncontrolled_change", "trajectory_id": "reference", "verifier_json": "{\"basis\": \"vcode\", \"gradable\": true, \"passed\": true, \"reward\": 1.0}"}
+{"benchmark_slug": "opsbench-14", "disclosure": "A seeded random tool policy against the same task and fresh world. It is a discrimination floor, not an agent or model result.", "domain": "Enterprise operations", "kind": "non-model failure floor", "label": "Seeded random floor", "model_id": null, "status": "failed", "step_count": 6, "steps_json": "[{\"arguments\": {\"control_ref\": \"active\", \"description\": \"pending\", \"doc_id\": 2, \"limit\": 11, \"query\": \"active\", \"step_number\": 6}, \"error\": null, \"index\": 1, \"ok\": true, \"output\": [], \"tool\": \"search_process_steps\", \"type\": \"tool_call\"}, {\"arguments\": {\"doc_id\": 16, \"request_id\": 14, \"requested_by\": \"pending\", \"risk_level\": \"42\", \"status\": \"active\", \"title\": \"pending\"}, \"error\": null, \"index\": 2, \"ok\": true, \"output\": 0, \"tool\": \"count_change_requests\", \"type\": \"tool_call\"}, {\"arguments\": {\"category\": \"pending\", \"name\": \"test\", \"renewal_on\": \"pending\", \"status\": \"42\", \"tier\": \"42\", \"vendor_id\": 19}, \"error\": null, \"index\": 3, \"ok\": true, \"output\": 0, \"tool\": \"count_vendors\", \"type\": \"tool_call\"}, {\"arguments\": {\"control_ref\": \"test\", \"description\": \"42\", \"doc_id\": 4, \"step_number\": 20}, \"error\": null, \"index\": 4, \"ok\": true, \"output\": {\"control_ref\": \"test\", \"description\": \"42\", \"doc_id\": 4, \"step_id\": 9212, \"step_number\": 20}, \"tool\": \"create_process_step\", \"type\": \"tool_call\"}, {\"arguments\": {\"reviewed_on\": \"active\", \"reviewer_name\": \"42\", \"score\": 23.69, \"status\": \"pending\", \"vendor_id\": 9}, \"error\": \"CHECK constraint failed: status\", \"index\": 5, \"ok\": false, \"output\": null, \"tool\": \"create_vendor_review\", \"type\": \"tool_call\"}, {\"arguments\": {\"doc_id\": 18}, \"error\": null, \"index\": 6, \"ok\": true, \"output\": false, \"tool\": \"delete_process_doc\", \"type\": \"tool_call\"}]", "task_id": "ops_reject_uncontrolled_change", "trajectory_id": "random", "verifier_json": "{\"basis\": \"vcode\", \"gradable\": true, \"passed\": false, \"reward\": 0.0}"}
+{"benchmark_slug": "repobench-500", "disclosure": "A checked-in reference trace replayed against a fresh SQLite copy. This proves the task can pass; it is not a model submission.", "domain": "Software engineering", "kind": "deterministic solvability ceiling", "label": "Reference solver", "model_id": null, "status": "passed", "step_count": 4, "steps_json": "[{\"arguments\": {\"incident_id\": 9401}, \"error\": null, \"index\": 1, \"ok\": true, \"output\": {\"detected_minute\": 100, \"incident_id\": 9401, \"resolved_minute\": null, \"service_id\": 9001, \"severity\": \"sev1\", \"status\": \"open\", \"summary\": \"Ingest queue saturated, consumers stalled\"}, \"tool\": \"get_incident\", \"type\": \"tool_call\"}, {\"arguments\": {\"incident_id\": 9401, \"resolved_minute\": 175, \"status\": \"resolved\"}, \"error\": null, \"index\": 2, \"ok\": true, \"output\": {\"detected_minute\": 100, \"incident_id\": 9401, \"resolved_minute\": 175, \"service_id\": 9001, \"severity\": \"sev1\", \"status\": \"resolved\", \"summary\": \"Ingest queue saturated, consumers stalled\"}, \"tool\": \"update_incident\", \"type\": \"tool_call\"}, {\"arguments\": {\"decided_on\": \"2026-07-21\", \"decision\": \"Incident 9401 lasted 75 minutes; rate-limit producers to keep consumers from stalling.\", \"service_id\": 9001, \"status\": \"accepted\", \"title\": \"Rate-limit the ingest queue\"}, \"error\": null, \"index\": 3, \"ok\": true, \"output\": {\"adr_id\": 9302, \"decided_on\": \"2026-07-21\", \"decision\": \"Incident 9401 lasted 75 minutes; rate-limit producers to keep consumers from stalling.\", \"service_id\": 9001, \"status\": \"accepted\", \"title\": \"Rate-limit the ingest queue\"}, \"tool\": \"create_adr_record\", \"type\": \"tool_call\"}, {\"arguments\": {\"author\": \"Rowan Novak\", \"body\": \"Incident 9401 resolved; writing the rate-limit ADR follow-up.\", \"day_label\": \"tuesday\", \"service_id\": 9001}, \"error\": null, \"index\": 4, \"ok\": true, \"output\": {\"author\": \"Rowan Novak\", \"body\": \"Incident 9401 resolved; writing the rate-limit ADR follow-up.\", \"day_label\": \"tuesday\", \"note_id\": 9503, \"service_id\": 9001}, \"tool\": \"create_standup_note\", \"type\": \"tool_call\"}]", "task_id": "eng_resolve_ingest_incident", "trajectory_id": "reference", "verifier_json": "{\"basis\": \"vcode\", \"gradable\": true, \"passed\": true, \"reward\": 1.0}"}
+{"benchmark_slug": "repobench-500", "disclosure": "A seeded random tool policy against the same task and fresh world. It is a discrimination floor, not an agent or model result.", "domain": "Software engineering", "kind": "non-model failure floor", "label": "Seeded random floor", "model_id": null, "status": "failed", "step_count": 6, "steps_json": "[{\"arguments\": {\"default_branch\": \"42\", \"name\": \"42\", \"repo_id\": 20, \"service_id\": 14, \"visibility\": \"test\"}, \"error\": null, \"index\": 1, \"ok\": true, \"output\": 0, \"tool\": \"count_repositories\", \"type\": \"tool_call\"}, {\"arguments\": {\"detected_minute\": 6, \"incident_id\": 10, \"resolved_minute\": 10, \"service_id\": 19, \"severity\": \"pending\", \"status\": \"42\", \"summary\": \"active\"}, \"error\": null, \"index\": 2, \"ok\": true, \"output\": 0, \"tool\": \"count_incidents\", \"type\": \"tool_call\"}, {\"arguments\": {\"name\": \"42\", \"oncall_name\": \"pending\", \"owner_team\": \"42\", \"tier\": \"active\"}, \"error\": null, \"index\": 3, \"ok\": true, \"output\": {\"name\": \"42\", \"oncall_name\": \"pending\", \"owner_team\": \"42\", \"service_id\": 9003, \"tier\": \"active\"}, \"tool\": \"create_service\", \"type\": \"tool_call\"}, {\"arguments\": {\"note_id\": 8}, \"error\": null, \"index\": 4, \"ok\": true, \"output\": true, \"tool\": \"delete_standup_note\", \"type\": \"tool_call\"}, {\"arguments\": {\"limit\": 11, \"service_id\": 6}, \"error\": null, \"index\": 5, \"ok\": true, \"output\": [{\"adr_id\": 1, \"decided_on\": \"Regional Decided On E18\", \"decision\": \"Secondary Decision E10\", \"service_id\": 6, \"status\": \"placed\", \"title\": \"Consolidated Regional Adr Record\"}, {\"adr_id\": 3, \"decided_on\": \"Regional Decided On C71\", \"decision\": \"Backup Decision E75\", \"service_id\": 6, \"status\": \"accepted\", \"title\": \"Standard Partner Adr Record\"}], \"tool\": \"list_adr_records_for_service\", \"type\": \"tool_call\"}, {\"arguments\": {\"limit\": 10, \"service_id\": 20}, \"error\": null, \"index\": 6, \"ok\": true, \"output\": [], \"tool\": \"list_adr_records_for_service\", \"type\": \"tool_call\"}]", "task_id": "eng_resolve_ingest_incident", "trajectory_id": "random", "verifier_json": "{\"basis\": \"vcode\", \"gradable\": true, \"passed\": false, \"reward\": 0.0}"}

evidence/defensebench-100.json

@@ -0,0 +1,609 @@
+{
+  "schemaVersion": "blobfish.benchmark-evidence.v1",
+  "benchmarkSlug": "defensebench-100",
+  "observedAt": "2026-08-26",
+  "source": {
+    "datasetId": "polyvorlabs/cyberdefense-bench",
+    "tasks": 100,
+    "harborUrl": "https://hub.harborframework.com/datasets/polyvorlabs/cyberdefense-bench",
+    "inspection": "Harbor CLI 0.21.0 dataset metadata and package contract",
+    "relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks."
+  },
+  "world": {
+    "path": "benchmark/worlds/defensebench-100",
+    "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/defensebench-100",
+    "worldId": "env_blobfish_benchmarks_defensebench_access__d2a0878c",
+    "worldDigest": "e1df820514e93081e497aa1aa2fd50423e3dbf0cd766e4850838096a1bc641f6",
+    "vertical": "enterprise_search",
+    "tables": 6,
+    "rows": 121,
+    "tools": 42,
+    "tasks": 4,
+    "readOnlyTools": 23,
+    "mutatingTools": 19,
+    "entities": [
+      {
+        "name": "sources",
+        "rows": 20
+      },
+      {
+        "name": "employees",
+        "rows": 21
+      },
+      {
+        "name": "indexed_documents",
+        "rows": 20
+      },
+      {
+        "name": "document_chunks",
+        "rows": 20
+      },
+      {
+        "name": "access_grants",
+        "rows": 22
+      },
+      {
+        "name": "search_queries",
+        "rows": 18
+      }
+    ]
+  },
+  "evaluation": {
+    "contract": "Fresh database copy, recorded tool trace, deterministic VCode or exact answer verification",
+    "referenceRunId": "release-oracle",
+    "floorRunId": "release-random",
+    "baselines": [
+      {
+        "id": "reference",
+        "label": "Reference solver",
+        "kind": "Solvability ceiling",
+        "disclosure": "Deterministic trace replay; not a model rank.",
+        "tasksAttempted": 4,
+        "tasksPassed": 4,
+        "passRate": 1.0,
+        "meanReward": 1.0,
+        "toolCalls": 11,
+        "successfulToolCalls": 11,
+        "episodeErrors": 0
+      },
+      {
+        "id": "random",
+        "label": "Seeded random floor",
+        "kind": "Discrimination floor",
+        "disclosure": "Seeded random tool policy; not a model rank.",
+        "tasksAttempted": 4,
+        "tasksPassed": 0,
+        "passRate": 0.0,
+        "meanReward": 0.0,
+        "toolCalls": 24,
+        "successfulToolCalls": 23,
+        "episodeErrors": 0
+      }
+    ],
+    "claim": "All companion tasks are solvable by their reference traces and reject the seeded random floor. No paid or frontier model run is claimed."
+  },
+  "tasks": [
+    {
+      "taskId": "es_lock_down_confidential_document",
+      "instruction": "Document 9201 ('FY27 Budget Model') is confidential and may only be readable by the Finance team. Review its access grants: revoke every still-active grant held by an employee who is not on the Finance team by setting that grant's status to 'revoked', leave Finance grants and already-revoked grants alone, do not touch grants on other documents, and finally set document 9201's status to 'restricted'.",
+      "split": "train",
+      "generatorDifficultyLabel": "hard",
+      "difficultyDisclosure": "Generator label only; no empirical model difficulty claim is published.",
+      "requiredTools": [
+        "search_access_grants",
+        "search_employees",
+        "update_access_grant",
+        "update_indexed_document"
+      ],
+      "referenceSteps": 5,
+      "referencePassed": true,
+      "floorPassed": false,
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "vcode"
+      }
+    },
+    {
+      "taskId": "es_mark_disconnected_source_stale",
+      "instruction": "A connector has been disconnected. Every indexed document that came from a source with connected = 0 must be set to status 'stale'; documents from connected sources stay 'indexed'.",
+      "split": "train",
+      "generatorDifficultyLabel": "medium",
+      "difficultyDisclosure": "Generator label only; no empirical model difficulty claim is published.",
+      "requiredTools": [
+        "search_sources",
+        "search_indexed_documents",
+        "update_indexed_document"
+      ],
+      "referenceSteps": 3,
+      "referencePassed": true,
+      "floorPassed": false,
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "vcode"
+      }
+    },
+    {
+      "taskId": "es_duplicate_person_name",
+      "instruction": "Two different employees share a display name, which makes people search ambiguous. How many employees are named exactly 'Rowan Ferreira'? Answer with the number.",
+      "split": "train",
+      "generatorDifficultyLabel": "easy",
+      "difficultyDisclosure": "Generator label only; no empirical model difficulty claim is published.",
+      "requiredTools": [
+        "count_employees"
+      ],
+      "referenceSteps": 1,
+      "referencePassed": true,
+      "floorPassed": false,
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "answer_match (vcode does not discriminate)"
+      }
+    },
+    {
+      "taskId": "es_dedupe_document_chunk",
+      "instruction": "A double sync indexed the same passage of document 9201 twice (chunk_id 9301 and 9302), so it is over-weighted in retrieval. Delete the duplicate chunk 9302 and keep 9301.",
+      "split": "heldout",
+      "generatorDifficultyLabel": "medium",
+      "difficultyDisclosure": "Generator label only; no empirical model difficulty claim is published.",
+      "requiredTools": [
+        "search_document_chunks",
+        "delete_document_chunk"
+      ],
+      "referenceSteps": 2,
+      "referencePassed": true,
+      "floorPassed": false,
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "vcode"
+      }
+    }
+  ],
+  "assets": [
+    {
+      "path": "environment.db",
+      "label": "Seeded SQLite snapshot",
+      "kind": "SQLite",
+      "description": "Pristine state copied for every episode.",
+      "bytes": 32768,
+      "sha256": "4940300240cbe15d35d1fd45016398abe150ca2ee0361d04b2a17432f8a26233",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/defensebench-100/environment.db"
+    },
+    {
+      "path": "schema.sql",
+      "label": "Relational schema",
+      "kind": "SQL",
+      "description": "Human-readable table and constraint definitions.",
+      "bytes": 1251,
+      "sha256": "08c2e8abf194ee2764b390f01b9e8185c8653950c9582b3e2210e0f5f4a7b70b",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/defensebench-100/schema.sql"
+    },
+    {
+      "path": "seed.sql",
+      "label": "Seed data",
+      "kind": "SQL",
+      "description": "Human-readable source for the released world state.",
+      "bytes": 14152,
+      "sha256": "9c2b62927c80c430b0808605fbdefa36c200ef146cd539f6d02d6b1acb66573c",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/defensebench-100/seed.sql"
+    },
+    {
+      "path": "tools.json",
+      "label": "MCP tool catalog",
+      "kind": "JSON",
+      "description": "Public tool names, schemas, and implementations.",
+      "bytes": 90244,
+      "sha256": "46021a577b0d1e7e3d8c5befed8ac5f6ea9d3d84ce61fc8e34a446fe59b35c43",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/defensebench-100/tools.json"
+    },
+    {
+      "path": "tasks.json",
+      "label": "Task manifest",
+      "kind": "JSON",
+      "description": "Instructions, splits, reference traces, and isolated verifier programs.",
+      "bytes": 106978,
+      "sha256": "44ae2da2ce9ef5faccb06335c0a242eabcdd8e41626925131bb01665d1165aa5",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/defensebench-100/tasks.json"
+    },
+    {
+      "path": "mcp/manifest.json",
+      "label": "MCP manifest",
+      "kind": "JSON",
+      "description": "Stdio server declaration for the world.",
+      "bytes": 1969,
+      "sha256": "3c809b60c7ab96026e4ee0f4ac5ba0d04b011005465adff97185e0f98d1893ac",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/defensebench-100/mcp/manifest.json"
+    },
+    {
+      "path": "world.json",
+      "label": "World manifest",
+      "kind": "JSON",
+      "description": "Content identity, counts, split, and admission evidence.",
+      "bytes": 2540,
+      "sha256": "c8a86f80f071188223d082706306fb84984c1fbb6cce9341f4ab0f8e4a48ef8a",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/defensebench-100/world.json"
+    }
+  ],
+  "trajectories": [
+    {
+      "id": "reference",
+      "label": "Reference solver",
+      "kind": "deterministic solvability ceiling",
+      "modelId": null,
+      "taskId": "es_lock_down_confidential_document",
+      "status": "passed",
+      "steps": [
+        {
+          "index": 1,
+          "type": "tool_call",
+          "tool": "search_access_grants",
+          "arguments": {
+            "document_id": 9201,
+            "limit": 50
+          },
+          "ok": true,
+          "output": [
+            {
+              "grant_id": 9401,
+              "document_id": 9201,
+              "employee_id": 9101,
+              "level": "read",
+              "status": "active"
+            },
+            {
+              "grant_id": 9402,
+              "document_id": 9201,
+              "employee_id": 9102,
+              "level": "read",
+              "status": "active"
+            },
+            {
+              "grant_id": 9403,
+              "document_id": 9201,
+              "employee_id": 9103,
+              "level": "read",
+              "status": "active"
+            },
+            {
+              "grant_id": 9404,
+              "document_id": 9201,
+              "employee_id": 9104,
+              "level": "read",
+              "status": "revoked"
+            }
+          ],
+          "error": null
+        },
+        {
+          "index": 2,
+          "type": "tool_call",
+          "tool": "search_employees",
+          "arguments": {
+            "limit": 50
+          },
+          "ok": true,
+          "output": [
+            {
+              "employee_id": 1,
+              "name": "Standard Internal Employee",
+              "email": "elizabethlee@gmail.com",
+              "team": "Standard Team E80"
+            },
+            {
+              "employee_id": 2,
+              "name": "Recurring Internal Employee",
+              "email": "john_jackson713@work.org",
+              "team": "Consolidated Team C88"
+            },
+            {
+              "employee_id": 3,
+              "name": "Consolidated Partner Employee",
+              "email": "jessicajackson@gmail.com",
+              "team": "Consolidated Team A59"
+            },
+            {
+              "employee_id": 4,
+              "name": "Standard Domestic Employee",
+              "email": "elizabeth_gonzalez83@work.org",
+              "team": "Preferred Team A46"
+            },
+            {
+              "employee_id": 5,
+              "name": "Quarterly Regional Employee",
+              "email": "barbara.martinez606@work.org",
+              "team": "Backup Team A22"
+            },
+            {
+              "employee_id": 6,
+              "name": "Standard Partner Employee",
+              "email": "lisawilson@company.com",
+              "team": "General Team E78"
+            },
+            {
+              "employee_id": 7,
+              "name": "Recurring Internal Employee",
+              "email": "sandra_thompson1@outlook.com",
+              "team": "General Team D87"
+            },
+            {
+              "employee_id": 8,
+              "name": "Quarterly Regional Employee",
+              "email": "kennethwalker478@work.org",
+              "team": "Backup Team A56"
+            },
+            {
+              "employee_id": 9,
+              "name": "Expedited Partner Employee",
+              "email": "karen_gonzalez@outlook.com",
+              "team": "Standard Team C17"
+            },
+            {
+              "employee_id": 10,
+              "name": "Consolidated Retail Employee",
+              "email": "sarah_wilson@outlook.com",
+              "team": "Backup Team B44"
+            },
+            {
+              "employee_id": 11,
+              "name": "Expedited Partner Employee",
+              "email": "betty.smith@outlook.com",
+              "team": "Preferred Team A60"
+            },
+            {
+              "employee_id": 12,
+              "name": "Consolidated Partner Employee",
+              "email": "sandra.martin63@work.org",
+              "team": "Standard Team D35"
+            },
+            {
+              "employee_id": 13,
+              "name": "Expedited Export Employee",
+              "email": "karenperez@outlook.com",
+              "team": "Preferred Team B21"
+            },
+            {
+              "employee_id": 14,
+              "name": "Expedited Export Employee",
+              "email": "joshua.smith@company.com",
+              "team": "Preferred Team C80"
+            },
+            {
+              "employee_id": 15,
+              "name": "Annual Export Employee",
+              "email": "jessicarobinson@outlook.com",
+              "team": "Backup Team F96"
+            },
+            {
+              "employee_id": 16,
+              "name": "Quarterly Partner Employee",
+              "email": "lisa_wilson@work.org",
+              "team": "Secondary Team F88"
+            },
+            {
+              "employee_id": 17,
+              "name": "Recurring Domestic Employee",
+              "email": "david_brown782@yahoo.com",
+              "team": "Standard Team C36"
+            },
+            {
+              "employee_id": 9101,
+              "name": "Ines Bakker",
+              "email": "ines@acme.example.com",
+              "team": "Finance"
+            },
+            {
+              "employee_id": 9102,
+              "name": "Rowan Ferreira",
+              "email": "rowan.f@acme.example.com",
+              "team": "Sales"
+            },
+            {
+              "employee_id": 9103,
+              "name": "Rowan Ferreira",
+              "email": "r.ferreira@acme.example.com",
+              "team": "Engineering"
+            },
+            {
+              "employee_id": 9104,
+              "name": "Tobias Lund",
+              "email": "tobias@acme.example.com",
+              "team": "Finance"
+            }
+          ],
+          "error": null
+        },
+        {
+          "index": 3,
+          "type": "tool_call",
+          "tool": "update_access_grant",
+          "arguments": {
+            "grant_id": 9402,
+            "status": "revoked"
+          },
+          "ok": true,
+          "output": {
+            "grant_id": 9402,
+            "document_id": 9201,
+            "employee_id": 9102,
+            "level": "read",
+            "status": "revoked"
+          },
+          "error": null
+        },
+        {
+          "index": 4,
+          "type": "tool_call",
+          "tool": "update_access_grant",
+          "arguments": {
+            "grant_id": 9403,
+            "status": "revoked"
+          },
+          "ok": true,
+          "output": {
+            "grant_id": 9403,
+            "document_id": 9201,
+            "employee_id": 9103,
+            "level": "read",
+            "status": "revoked"
+          },
+          "error": null
+        },
+        {
+          "index": 5,
+          "type": "tool_call",
+          "tool": "update_indexed_document",
+          "arguments": {
+            "document_id": 9201,
+            "status": "restricted"
+          },
+          "ok": true,
+          "output": {
+            "document_id": 9201,
+            "source_id": 9001,
+            "title": "FY27 Budget Model",
+            "owner_email": "ines@acme.example.com",
+            "updated_on": "2026-07-10",
+            "sensitivity": "confidential",
+            "status": "restricted"
+          },
+          "error": null
+        }
+      ],
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "vcode"
+      },
+      "disclosure": "A checked-in reference trace replayed against a fresh SQLite copy. This proves the task can pass; it is not a model submission."
+    },
+    {
+      "id": "random",
+      "label": "Seeded random floor",
+      "kind": "non-model failure floor",
+      "modelId": null,
+      "taskId": "es_lock_down_confidential_document",
+      "status": "failed",
+      "steps": [
+        {
+          "index": 1,
+          "type": "tool_call",
+          "tool": "update_source",
+          "arguments": {
+            "source_id": 3,
+            "name": "pending",
+            "kind": "test",
+            "connected": 8,
+            "last_synced_on": "test"
+          },
+          "ok": true,
+          "output": {
+            "source_id": 3,
+            "name": "pending",
+            "kind": "test",
+            "connected": 8,
+            "last_synced_on": "test"
+          },
+          "error": null
+        },
+        {
+          "index": 2,
+          "type": "tool_call",
+          "tool": "list_access_grants_for_employee",
+          "arguments": {
+            "employee_id": 15,
+            "limit": 10
+          },
+          "ok": true,
+          "output": [],
+          "error": null
+        },
+        {
+          "index": 3,
+          "type": "tool_call",
+          "tool": "list_indexed_documents_for_source",
+          "arguments": {
+            "source_id": 15,
+            "limit": 6
+          },
+          "ok": true,
+          "output": [
+            {
+              "document_id": 16,
+              "source_id": 15,
+              "title": "Standard Regional Indexed Document",
+              "owner_email": "john.lewis377@gmail.com",
+              "updated_on": "2025-03-07",
+              "sensitivity": "internal",
+              "status": "indexed"
+            }
+          ],
+          "error": null
+        },
+        {
+          "index": 4,
+          "type": "tool_call",
+          "tool": "search_indexed_documents",
+          "arguments": {
+            "query": "active",
+            "limit": 18,
+            "source_id": 8,
+            "title": "active",
+            "owner_email": "pending",
+            "updated_on": "active",
+            "sensitivity": "pending",
+            "status": "active"
+          },
+          "ok": true,
+          "output": [],
+          "error": null
+        },
+        {
+          "index": 5,
+          "type": "tool_call",
+          "tool": "search_search_queries",
+          "arguments": {
+            "query": "active",
+            "limit": 13,
+            "employee_id": 14,
+            "query_text": "pending",
+            "ran_on": "test",
+            "result_count": 17
+          },
+          "ok": true,
+          "output": [],
+          "error": null
+        },
+        {
+          "index": 6,
+          "type": "tool_call",
+          "tool": "count_sources",
+          "arguments": {
+            "source_id": 20,
+            "name": "pending",
+            "kind": "active",
+            "connected": 5,
+            "last_synced_on": "test"
+          },
+          "ok": true,
+          "output": 0,
+          "error": null
+        }
+      ],
+      "verifier": {
+        "passed": false,
+        "reward": 0.0,
+        "gradable": true,
+        "basis": "vcode"
+      },
+      "disclosure": "A seeded random tool policy against the same task and fresh world. It is a discrimination floor, not an agent or model result."
+    }
+  ]
+}

evidence/discoverybench-102.json

@@ -0,0 +1,461 @@
+{
+  "schemaVersion": "blobfish.benchmark-evidence.v1",
+  "benchmarkSlug": "discoverybench-102",
+  "observedAt": "2026-08-26",
+  "source": {
+    "datasetId": "scienceagentbench/scienceagentbench",
+    "tasks": 102,
+    "harborUrl": "https://hub.harborframework.com/datasets/scienceagentbench/scienceagentbench",
+    "inspection": "Harbor CLI 0.21.0 dataset metadata and package contract",
+    "relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks."
+  },
+  "world": {
+    "path": "benchmark/worlds/discoverybench-102",
+    "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/discoverybench-102",
+    "worldId": "env_blobfish_benchmarks_discoverybench_scien_f7e1dbce",
+    "worldDigest": "475b17df4a7efe8265400a4d47889288764829150a4dc63e9b1188a06948cd29",
+    "vertical": "bio_research",
+    "tables": 6,
+    "rows": 119,
+    "tools": 43,
+    "tasks": 4,
+    "readOnlyTools": 23,
+    "mutatingTools": 20,
+    "entities": [
+      {
+        "name": "targets",
+        "rows": 19
+      },
+      {
+        "name": "compounds",
+        "rows": 20
+      },
+      {
+        "name": "assays",
+        "rows": 19
+      },
+      {
+        "name": "assay_runs",
+        "rows": 19
+      },
+      {
+        "name": "assay_results",
+        "rows": 24
+      },
+      {
+        "name": "literature_refs",
+        "rows": 18
+      }
+    ]
+  },
+  "evaluation": {
+    "contract": "Fresh database copy, recorded tool trace, deterministic VCode or exact answer verification",
+    "referenceRunId": "release-oracle",
+    "floorRunId": "release-random",
+    "baselines": [
+      {
+        "id": "reference",
+        "label": "Reference solver",
+        "kind": "Solvability ceiling",
+        "disclosure": "Deterministic trace replay; not a model rank.",
+        "tasksAttempted": 4,
+        "tasksPassed": 4,
+        "passRate": 1.0,
+        "meanReward": 1.0,
+        "toolCalls": 9,
+        "successfulToolCalls": 9,
+        "episodeErrors": 0
+      },
+      {
+        "id": "random",
+        "label": "Seeded random floor",
+        "kind": "Discrimination floor",
+        "disclosure": "Seeded random tool policy; not a model rank.",
+        "tasksAttempted": 4,
+        "tasksPassed": 0,
+        "passRate": 0.0,
+        "meanReward": 0.0,
+        "toolCalls": 24,
+        "successfulToolCalls": 20,
+        "episodeErrors": 0
+      }
+    ],
+    "claim": "All companion tasks are solvable by their reference traces and reject the seeded random floor. No paid or frontier model run is claimed."
+  },
+  "tasks": [
+    {
+      "taskId": "bio_promote_hit_compound",
+      "instruction": "Call the potency of assay run 9301. Compute the mean ic50_nm across its replicates that PASSED QC (qc_flag 'pass') — replicates flagged 'fail' are excluded. A mean below 100 nM makes the compound a hit, so: set compound 9101's status to 'hit' and record the mean in its screening_note, set assay run 9301's status to 'complete', and queue confirmation by creating an assay run on assay 9201 for compound 9101 with run_on '2026-08-01', plate_label 'CONF-001' and status 'pending'.",
+      "split": "train",
+      "generatorDifficultyLabel": "hard",
+      "difficultyDisclosure": "Generator label only; no empirical model difficulty claim is published.",
+      "requiredTools": [
+        "search_assay_results",
+        "update_compound",
+        "update_assay_run",
+        "create_assay_run"
+      ],
+      "referenceSteps": 4,
+      "referencePassed": true,
+      "floorPassed": false,
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "vcode"
+      }
+    },
+    {
+      "taskId": "bio_invalidate_failed_run",
+      "instruction": "Quality control: an assay run where EVERY replicate carries qc_flag 'fail' produced no usable data and must be set to status 'invalid'. Runs with at least one passing replicate are left alone.",
+      "split": "train",
+      "generatorDifficultyLabel": "medium",
+      "difficultyDisclosure": "Generator label only; no empirical model difficulty claim is published.",
+      "requiredTools": [
+        "search_assay_results",
+        "update_assay_run"
+      ],
+      "referenceSteps": 2,
+      "referencePassed": true,
+      "floorPassed": false,
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "vcode"
+      }
+    },
+    {
+      "taskId": "bio_duplicate_compound_name",
+      "instruction": "Two structurally distinct compounds were filed under the same common name. How many compounds are named exactly 'Velatinib'? Answer with the number.",
+      "split": "train",
+      "generatorDifficultyLabel": "easy",
+      "difficultyDisclosure": "Generator label only; no empirical model difficulty claim is published.",
+      "requiredTools": [
+        "count_compounds"
+      ],
+      "referenceSteps": 1,
+      "referencePassed": true,
+      "floorPassed": false,
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "answer_match (vcode does not discriminate)"
+      }
+    },
+    {
+      "taskId": "bio_dedupe_assay_result",
+      "instruction": "Run 9302 has replicate 1 recorded twice (result_id 9411 and 9412) because the plate reader exported twice. Delete the duplicate 9412 and keep 9411.",
+      "split": "heldout",
+      "generatorDifficultyLabel": "medium",
+      "difficultyDisclosure": "Generator label only; no empirical model difficulty claim is published.",
+      "requiredTools": [
+        "search_assay_results",
+        "delete_assay_result"
+      ],
+      "referenceSteps": 2,
+      "referencePassed": true,
+      "floorPassed": false,
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "vcode"
+      }
+    }
+  ],
+  "assets": [
+    {
+      "path": "environment.db",
+      "label": "Seeded SQLite snapshot",
+      "kind": "SQLite",
+      "description": "Pristine state copied for every episode.",
+      "bytes": 32768,
+      "sha256": "44b907f65127934c49e2569a0226486c470231e635d324cec1aa09ebf4a6df05",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/discoverybench-102/environment.db"
+    },
+    {
+      "path": "schema.sql",
+      "label": "Relational schema",
+      "kind": "SQL",
+      "description": "Human-readable table and constraint definitions.",
+      "bytes": 1497,
+      "sha256": "4dfc04a1e3d89b3839449644c07633a4bc2b86d65bc8d6d459ac0c3bb4989c83",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/discoverybench-102/schema.sql"
+    },
+    {
+      "path": "seed.sql",
+      "label": "Seed data",
+      "kind": "SQL",
+      "description": "Human-readable source for the released world state.",
+      "bytes": 11964,
+      "sha256": "e3e3c20b66558ba03fff8da8292f92b28b031e1f2dc7a2310e5e702cff3398bc",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/discoverybench-102/seed.sql"
+    },
+    {
+      "path": "tools.json",
+      "label": "MCP tool catalog",
+      "kind": "JSON",
+      "description": "Public tool names, schemas, and implementations.",
+      "bytes": 97373,
+      "sha256": "511ab9dc9a947171a0a70b0d75d1201d40298c285f217b904ba69ef631ae5e5a",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/discoverybench-102/tools.json"
+    },
+    {
+      "path": "tasks.json",
+      "label": "Task manifest",
+      "kind": "JSON",
+      "description": "Instructions, splits, reference traces, and isolated verifier programs.",
+      "bytes": 103570,
+      "sha256": "73b32ea4a6834a886e55ae6cf26412bec6eadff9e06959912dbf905ab98881aa",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/discoverybench-102/tasks.json"
+    },
+    {
+      "path": "mcp/manifest.json",
+      "label": "MCP manifest",
+      "kind": "JSON",
+      "description": "Stdio server declaration for the world.",
+      "bytes": 1862,
+      "sha256": "6dfe7eff3b010b684dadaac82edc606243c2582461bb691c512da2f3d7c074a4",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/discoverybench-102/mcp/manifest.json"
+    },
+    {
+      "path": "world.json",
+      "label": "World manifest",
+      "kind": "JSON",
+      "description": "Content identity, counts, split, and admission evidence.",
+      "bytes": 2492,
+      "sha256": "9788c5cbc28caa5c426e82ab59d220ad7d7aaccf913b4a77dc0e7a8e4f4c6f14",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/discoverybench-102/world.json"
+    }
+  ],
+  "trajectories": [
+    {
+      "id": "reference",
+      "label": "Reference solver",
+      "kind": "deterministic solvability ceiling",
+      "modelId": null,
+      "taskId": "bio_promote_hit_compound",
+      "status": "passed",
+      "steps": [
+        {
+          "index": 1,
+          "type": "tool_call",
+          "tool": "search_assay_results",
+          "arguments": {
+            "run_id": 9301,
+            "limit": 50
+          },
+          "ok": true,
+          "output": [
+            {
+              "result_id": 9401,
+              "run_id": 9301,
+              "replicate": 1,
+              "ic50_nm": 20.0,
+              "percent_inhibition": 91.0,
+              "qc_flag": "pass"
+            },
+            {
+              "result_id": 9402,
+              "run_id": 9301,
+              "replicate": 2,
+              "ic50_nm": 24.0,
+              "percent_inhibition": 89.0,
+              "qc_flag": "pass"
+            },
+            {
+              "result_id": 9403,
+              "run_id": 9301,
+              "replicate": 3,
+              "ic50_nm": 22.0,
+              "percent_inhibition": 90.0,
+              "qc_flag": "pass"
+            },
+            {
+              "result_id": 9404,
+              "run_id": 9301,
+              "replicate": 4,
+              "ic50_nm": 400.0,
+              "percent_inhibition": 11.0,
+              "qc_flag": "fail"
+            }
+          ],
+          "error": null
+        },
+        {
+          "index": 2,
+          "type": "tool_call",
+          "tool": "update_compound",
+          "arguments": {
+            "compound_id": 9101,
+            "status": "hit",
+            "screening_note": "Mean QC-passing IC50 22 nM on run 9301."
+          },
+          "ok": true,
+          "output": {
+            "compound_id": 9101,
+            "code": "BF-1042",
+            "name": "Velatinib",
+            "molecular_weight": 421.5,
+            "status": "hit",
+            "screening_note": "Mean QC-passing IC50 22 nM on run 9301."
+          },
+          "error": null
+        },
+        {
+          "index": 3,
+          "type": "tool_call",
+          "tool": "update_assay_run",
+          "arguments": {
+            "run_id": 9301,
+            "status": "complete"
+          },
+          "ok": true,
+          "output": {
+            "run_id": 9301,
+            "assay_id": 9201,
+            "compound_id": 9101,
+            "run_on": "2026-07-14",
+            "plate_label": "PLT-0091",
+            "status": "complete"
+          },
+          "error": null
+        },
+        {
+          "index": 4,
+          "type": "tool_call",
+          "tool": "create_assay_run",
+          "arguments": {
+            "assay_id": 9201,
+            "compound_id": 9101,
+            "run_on": "2026-08-01",
+            "plate_label": "CONF-001",
+            "status": "pending"
+          },
+          "ok": true,
+          "output": {
+            "run_id": 9303,
+            "assay_id": 9201,
+            "compound_id": 9101,
+            "run_on": "2026-08-01",
+            "plate_label": "CONF-001",
+            "status": "pending"
+          },
+          "error": null
+        }
+      ],
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "vcode"
+      },
+      "disclosure": "A checked-in reference trace replayed against a fresh SQLite copy. This proves the task can pass; it is not a model submission."
+    },
+    {
+      "id": "random",
+      "label": "Seeded random floor",
+      "kind": "non-model failure floor",
+      "modelId": null,
+      "taskId": "bio_promote_hit_compound",
+      "status": "failed",
+      "steps": [
+        {
+          "index": 1,
+          "type": "tool_call",
+          "tool": "count_assay_runs",
+          "arguments": {
+            "run_id": 6,
+            "assay_id": 13,
+            "compound_id": 16,
+            "run_on": "pending",
+            "plate_label": "active",
+            "status": "pending"
+          },
+          "ok": true,
+          "output": 0,
+          "error": null
+        },
+        {
+          "index": 2,
+          "type": "tool_call",
+          "tool": "get_literature_ref",
+          "arguments": {
+            "ref_id": 20
+          },
+          "ok": true,
+          "output": null,
+          "error": null
+        },
+        {
+          "index": 3,
+          "type": "tool_call",
+          "tool": "create_literature_ref",
+          "arguments": {
+            "target_id": 7,
+            "citation": "42",
+            "doi": "active",
+            "relevance": "active"
+          },
+          "ok": false,
+          "output": null,
+          "error": "CHECK constraint failed: relevance"
+        },
+        {
+          "index": 4,
+          "type": "tool_call",
+          "tool": "get_assay",
+          "arguments": {
+            "assay_id": 3
+          },
+          "ok": true,
+          "output": {
+            "assay_id": 3,
+            "target_id": 2,
+            "name": "Priority Retail Assay",
+            "assay_type": "premium",
+            "readout": "General Readout B61",
+            "status": "suspended"
+          },
+          "error": null
+        },
+        {
+          "index": 5,
+          "type": "tool_call",
+          "tool": "delete_assay",
+          "arguments": {
+            "assay_id": 2
+          },
+          "ok": true,
+          "output": true,
+          "error": null
+        },
+        {
+          "index": 6,
+          "type": "tool_call",
+          "tool": "count_compounds",
+          "arguments": {
+            "compound_id": 16,
+            "code": "test",
+            "name": "pending",
+            "molecular_weight": 77.86,
+            "status": "active",
+            "screening_note": "active"
+          },
+          "ok": true,
+          "output": 0,
+          "error": null
+        }
+      ],
+      "verifier": {
+        "passed": false,
+        "reward": 0.0,
+        "gradable": true,
+        "basis": "vcode"
+      },
+      "disclosure": "A seeded random tool policy against the same task and fresh world. It is a discrimination floor, not an agent or model result."
+    }
+  ]
+}

evidence/filingbench-50.json

@@ -0,0 +1,477 @@
+{
+  "schemaVersion": "blobfish.benchmark-evidence.v1",
+  "benchmarkSlug": "filingbench-50",
+  "observedAt": "2026-08-26",
+  "source": {
+    "datasetId": "vals/financeagent",
+    "tasks": 50,
+    "harborUrl": "https://hub.harborframework.com/datasets/vals/financeagent",
+    "inspection": "Harbor CLI 0.21.0 dataset metadata and package contract",
+    "relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks."
+  },
+  "world": {
+    "path": "benchmark/worlds/filingbench-50",
+    "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/filingbench-50",
+    "worldId": "env_blobfish_benchmarks_filingbench_financia_599d3443",
+    "worldDigest": "cd3085c1e21b8793bd4ba01fe48bd2a711e793f8233d931b4944fa1bafadaf41",
+    "vertical": "financial_analysis",
+    "tables": 6,
+    "rows": 116,
+    "tools": 44,
+    "tasks": 4,
+    "readOnlyTools": 23,
+    "mutatingTools": 21,
+    "entities": [
+      {
+        "name": "fa_companies",
+        "rows": 20
+      },
+      {
+        "name": "financial_periods",
+        "rows": 21
+      },
+      {
+        "name": "dcf_models",
+        "rows": 19
+      },
+      {
+        "name": "dcf_assumptions",
+        "rows": 20
+      },
+      {
+        "name": "lbo_models",
+        "rows": 18
+      },
+      {
+        "name": "review_notes",
+        "rows": 18
+      }
+    ]
+  },
+  "evaluation": {
+    "contract": "Fresh database copy, recorded tool trace, deterministic VCode or exact answer verification",
+    "referenceRunId": "release-oracle",
+    "floorRunId": "release-random",
+    "baselines": [
+      {
+        "id": "reference",
+        "label": "Reference solver",
+        "kind": "Solvability ceiling",
+        "disclosure": "Deterministic trace replay; not a model rank.",
+        "tasksAttempted": 4,
+        "tasksPassed": 4,
+        "passRate": 1.0,
+        "meanReward": 1.0,
+        "toolCalls": 9,
+        "successfulToolCalls": 9,
+        "episodeErrors": 0
+      },
+      {
+        "id": "random",
+        "label": "Seeded random floor",
+        "kind": "Discrimination floor",
+        "disclosure": "Seeded random tool policy; not a model rank.",
+        "tasksAttempted": 4,
+        "tasksPassed": 0,
+        "passRate": 0.0,
+        "meanReward": 0.0,
+        "toolCalls": 24,
+        "successfulToolCalls": 19,
+        "episodeErrors": 0
+      }
+    ],
+    "claim": "All companion tasks are solvable by their reference traces and reject the seeded random floor. No paid or frontier model run is claimed."
+  },
+  "tasks": [
+    {
+      "taskId": "fa_finalize_solstice_dcf",
+      "instruction": "Finalise the Solstice base-case DCF (model_id 9201, company 9001). Total the free_cash_flow across every financial period belonging to company 9001 — other companies' periods do not count. Set model 9201's enterprise_value to that total and its status to 'final', mark its WACC assumption (assumption_id 9301) as status 'verified', and file the review: create a review note on model 9201 from reviewer 'Head of Research' with status 'open' and a body containing the summed cash flow.",
+      "split": "train",
+      "generatorDifficultyLabel": "hard",
+      "difficultyDisclosure": "Generator label only; no empirical model difficulty claim is published.",
+      "requiredTools": [
+        "search_financial_periods",
+        "update_dcf_model",
+        "update_dcf_assumption",
+        "create_review_note"
+      ],
+      "referenceSteps": 4,
+      "referencePassed": true,
+      "floorPassed": false,
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "vcode"
+      }
+    },
+    {
+      "taskId": "fa_revert_model_without_discount_rate",
+      "instruction": "Model governance: a DCF may not carry status 'final' with a discount_rate of 0 — there is nothing to discount by. Find that model and set its status back to 'draft'. Models that are already draft are left alone.",
+      "split": "train",
+      "generatorDifficultyLabel": "medium",
+      "difficultyDisclosure": "Generator label only; no empirical model difficulty claim is published.",
+      "requiredTools": [
+        "search_dcf_models",
+        "update_dcf_model"
+      ],
+      "referenceSteps": 2,
+      "referencePassed": true,
+      "floorPassed": false,
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "vcode"
+      }
+    },
+    {
+      "taskId": "fa_duplicate_company_listing",
+      "instruction": "The same issuer appears twice under two listings. How many companies are named exactly 'Solstice Energy'? Answer with the number.",
+      "split": "train",
+      "generatorDifficultyLabel": "easy",
+      "difficultyDisclosure": "Generator label only; no empirical model difficulty claim is published.",
+      "requiredTools": [
+        "count_fa_companies"
+      ],
+      "referenceSteps": 1,
+      "referencePassed": true,
+      "floorPassed": false,
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "answer_match (vcode does not discriminate)"
+      }
+    },
+    {
+      "taskId": "fa_dedupe_dcf_assumption",
+      "instruction": "Model 9202 carries the 'Terminal growth' assumption twice (assumption_id 9311 and 9312). Delete the duplicate 9312 and keep 9311.",
+      "split": "heldout",
+      "generatorDifficultyLabel": "medium",
+      "difficultyDisclosure": "Generator label only; no empirical model difficulty claim is published.",
+      "requiredTools": [
+        "search_dcf_assumptions",
+        "delete_dcf_assumption"
+      ],
+      "referenceSteps": 2,
+      "referencePassed": true,
+      "floorPassed": false,
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "vcode"
+      }
+    }
+  ],
+  "assets": [
+    {
+      "path": "environment.db",
+      "label": "Seeded SQLite snapshot",
+      "kind": "SQLite",
+      "description": "Pristine state copied for every episode.",
+      "bytes": 32768,
+      "sha256": "f70a5b84565d5aa98bbcb1aeff9c7a0742213f85c3ad876a46b0d59265875e51",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/filingbench-50/environment.db"
+    },
+    {
+      "path": "schema.sql",
+      "label": "Relational schema",
+      "kind": "SQL",
+      "description": "Human-readable table and constraint definitions.",
+      "bytes": 1574,
+      "sha256": "510bf7e2943495cb8a329c4ccce54ffc7376226dedd2b9c11288d8fa78dc40dd",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/filingbench-50/schema.sql"
+    },
+    {
+      "path": "seed.sql",
+      "label": "Seed data",
+      "kind": "SQL",
+      "description": "Human-readable source for the released world state.",
+      "bytes": 14116,
+      "sha256": "c888acecc74c67054a759370315e8573f3315579c287ac540b21b57d37a3dc44",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/filingbench-50/seed.sql"
+    },
+    {
+      "path": "tools.json",
+      "label": "MCP tool catalog",
+      "kind": "JSON",
+      "description": "Public tool names, schemas, and implementations.",
+      "bytes": 104507,
+      "sha256": "6653f9d2a6e133047d7bb9f30b8ef81a74debb4efc94b94785664894036d4414",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/filingbench-50/tools.json"
+    },
+    {
+      "path": "tasks.json",
+      "label": "Task manifest",
+      "kind": "JSON",
+      "description": "Instructions, splits, reference traces, and isolated verifier programs.",
+      "bytes": 106953,
+      "sha256": "96388b0f39e51ed14f718001371783a506e5654586235e0ae1efd4decfbaf80b",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/filingbench-50/tasks.json"
+    },
+    {
+      "path": "mcp/manifest.json",
+      "label": "MCP manifest",
+      "kind": "JSON",
+      "description": "Stdio server declaration for the world.",
+      "bytes": 2039,
+      "sha256": "e92914c7b33fd3e5a50cf7c908177962734d3e7e775da258ca4e62f1437fd301",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/filingbench-50/mcp/manifest.json"
+    },
+    {
+      "path": "world.json",
+      "label": "World manifest",
+      "kind": "JSON",
+      "description": "Content identity, counts, split, and admission evidence.",
+      "bytes": 2527,
+      "sha256": "8475c86667f578b550dcc0bf5c449981036ea1a8c15dc9613906bb30165f972e",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/filingbench-50/world.json"
+    }
+  ],
+  "trajectories": [
+    {
+      "id": "reference",
+      "label": "Reference solver",
+      "kind": "deterministic solvability ceiling",
+      "modelId": null,
+      "taskId": "fa_finalize_solstice_dcf",
+      "status": "passed",
+      "steps": [
+        {
+          "index": 1,
+          "type": "tool_call",
+          "tool": "search_financial_periods",
+          "arguments": {
+            "company_id": 9001,
+            "limit": 50
+          },
+          "ok": true,
+          "output": [
+            {
+              "period_id": 9101,
+              "company_id": 9001,
+              "period_label": "2024",
+              "revenue": 800.0,
+              "ebitda": 200.0,
+              "capex": 60.0,
+              "free_cash_flow": 100.0
+            },
+            {
+              "period_id": 9102,
+              "company_id": 9001,
+              "period_label": "2025",
+              "revenue": 900.0,
+              "ebitda": 230.0,
+              "capex": 65.0,
+              "free_cash_flow": 120.0
+            },
+            {
+              "period_id": 9103,
+              "company_id": 9001,
+              "period_label": "2026",
+              "revenue": 1000.0,
+              "ebitda": 260.0,
+              "capex": 70.0,
+              "free_cash_flow": 140.0
+            }
+          ],
+          "error": null
+        },
+        {
+          "index": 2,
+          "type": "tool_call",
+          "tool": "update_dcf_model",
+          "arguments": {
+            "model_id": 9201,
+            "enterprise_value": 360.0,
+            "status": "final"
+          },
+          "ok": true,
+          "output": {
+            "model_id": 9201,
+            "company_id": 9001,
+            "name": "Solstice base case",
+            "discount_rate": 0.09,
+            "terminal_growth": 0.02,
+            "status": "final",
+            "enterprise_value": 360.0
+          },
+          "error": null
+        },
+        {
+          "index": 3,
+          "type": "tool_call",
+          "tool": "update_dcf_assumption",
+          "arguments": {
+            "assumption_id": 9301,
+            "status": "verified"
+          },
+          "ok": true,
+          "output": {
+            "assumption_id": 9301,
+            "model_id": 9201,
+            "name": "WACC",
+            "value": 0.09,
+            "source": "Capital IQ",
+            "status": "verified"
+          },
+          "error": null
+        },
+        {
+          "index": 4,
+          "type": "tool_call",
+          "tool": "create_review_note",
+          "arguments": {
+            "model_id": 9201,
+            "reviewer_name": "Head of Research",
+            "body": "Unlevered free cash flow across the period set totals 360.",
+            "status": "open"
+          },
+          "ok": true,
+          "output": {
+            "note_id": 9502,
+            "model_id": 9201,
+            "reviewer_name": "Head of Research",
+            "body": "Unlevered free cash flow across the period set totals 360.",
+            "status": "open"
+          },
+          "error": null
+        }
+      ],
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "vcode"
+      },
+      "disclosure": "A checked-in reference trace replayed against a fresh SQLite copy. This proves the task can pass; it is not a model submission."
+    },
+    {
+      "id": "random",
+      "label": "Seeded random floor",
+      "kind": "non-model failure floor",
+      "modelId": null,
+      "taskId": "fa_finalize_solstice_dcf",
+      "status": "failed",
+      "steps": [
+        {
+          "index": 1,
+          "type": "tool_call",
+          "tool": "create_dcf_model",
+          "arguments": {
+            "company_id": 4,
+            "name": "pending",
+            "discount_rate": 67.54,
+            "terminal_growth": 78.06,
+            "status": "test",
+            "enterprise_value": 8.06
+          },
+          "ok": false,
+          "output": null,
+          "error": "CHECK constraint failed: status"
+        },
+        {
+          "index": 2,
+          "type": "tool_call",
+          "tool": "create_dcf_assumption",
+          "arguments": {
+            "model_id": 9,
+            "name": "active",
+            "value": 57.89,
+            "source": "pending",
+            "status": "active"
+          },
+          "ok": true,
+          "output": {
+            "assumption_id": 9313,
+            "model_id": 9,
+            "name": "active",
+            "value": 57.89,
+            "source": "pending",
+            "status": "active"
+          },
+          "error": null
+        },
+        {
+          "index": 3,
+          "type": "tool_call",
+          "tool": "count_fa_companies",
+          "arguments": {
+            "company_id": 19,
+            "name": "test",
+            "ticker": "42",
+            "sector": "test"
+          },
+          "ok": true,
+          "output": 0,
+          "error": null
+        },
+        {
+          "index": 4,
+          "type": "tool_call",
+          "tool": "get_review_note",
+          "arguments": {
+            "note_id": 16
+          },
+          "ok": true,
+          "output": {
+            "note_id": 16,
+            "model_id": 10,
+            "reviewer_name": "Standard Internal Review Note",
+            "body": "{\"checksum\": \"042d186d\", \"received_via\": \"manual_entry\", \"record_type\": \"review_note\", \"reference\": \"RN-001004\", \"source\": \"web_portal\", \"version\": 1}",
+            "status": "closed"
+          },
+          "error": null
+        },
+        {
+          "index": 5,
+          "type": "tool_call",
+          "tool": "search_dcf_models",
+          "arguments": {
+            "query": "active",
+            "limit": 19,
+            "company_id": 7,
+            "name": "active",
+            "discount_rate": 39.62,
+            "terminal_growth": 89.02,
+            "status": "pending",
+            "enterprise_value": 91.65
+          },
+          "ok": true,
+          "output": [],
+          "error": null
+        },
+        {
+          "index": 6,
+          "type": "tool_call",
+          "tool": "list_dcf_assumptions_for_dcf_model",
+          "arguments": {
+            "model_id": 6,
+            "limit": 19
+          },
+          "ok": true,
+          "output": [
+            {
+              "assumption_id": 7,
+              "model_id": 6,
+              "name": "Quarterly Regional Dcf Assumption",
+              "value": 83.46,
+              "source": "General Source F92",
+              "status": "unverified"
+            }
+          ],
+          "error": null
+        }
+      ],
+      "verifier": {
+        "passed": false,
+        "reward": 0.0,
+        "gradable": true,
+        "basis": "vcode"
+      },
+      "disclosure": "A seeded random tool policy against the same task and fresh world. It is a discrimination floor, not an agent or model result."
+    }
+  ]
+}

evidence/opsbench-14.json

@@ -0,0 +1,461 @@
+{
+  "schemaVersion": "blobfish.benchmark-evidence.v1",
+  "benchmarkSlug": "opsbench-14",
+  "observedAt": "2026-08-26",
+  "source": {
+    "datasetId": "Enterprise-Bench/l1-l2-bench",
+    "tasks": 14,
+    "harborUrl": "https://hub.harborframework.com/datasets/Enterprise-Bench/l1-l2-bench",
+    "inspection": "Harbor CLI 0.21.0 dataset metadata and package contract",
+    "relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks."
+  },
+  "world": {
+    "path": "benchmark/worlds/opsbench-14",
+    "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/opsbench-14",
+    "worldId": "env_blobfish_benchmarks_opsbench_enterprise__8377a7f8",
+    "worldDigest": "39937737e9b493a2e1a4a10ba85780b15de76a222cf180538e40969cb6e3f1a8",
+    "vertical": "operations",
+    "tables": 6,
+    "rows": 116,
+    "tools": 42,
+    "tasks": 4,
+    "readOnlyTools": 21,
+    "mutatingTools": 21,
+    "entities": [
+      {
+        "name": "vendors",
+        "rows": 20
+      },
+      {
+        "name": "process_docs",
+        "rows": 19
+      },
+      {
+        "name": "process_steps",
+        "rows": 22
+      },
+      {
+        "name": "change_requests",
+        "rows": 19
+      },
+      {
+        "name": "compliance_controls",
+        "rows": 18
+      },
+      {
+        "name": "vendor_reviews",
+        "rows": 18
+      }
+    ]
+  },
+  "evaluation": {
+    "contract": "Fresh database copy, recorded tool trace, deterministic VCode or exact answer verification",
+    "referenceRunId": "release-oracle",
+    "floorRunId": "release-random",
+    "baselines": [
+      {
+        "id": "reference",
+        "label": "Reference solver",
+        "kind": "Solvability ceiling",
+        "disclosure": "Deterministic trace replay; not a model rank.",
+        "tasksAttempted": 4,
+        "tasksPassed": 4,
+        "passRate": 1.0,
+        "meanReward": 1.0,
+        "toolCalls": 9,
+        "successfulToolCalls": 9,
+        "episodeErrors": 0
+      },
+      {
+        "id": "random",
+        "label": "Seeded random floor",
+        "kind": "Discrimination floor",
+        "disclosure": "Seeded random tool policy; not a model rank.",
+        "tasksAttempted": 4,
+        "tasksPassed": 0,
+        "passRate": 0.0,
+        "meanReward": 0.0,
+        "toolCalls": 24,
+        "successfulToolCalls": 22,
+        "episodeErrors": 0
+      }
+    ],
+    "claim": "All companion tasks are solvable by their reference traces and reject the seeded random floor. No paid or frontier model run is claimed."
+  },
+  "tasks": [
+    {
+      "taskId": "ops_reject_uncontrolled_change",
+      "instruction": "Change request 9301 is a high-risk change to the Vendor Onboarding process (doc_id 9101). Policy: a high-risk change cannot be approved while any step of that process has no control_ref mapped. Count the DISTINCT steps of doc 9101 with an empty control_ref — note step 9202 is a duplicate of 9201 and counts once. Because at least one step is uncontrolled: set change request 9301's status to 'rejected', set process doc 9101's status to 'needs_controls', and open the remediation by creating a compliance control named 'Step control backfill' with framework 'SOC2' and status 'untested'.",
+      "split": "heldout",
+      "generatorDifficultyLabel": "hard",
+      "difficultyDisclosure": "Generator label only; no empirical model difficulty claim is published.",
+      "requiredTools": [
+        "search_process_steps",
+        "update_change_request",
+        "update_process_doc",
+        "create_compliance_control"
+      ],
+      "referenceSteps": 4,
+      "referencePassed": true,
+      "floorPassed": false,
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "vcode"
+      }
+    },
+    {
+      "taskId": "ops_flag_overdue_vendor_renewal",
+      "instruction": "Today is 2026-07-20. Any active vendor whose renewal_on date has already passed must be set to status 'renewal_overdue'. Vendors whose renewal is still ahead stay 'active'.",
+      "split": "train",
+      "generatorDifficultyLabel": "medium",
+      "difficultyDisclosure": "Generator label only; no empirical model difficulty claim is published.",
+      "requiredTools": [
+        "search_vendors",
+        "update_vendor"
+      ],
+      "referenceSteps": 2,
+      "referencePassed": true,
+      "floorPassed": false,
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "vcode"
+      }
+    },
+    {
+      "taskId": "ops_duplicate_vendor_name",
+      "instruction": "Two legal entities are registered under one trading name, so spend is split. How many vendors are named exactly 'Cascade Facilities Ltd'? Answer with the number.",
+      "split": "heldout",
+      "generatorDifficultyLabel": "easy",
+      "difficultyDisclosure": "Generator label only; no empirical model difficulty claim is published.",
+      "requiredTools": [
+        "count_vendors"
+      ],
+      "referenceSteps": 1,
+      "referencePassed": true,
+      "floorPassed": false,
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "answer_match (vcode does not discriminate)"
+      }
+    },
+    {
+      "taskId": "ops_dedupe_process_step",
+      "instruction": "A bad merge left step 1 of the Vendor Onboarding process recorded twice (step_id 9201 and 9202). Delete the duplicate 9202 and keep 9201.",
+      "split": "train",
+      "generatorDifficultyLabel": "medium",
+      "difficultyDisclosure": "Generator label only; no empirical model difficulty claim is published.",
+      "requiredTools": [
+        "search_process_steps",
+        "delete_process_step"
+      ],
+      "referenceSteps": 2,
+      "referencePassed": true,
+      "floorPassed": false,
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "vcode"
+      }
+    }
+  ],
+  "assets": [
+    {
+      "path": "environment.db",
+      "label": "Seeded SQLite snapshot",
+      "kind": "SQLite",
+      "description": "Pristine state copied for every episode.",
+      "bytes": 32768,
+      "sha256": "d6ca2df3416d0c441903d697c21418639839a44e6f43dd92cc0451baaead43fb",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/opsbench-14/environment.db"
+    },
+    {
+      "path": "schema.sql",
+      "label": "Relational schema",
+      "kind": "SQL",
+      "description": "Human-readable table and constraint definitions.",
+      "bytes": 1442,
+      "sha256": "0e2ed2c463b00b1896d51e02a8abd3b9e743bfcdc73817ed8a09624f5fb8109d",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/opsbench-14/schema.sql"
+    },
+    {
+      "path": "seed.sql",
+      "label": "Seed data",
+      "kind": "SQL",
+      "description": "Human-readable source for the released world state.",
+      "bytes": 13353,
+      "sha256": "9457c0e86a30ba262563ebe9bbac22edeb38d07702018689154ec6e783373934",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/opsbench-14/seed.sql"
+    },
+    {
+      "path": "tools.json",
+      "label": "MCP tool catalog",
+      "kind": "JSON",
+      "description": "Public tool names, schemas, and implementations.",
+      "bytes": 99191,
+      "sha256": "c8fbd5a4df7d08a3bdd0ad292f2cbc362f118e03f3c4b4c2da5f2ed7fbe0a315",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/opsbench-14/tools.json"
+    },
+    {
+      "path": "tasks.json",
+      "label": "Task manifest",
+      "kind": "JSON",
+      "description": "Instructions, splits, reference traces, and isolated verifier programs.",
+      "bytes": 106649,
+      "sha256": "0994054741ca6f7517a42584643e4570f434f7347dcf4fd59587f5525f97b1f8",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/opsbench-14/tasks.json"
+    },
+    {
+      "path": "mcp/manifest.json",
+      "label": "MCP manifest",
+      "kind": "JSON",
+      "description": "Stdio server declaration for the world.",
+      "bytes": 1989,
+      "sha256": "d431118d99a969b59a6f86bad4a178ddf673da1501091a11ba744a35df285dd5",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/opsbench-14/mcp/manifest.json"
+    },
+    {
+      "path": "world.json",
+      "label": "World manifest",
+      "kind": "JSON",
+      "description": "Content identity, counts, split, and admission evidence.",
+      "bytes": 2539,
+      "sha256": "42447057cbe5a476888045d6601ad7c428d05a7aa3c2bc144cebb176bd85dd1b",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/opsbench-14/world.json"
+    }
+  ],
+  "trajectories": [
+    {
+      "id": "reference",
+      "label": "Reference solver",
+      "kind": "deterministic solvability ceiling",
+      "modelId": null,
+      "taskId": "ops_reject_uncontrolled_change",
+      "status": "passed",
+      "steps": [
+        {
+          "index": 1,
+          "type": "tool_call",
+          "tool": "search_process_steps",
+          "arguments": {
+            "doc_id": 9101,
+            "limit": 50
+          },
+          "ok": true,
+          "output": [
+            {
+              "step_id": 9201,
+              "doc_id": 9101,
+              "step_number": 1,
+              "description": "Collect vendor tax documentation",
+              "control_ref": "CTRL-014"
+            },
+            {
+              "step_id": 9202,
+              "doc_id": 9101,
+              "step_number": 1,
+              "description": "Collect vendor tax documentation",
+              "control_ref": "CTRL-014"
+            },
+            {
+              "step_id": 9203,
+              "doc_id": 9101,
+              "step_number": 2,
+              "description": "Run sanctions screening",
+              "control_ref": "CTRL-021"
+            },
+            {
+              "step_id": 9204,
+              "doc_id": 9101,
+              "step_number": 3,
+              "description": "Grant system access",
+              "control_ref": ""
+            }
+          ],
+          "error": null
+        },
+        {
+          "index": 2,
+          "type": "tool_call",
+          "tool": "update_change_request",
+          "arguments": {
+            "request_id": 9301,
+            "status": "rejected"
+          },
+          "ok": true,
+          "output": {
+            "request_id": 9301,
+            "doc_id": 9101,
+            "title": "Skip sanctions screening for tier3",
+            "risk_level": "high",
+            "status": "rejected",
+            "requested_by": "Riley Garcia"
+          },
+          "error": null
+        },
+        {
+          "index": 3,
+          "type": "tool_call",
+          "tool": "update_process_doc",
+          "arguments": {
+            "doc_id": 9101,
+            "status": "needs_controls"
+          },
+          "ok": true,
+          "output": {
+            "doc_id": 9101,
+            "name": "Vendor Onboarding",
+            "owner_name": "Jordan Patel",
+            "status": "needs_controls",
+            "version": 3
+          },
+          "error": null
+        },
+        {
+          "index": 4,
+          "type": "tool_call",
+          "tool": "create_compliance_control",
+          "arguments": {
+            "name": "Step control backfill",
+            "framework": "SOC2",
+            "status": "untested",
+            "last_tested_on": ""
+          },
+          "ok": true,
+          "output": {
+            "control_id": 9402,
+            "name": "Step control backfill",
+            "framework": "SOC2",
+            "status": "untested",
+            "last_tested_on": ""
+          },
+          "error": null
+        }
+      ],
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "vcode"
+      },
+      "disclosure": "A checked-in reference trace replayed against a fresh SQLite copy. This proves the task can pass; it is not a model submission."
+    },
+    {
+      "id": "random",
+      "label": "Seeded random floor",
+      "kind": "non-model failure floor",
+      "modelId": null,
+      "taskId": "ops_reject_uncontrolled_change",
+      "status": "failed",
+      "steps": [
+        {
+          "index": 1,
+          "type": "tool_call",
+          "tool": "search_process_steps",
+          "arguments": {
+            "query": "active",
+            "limit": 11,
+            "doc_id": 2,
+            "step_number": 6,
+            "description": "pending",
+            "control_ref": "active"
+          },
+          "ok": true,
+          "output": [],
+          "error": null
+        },
+        {
+          "index": 2,
+          "type": "tool_call",
+          "tool": "count_change_requests",
+          "arguments": {
+            "request_id": 14,
+            "doc_id": 16,
+            "title": "pending",
+            "risk_level": "42",
+            "status": "active",
+            "requested_by": "pending"
+          },
+          "ok": true,
+          "output": 0,
+          "error": null
+        },
+        {
+          "index": 3,
+          "type": "tool_call",
+          "tool": "count_vendors",
+          "arguments": {
+            "vendor_id": 19,
+            "name": "test",
+            "category": "pending",
+            "tier": "42",
+            "status": "42",
+            "renewal_on": "pending"
+          },
+          "ok": true,
+          "output": 0,
+          "error": null
+        },
+        {
+          "index": 4,
+          "type": "tool_call",
+          "tool": "create_process_step",
+          "arguments": {
+            "doc_id": 4,
+            "step_number": 20,
+            "description": "42",
+            "control_ref": "test"
+          },
+          "ok": true,
+          "output": {
+            "step_id": 9212,
+            "doc_id": 4,
+            "step_number": 20,
+            "description": "42",
+            "control_ref": "test"
+          },
+          "error": null
+        },
+        {
+          "index": 5,
+          "type": "tool_call",
+          "tool": "create_vendor_review",
+          "arguments": {
+            "vendor_id": 9,
+            "reviewer_name": "42",
+            "score": 23.69,
+            "status": "pending",
+            "reviewed_on": "active"
+          },
+          "ok": false,
+          "output": null,
+          "error": "CHECK constraint failed: status"
+        },
+        {
+          "index": 6,
+          "type": "tool_call",
+          "tool": "delete_process_doc",
+          "arguments": {
+            "doc_id": 18
+          },
+          "ok": true,
+          "output": false,
+          "error": null
+        }
+      ],
+      "verifier": {
+        "passed": false,
+        "reward": 0.0,
+        "gradable": true,
+        "basis": "vcode"
+      },
+      "disclosure": "A seeded random tool policy against the same task and fresh world. It is a discrimination floor, not an agent or model result."
+    }
+  ]
+}

evidence/repobench-500.json

@@ -0,0 +1,455 @@
+{
+  "schemaVersion": "blobfish.benchmark-evidence.v1",
+  "benchmarkSlug": "repobench-500",
+  "observedAt": "2026-08-26",
+  "source": {
+    "datasetId": "swe-bench/swe-bench-verified",
+    "tasks": 500,
+    "harborUrl": "https://hub.harborframework.com/datasets/swe-bench/swe-bench-verified",
+    "inspection": "Harbor CLI 0.21.0 dataset metadata and package contract",
+    "relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks."
+  },
+  "world": {
+    "path": "benchmark/worlds/repobench-500",
+    "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/repobench-500",
+    "worldId": "env_blobfish_benchmarks_repobench_software_e_0af840ee",
+    "worldDigest": "482b4cabbd31f60b18d71d5d709ae725e85c1a57f5736fdee8a40b4677546a4e",
+    "vertical": "engineering",
+    "tables": 6,
+    "rows": 115,
+    "tools": 44,
+    "tasks": 4,
+    "readOnlyTools": 23,
+    "mutatingTools": 21,
+    "entities": [
+      {
+        "name": "services",
+        "rows": 19
+      },
+      {
+        "name": "repositories",
+        "rows": 20
+      },
+      {
+        "name": "pull_requests",
+        "rows": 20
+      },
+      {
+        "name": "adr_records",
+        "rows": 18
+      },
+      {
+        "name": "incidents",
+        "rows": 19
+      },
+      {
+        "name": "standup_notes",
+        "rows": 19
+      }
+    ]
+  },
+  "evaluation": {
+    "contract": "Fresh database copy, recorded tool trace, deterministic VCode or exact answer verification",
+    "referenceRunId": "release-oracle",
+    "floorRunId": "release-random",
+    "baselines": [
+      {
+        "id": "reference",
+        "label": "Reference solver",
+        "kind": "Solvability ceiling",
+        "disclosure": "Deterministic trace replay; not a model rank.",
+        "tasksAttempted": 4,
+        "tasksPassed": 4,
+        "passRate": 1.0,
+        "meanReward": 1.0,
+        "toolCalls": 9,
+        "successfulToolCalls": 9,
+        "episodeErrors": 0
+      },
+      {
+        "id": "random",
+        "label": "Seeded random floor",
+        "kind": "Discrimination floor",
+        "disclosure": "Seeded random tool policy; not a model rank.",
+        "tasksAttempted": 4,
+        "tasksPassed": 0,
+        "passRate": 0.0,
+        "meanReward": 0.0,
+        "toolCalls": 24,
+        "successfulToolCalls": 20,
+        "episodeErrors": 0
+      }
+    ],
+    "claim": "All companion tasks are solvable by their reference traces and reject the seeded random floor. No paid or frontier model run is claimed."
+  },
+  "tasks": [
+    {
+      "taskId": "eng_resolve_ingest_incident",
+      "instruction": "Incident 9401 on the ingest service was mitigated at minute 175. Close it out: set incident 9401's resolved_minute to 175 and status to 'resolved', work out how long it lasted (resolved_minute minus detected_minute), then record the follow-up decision by creating an ADR on service 9001 titled 'Rate-limit the ingest queue' with status 'accepted', decided_on '2026-07-21' and a decision that states the incident duration as 'N minutes'; finally post a standup note on service 9001 by author 'Rowan Novak' for day_label 'tuesday' whose body references incident 9401.",
+      "split": "train",
+      "generatorDifficultyLabel": "hard",
+      "difficultyDisclosure": "Generator label only; no empirical model difficulty claim is published.",
+      "requiredTools": [
+        "get_incident",
+        "update_incident",
+        "create_adr_record",
+        "create_standup_note"
+      ],
+      "referenceSteps": 4,
+      "referencePassed": true,
+      "floorPassed": false,
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "vcode"
+      }
+    },
+    {
+      "taskId": "eng_reopen_pr_with_red_ci",
+      "instruction": "Merge policy: a pull request may not sit in status 'approved' while its checks_status is 'failing'. Find the offending PR and set its status back to 'open'. PRs whose checks pass keep their approved status.",
+      "split": "train",
+      "generatorDifficultyLabel": "medium",
+      "difficultyDisclosure": "Generator label only; no empirical model difficulty claim is published.",
+      "requiredTools": [
+        "search_pull_requests",
+        "update_pull_request"
+      ],
+      "referenceSteps": 2,
+      "referencePassed": true,
+      "floorPassed": false,
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "vcode"
+      }
+    },
+    {
+      "taskId": "eng_duplicate_repository_name",
+      "instruction": "After a service split, two services own a repository with the same name. How many repositories are named exactly 'ingest-worker'? Answer with the number.",
+      "split": "train",
+      "generatorDifficultyLabel": "easy",
+      "difficultyDisclosure": "Generator label only; no empirical model difficulty claim is published.",
+      "requiredTools": [
+        "count_repositories"
+      ],
+      "referenceSteps": 1,
+      "referencePassed": true,
+      "floorPassed": false,
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "answer_match (vcode does not discriminate)"
+      }
+    },
+    {
+      "taskId": "eng_dedupe_standup_note",
+      "instruction": "A flaky bot posted Monday's standup note twice on service 9001 (note_id 9501 and 9502). Delete the duplicate 9502 and keep 9501.",
+      "split": "heldout",
+      "generatorDifficultyLabel": "medium",
+      "difficultyDisclosure": "Generator label only; no empirical model difficulty claim is published.",
+      "requiredTools": [
+        "search_standup_notes",
+        "delete_standup_note"
+      ],
+      "referenceSteps": 2,
+      "referencePassed": true,
+      "floorPassed": false,
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "vcode"
+      }
+    }
+  ],
+  "assets": [
+    {
+      "path": "environment.db",
+      "label": "Seeded SQLite snapshot",
+      "kind": "SQLite",
+      "description": "Pristine state copied for every episode.",
+      "bytes": 32768,
+      "sha256": "12007f3262431078bc3f6476bec0de28d01e0b93acfcc2d86268d1ba36210303",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/repobench-500/environment.db"
+    },
+    {
+      "path": "schema.sql",
+      "label": "Relational schema",
+      "kind": "SQL",
+      "description": "Human-readable table and constraint definitions.",
+      "bytes": 1690,
+      "sha256": "76b797e52cbe538d5907b14387a8da0e6fae3c883e546d5af1c1504a3973255f",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/repobench-500/schema.sql"
+    },
+    {
+      "path": "seed.sql",
+      "label": "Seed data",
+      "kind": "SQL",
+      "description": "Human-readable source for the released world state.",
+      "bytes": 15051,
+      "sha256": "d24ba1b70830c020f4abfa3a702a5e6a1bcb5378202a1a3c605af09fbcdecf47",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/repobench-500/seed.sql"
+    },
+    {
+      "path": "tools.json",
+      "label": "MCP tool catalog",
+      "kind": "JSON",
+      "description": "Public tool names, schemas, and implementations.",
+      "bytes": 103816,
+      "sha256": "75077c064c2c026100e6e30008bac57e490f60918e72b48a0b8b18acfa1d61ea",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/repobench-500/tools.json"
+    },
+    {
+      "path": "tasks.json",
+      "label": "Task manifest",
+      "kind": "JSON",
+      "description": "Instructions, splits, reference traces, and isolated verifier programs.",
+      "bytes": 106569,
+      "sha256": "c05032fc9010fed55dd7dd853f10622fa4657c7fe1844160270e8644bd26fda0",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/repobench-500/tasks.json"
+    },
+    {
+      "path": "mcp/manifest.json",
+      "label": "MCP manifest",
+      "kind": "JSON",
+      "description": "Stdio server declaration for the world.",
+      "bytes": 1954,
+      "sha256": "1b85dad89ae04a265b46ef7b53d52a8660c7973ca1158442637d84cf25de5faf",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/repobench-500/mcp/manifest.json"
+    },
+    {
+      "path": "world.json",
+      "label": "World manifest",
+      "kind": "JSON",
+      "description": "Content identity, counts, split, and admission evidence.",
+      "bytes": 2517,
+      "sha256": "c38c5e25cf4ba718e434d8da896a9c90ee02b5f8d25a6c9357c4ee10b2f3ff7d",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/repobench-500/world.json"
+    }
+  ],
+  "trajectories": [
+    {
+      "id": "reference",
+      "label": "Reference solver",
+      "kind": "deterministic solvability ceiling",
+      "modelId": null,
+      "taskId": "eng_resolve_ingest_incident",
+      "status": "passed",
+      "steps": [
+        {
+          "index": 1,
+          "type": "tool_call",
+          "tool": "get_incident",
+          "arguments": {
+            "incident_id": 9401
+          },
+          "ok": true,
+          "output": {
+            "incident_id": 9401,
+            "service_id": 9001,
+            "summary": "Ingest queue saturated, consumers stalled",
+            "severity": "sev1",
+            "status": "open",
+            "detected_minute": 100,
+            "resolved_minute": null
+          },
+          "error": null
+        },
+        {
+          "index": 2,
+          "type": "tool_call",
+          "tool": "update_incident",
+          "arguments": {
+            "incident_id": 9401,
+            "resolved_minute": 175,
+            "status": "resolved"
+          },
+          "ok": true,
+          "output": {
+            "incident_id": 9401,
+            "service_id": 9001,
+            "summary": "Ingest queue saturated, consumers stalled",
+            "severity": "sev1",
+            "status": "resolved",
+            "detected_minute": 100,
+            "resolved_minute": 175
+          },
+          "error": null
+        },
+        {
+          "index": 3,
+          "type": "tool_call",
+          "tool": "create_adr_record",
+          "arguments": {
+            "service_id": 9001,
+            "title": "Rate-limit the ingest queue",
+            "decision": "Incident 9401 lasted 75 minutes; rate-limit producers to keep consumers from stalling.",
+            "status": "accepted",
+            "decided_on": "2026-07-21"
+          },
+          "ok": true,
+          "output": {
+            "adr_id": 9302,
+            "service_id": 9001,
+            "title": "Rate-limit the ingest queue",
+            "decision": "Incident 9401 lasted 75 minutes; rate-limit producers to keep consumers from stalling.",
+            "status": "accepted",
+            "decided_on": "2026-07-21"
+          },
+          "error": null
+        },
+        {
+          "index": 4,
+          "type": "tool_call",
+          "tool": "create_standup_note",
+          "arguments": {
+            "service_id": 9001,
+            "author": "Rowan Novak",
+            "day_label": "tuesday",
+            "body": "Incident 9401 resolved; writing the rate-limit ADR follow-up."
+          },
+          "ok": true,
+          "output": {
+            "note_id": 9503,
+            "service_id": 9001,
+            "author": "Rowan Novak",
+            "day_label": "tuesday",
+            "body": "Incident 9401 resolved; writing the rate-limit ADR follow-up."
+          },
+          "error": null
+        }
+      ],
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "vcode"
+      },
+      "disclosure": "A checked-in reference trace replayed against a fresh SQLite copy. This proves the task can pass; it is not a model submission."
+    },
+    {
+      "id": "random",
+      "label": "Seeded random floor",
+      "kind": "non-model failure floor",
+      "modelId": null,
+      "taskId": "eng_resolve_ingest_incident",
+      "status": "failed",
+      "steps": [
+        {
+          "index": 1,
+          "type": "tool_call",
+          "tool": "count_repositories",
+          "arguments": {
+            "repo_id": 20,
+            "service_id": 14,
+            "name": "42",
+            "default_branch": "42",
+            "visibility": "test"
+          },
+          "ok": true,
+          "output": 0,
+          "error": null
+        },
+        {
+          "index": 2,
+          "type": "tool_call",
+          "tool": "count_incidents",
+          "arguments": {
+            "incident_id": 10,
+            "service_id": 19,
+            "summary": "active",
+            "severity": "pending",
+            "status": "42",
+            "detected_minute": 6,
+            "resolved_minute": 10
+          },
+          "ok": true,
+          "output": 0,
+          "error": null
+        },
+        {
+          "index": 3,
+          "type": "tool_call",
+          "tool": "create_service",
+          "arguments": {
+            "name": "42",
+            "tier": "active",
+            "owner_team": "42",
+            "oncall_name": "pending"
+          },
+          "ok": true,
+          "output": {
+            "service_id": 9003,
+            "name": "42",
+            "tier": "active",
+            "owner_team": "42",
+            "oncall_name": "pending"
+          },
+          "error": null
+        },
+        {
+          "index": 4,
+          "type": "tool_call",
+          "tool": "delete_standup_note",
+          "arguments": {
+            "note_id": 8
+          },
+          "ok": true,
+          "output": true,
+          "error": null
+        },
+        {
+          "index": 5,
+          "type": "tool_call",
+          "tool": "list_adr_records_for_service",
+          "arguments": {
+            "service_id": 6,
+            "limit": 11
+          },
+          "ok": true,
+          "output": [
+            {
+              "adr_id": 1,
+              "service_id": 6,
+              "title": "Consolidated Regional Adr Record",
+              "decision": "Secondary Decision E10",
+              "status": "placed",
+              "decided_on": "Regional Decided On E18"
+            },
+            {
+              "adr_id": 3,
+              "service_id": 6,
+              "title": "Standard Partner Adr Record",
+              "decision": "Backup Decision E75",
+              "status": "accepted",
+              "decided_on": "Regional Decided On C71"
+            }
+          ],
+          "error": null
+        },
+        {
+          "index": 6,
+          "type": "tool_call",
+          "tool": "list_adr_records_for_service",
+          "arguments": {
+            "service_id": 20,
+            "limit": 10
+          },
+          "ok": true,
+          "output": [],
+          "error": null
+        }
+      ],
+      "verifier": {
+        "passed": false,
+        "reward": 0.0,
+        "gradable": true,
+        "basis": "vcode"
+      },
+      "disclosure": "A seeded random tool policy against the same task and fresh world. It is a discrimination floor, not an agent or model result."
+    }
+  ]
+}

evidence/servicebench-375.json

@@ -0,0 +1,443 @@
+{
+  "schemaVersion": "blobfish.benchmark-evidence.v1",
+  "benchmarkSlug": "servicebench-375",
+  "observedAt": "2026-08-26",
+  "source": {
+    "datasetId": "sierra-research/tau3-bench",
+    "tasks": 375,
+    "harborUrl": "https://hub.harborframework.com/datasets/sierra-research/tau3-bench",
+    "inspection": "Harbor CLI 0.21.0 dataset metadata and package contract",
+    "relationship": "The source suite and Blobfish companion world share a professional domain, but the companion tasks are independent diagnostic tasks—not converted source tasks."
+  },
+  "world": {
+    "path": "benchmark/worlds/servicebench-375",
+    "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/servicebench-375",
+    "worldId": "env_blobfish_benchmarks_servicebench_custome_79e1b30f",
+    "worldDigest": "ab0d887316ac287e98532ff96cd1ec24f979922a054a8e02b58e4d2e75f37260",
+    "vertical": "customer_support",
+    "tables": 6,
+    "rows": 115,
+    "tools": 43,
+    "tasks": 4,
+    "readOnlyTools": 22,
+    "mutatingTools": 21,
+    "entities": [
+      {
+        "name": "support_customers",
+        "rows": 20
+      },
+      {
+        "name": "sla_policies",
+        "rows": 19
+      },
+      {
+        "name": "tickets",
+        "rows": 20
+      },
+      {
+        "name": "ticket_messages",
+        "rows": 19
+      },
+      {
+        "name": "escalations",
+        "rows": 18
+      },
+      {
+        "name": "kb_articles",
+        "rows": 19
+      }
+    ]
+  },
+  "evaluation": {
+    "contract": "Fresh database copy, recorded tool trace, deterministic VCode or exact answer verification",
+    "referenceRunId": "release-oracle",
+    "floorRunId": "release-random",
+    "baselines": [
+      {
+        "id": "reference",
+        "label": "Reference solver",
+        "kind": "Solvability ceiling",
+        "disclosure": "Deterministic trace replay; not a model rank.",
+        "tasksAttempted": 4,
+        "tasksPassed": 4,
+        "passRate": 1.0,
+        "meanReward": 1.0,
+        "toolCalls": 9,
+        "successfulToolCalls": 9,
+        "episodeErrors": 0
+      },
+      {
+        "id": "random",
+        "label": "Seeded random floor",
+        "kind": "Discrimination floor",
+        "disclosure": "Seeded random tool policy; not a model rank.",
+        "tasksAttempted": 4,
+        "tasksPassed": 0,
+        "passRate": 0.0,
+        "meanReward": 0.0,
+        "toolCalls": 24,
+        "successfulToolCalls": 18,
+        "episodeErrors": 0
+      }
+    ],
+    "claim": "All companion tasks are solvable by their reference traces and reject the seeded random floor. No paid or frontier model run is claimed."
+  },
+  "tasks": [
+    {
+      "taskId": "cs_escalate_sla_breach",
+      "instruction": "It is now minute 180. Ticket 9201 belongs to an enterprise-plan customer and still has no first response. Look up the first_response_minutes of the SLA policy for that customer's plan and work out how many minutes past the SLA deadline the response now is (deadline = opened_minute + first_response_minutes). Then set ticket 9201's priority to 'urgent' and status to 'escalated', create an escalation on ticket 9201 with reason 'sla_first_response_breach', tier 'tier2' and status 'open', and post a holding reply: a ticket message on 9201 from author 'agent' at sent_minute 180 whose body states the overdue amount as 'N minutes'.",
+      "split": "train",
+      "generatorDifficultyLabel": "hard",
+      "difficultyDisclosure": "Generator label only; no empirical model difficulty claim is published.",
+      "requiredTools": [
+        "search_sla_policies",
+        "update_ticket",
+        "create_escalation",
+        "create_ticket_message"
+      ],
+      "referenceSteps": 4,
+      "referencePassed": true,
+      "floorPassed": false,
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "vcode"
+      }
+    },
+    {
+      "taskId": "cs_close_resolved_tickets",
+      "instruction": "Tidy the queue: any ticket that already has a resolved_minute recorded must not still be sitting in status 'open' — set it to 'resolved'. Tickets without a resolved_minute stay open.",
+      "split": "train",
+      "generatorDifficultyLabel": "medium",
+      "difficultyDisclosure": "Generator label only; no empirical model difficulty claim is published.",
+      "requiredTools": [
+        "search_tickets",
+        "update_ticket"
+      ],
+      "referenceSteps": 2,
+      "referencePassed": true,
+      "floorPassed": false,
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "vcode"
+      }
+    },
+    {
+      "taskId": "cs_duplicate_customer_name",
+      "instruction": "Two subsidiaries were onboarded under one trading name, so routing is ambiguous. How many support customers are named exactly 'Bluepeak Robotics'? Answer with the number.",
+      "split": "train",
+      "generatorDifficultyLabel": "easy",
+      "difficultyDisclosure": "Generator label only; no empirical model difficulty claim is published.",
+      "requiredTools": [
+        "count_support_customers"
+      ],
+      "referenceSteps": 1,
+      "referencePassed": true,
+      "floorPassed": false,
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "answer_match (vcode does not discriminate)"
+      }
+    },
+    {
+      "taskId": "cs_dedupe_kb_article",
+      "instruction": "Two knowledge-base articles cover the same fix under the identical title 'Recovering an offline fleet controller' (article_id 9501 is published, 9502 is a duplicate draft). Delete the duplicate draft 9502 and keep the published article 9501.",
+      "split": "heldout",
+      "generatorDifficultyLabel": "medium",
+      "difficultyDisclosure": "Generator label only; no empirical model difficulty claim is published.",
+      "requiredTools": [
+        "search_kb_articles",
+        "delete_kb_article"
+      ],
+      "referenceSteps": 2,
+      "referencePassed": true,
+      "floorPassed": false,
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "vcode"
+      }
+    }
+  ],
+  "assets": [
+    {
+      "path": "environment.db",
+      "label": "Seeded SQLite snapshot",
+      "kind": "SQLite",
+      "description": "Pristine state copied for every episode.",
+      "bytes": 32768,
+      "sha256": "f1c39d517436dd11eff05bf2da9440635d53b397ca3966e386e01f7f4a596537",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/servicebench-375/environment.db"
+    },
+    {
+      "path": "schema.sql",
+      "label": "Relational schema",
+      "kind": "SQL",
+      "description": "Human-readable table and constraint definitions.",
+      "bytes": 1521,
+      "sha256": "aae113321a7e8003690d02fdf2e6540f74f9735211eaaefacbb8da160f481a2d",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/servicebench-375/schema.sql"
+    },
+    {
+      "path": "seed.sql",
+      "label": "Seed data",
+      "kind": "SQL",
+      "description": "Human-readable source for the released world state.",
+      "bytes": 15471,
+      "sha256": "9b5b097b96bcd595d8f3edc80d437e9881739f350848b4414840effa81dd7839",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/servicebench-375/seed.sql"
+    },
+    {
+      "path": "tools.json",
+      "label": "MCP tool catalog",
+      "kind": "JSON",
+      "description": "Public tool names, schemas, and implementations.",
+      "bytes": 101198,
+      "sha256": "bbfffdac46d3bd61a8eea7e17c6c74effca6ceb2e2342daaf03cbb65f70da2a5",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/servicebench-375/tools.json"
+    },
+    {
+      "path": "tasks.json",
+      "label": "Task manifest",
+      "kind": "JSON",
+      "description": "Instructions, splits, reference traces, and isolated verifier programs.",
+      "bytes": 105899,
+      "sha256": "1a7d4d40cff99e0ae7e7fa21dc6fd5122f4f5a730fc939fec2c499e8046a613f",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/servicebench-375/tasks.json"
+    },
+    {
+      "path": "mcp/manifest.json",
+      "label": "MCP manifest",
+      "kind": "JSON",
+      "description": "Stdio server declaration for the world.",
+      "bytes": 1959,
+      "sha256": "d7fc28450ed8676dec434717efb0d3e3b1b802ab58c4b904504d7b84192e83f5",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/servicebench-375/mcp/manifest.json"
+    },
+    {
+      "path": "world.json",
+      "label": "World manifest",
+      "kind": "JSON",
+      "description": "Content identity, counts, split, and admission evidence.",
+      "bytes": 2495,
+      "sha256": "79dadcb06d047317b285f5f996c9e7568a386854d5e4f997b5667ce39bfda67c",
+      "repositoryUrl": "https://github.com/blobfishai/blobfishai/tree/main/benchmark/worlds/servicebench-375/world.json"
+    }
+  ],
+  "trajectories": [
+    {
+      "id": "reference",
+      "label": "Reference solver",
+      "kind": "deterministic solvability ceiling",
+      "modelId": null,
+      "taskId": "cs_escalate_sla_breach",
+      "status": "passed",
+      "steps": [
+        {
+          "index": 1,
+          "type": "tool_call",
+          "tool": "search_sla_policies",
+          "arguments": {
+            "plan": "enterprise",
+            "limit": 10
+          },
+          "ok": true,
+          "output": [
+            {
+              "policy_id": 9101,
+              "name": "Enterprise SLA",
+              "plan": "enterprise",
+              "first_response_minutes": 30,
+              "resolve_minutes": 240
+            }
+          ],
+          "error": null
+        },
+        {
+          "index": 2,
+          "type": "tool_call",
+          "tool": "update_ticket",
+          "arguments": {
+            "ticket_id": 9201,
+            "priority": "urgent",
+            "status": "escalated"
+          },
+          "ok": true,
+          "output": {
+            "ticket_id": 9201,
+            "customer_id": 9001,
+            "subject": "Fleet controller offline",
+            "status": "escalated",
+            "priority": "urgent",
+            "opened_minute": 100,
+            "first_response_minute": null,
+            "resolved_minute": null
+          },
+          "error": null
+        },
+        {
+          "index": 3,
+          "type": "tool_call",
+          "tool": "create_escalation",
+          "arguments": {
+            "ticket_id": 9201,
+            "reason": "sla_first_response_breach",
+            "tier": "tier2",
+            "status": "open"
+          },
+          "ok": true,
+          "output": {
+            "escalation_id": 9402,
+            "ticket_id": 9201,
+            "reason": "sla_first_response_breach",
+            "tier": "tier2",
+            "status": "open"
+          },
+          "error": null
+        },
+        {
+          "index": 4,
+          "type": "tool_call",
+          "tool": "create_ticket_message",
+          "arguments": {
+            "ticket_id": 9201,
+            "author": "agent",
+            "body": "Apologies — our first response is 50 minutes past the enterprise SLA. An engineer is now on this.",
+            "sent_minute": 180,
+            "channel": "email"
+          },
+          "ok": true,
+          "output": {
+            "message_id": 9303,
+            "ticket_id": 9201,
+            "author": "agent",
+            "body": "Apologies — our first response is 50 minutes past the enterprise SLA. An engineer is now on this.",
+            "sent_minute": 180,
+            "channel": "email"
+          },
+          "error": null
+        }
+      ],
+      "verifier": {
+        "passed": true,
+        "reward": 1.0,
+        "gradable": true,
+        "basis": "vcode"
+      },
+      "disclosure": "A checked-in reference trace replayed against a fresh SQLite copy. This proves the task can pass; it is not a model submission."
+    },
+    {
+      "id": "random",
+      "label": "Seeded random floor",
+      "kind": "non-model failure floor",
+      "modelId": null,
+      "taskId": "cs_escalate_sla_breach",
+      "status": "failed",
+      "steps": [
+        {
+          "index": 1,
+          "type": "tool_call",
+          "tool": "create_support_customer",
+          "arguments": {
+            "name": "test",
+            "plan": "test",
+            "region": "active"
+          },
+          "ok": true,
+          "output": {
+            "customer_id": 9004,
+            "name": "test",
+            "plan": "test",
+            "region": "active"
+          },
+          "error": null
+        },
+        {
+          "index": 2,
+          "type": "tool_call",
+          "tool": "search_sla_policies",
+          "arguments": {
+            "query": "42",
+            "limit": 11,
+            "name": "pending",
+            "plan": "42",
+            "first_response_minutes": 12,
+            "resolve_minutes": 4
+          },
+          "ok": true,
+          "output": [],
+          "error": null
+        },
+        {
+          "index": 3,
+          "type": "tool_call",
+          "tool": "count_sla_policies",
+          "arguments": {
+            "policy_id": 14,
+            "name": "test",
+            "plan": "42",
+            "first_response_minutes": 20,
+            "resolve_minutes": 1
+          },
+          "ok": true,
+          "output": 0,
+          "error": null
+        },
+        {
+          "index": 4,
+          "type": "tool_call",
+          "tool": "count_escalations",
+          "arguments": {
+            "escalation_id": 14,
+            "ticket_id": 7,
+            "reason": "test",
+            "tier": "active",
+            "status": "test"
+          },
+          "ok": true,
+          "output": 0,
+          "error": null
+        },
+        {
+          "index": 5,
+          "type": "tool_call",
+          "tool": "delete_escalation",
+          "arguments": {
+            "escalation_id": 6
+          },
+          "ok": true,
+          "output": true,
+          "error": null
+        },
+        {
+          "index": 6,
+          "type": "tool_call",
+          "tool": "count_sla_policies",
+          "arguments": {
+            "policy_id": 6,
+            "name": "pending",
+            "plan": "42",
+            "first_response_minutes": 5,
+            "resolve_minutes": 13
+          },
+          "ok": true,
+          "output": 0,
+          "error": null
+        }
+      ],
+      "verifier": {
+        "passed": false,
+        "reward": 0.0,
+        "gradable": true,
+        "basis": "vcode"
+      },
+      "disclosure": "A seeded random tool policy against the same task and fresh world. It is a discrimination floor, not an agent or model result."
+    }
+  ]
+}

worlds/defensebench-100/mcp/manifest.json

@@ -0,0 +1,68 @@
+{
+  "schema_version": "blobfish.mcp-service-layout.v1",
+  "world_id": "env_blobfish_benchmarks_defensebench_access__d2a0878c",
+  "grouping": "tool.forge_service provenance only; no name/description inference (spec invariant 10)",
+  "systems": [],
+  "server_count": 1,
+  "tool_count": 42,
+  "servers": [
+    {
+      "name": "core",
+      "kind": "core",
+      "home": "core/",
+      "tool_count": 42,
+      "tools": [
+        "count_access_grants",
+        "count_document_chunks",
+        "count_employees",
+        "count_indexed_documents",
+        "count_search_queries",
+        "count_sources",
+        "create_access_grant",
+        "create_document_chunk",
+        "create_employee",
+        "create_indexed_document",
+        "create_search_query",
+        "create_source",
+        "delete_access_grant",
+        "delete_document_chunk",
+        "delete_employee",
+        "delete_indexed_document",
+        "delete_search_query",
+        "delete_source",
+        "get_access_grant",
+        "get_document_chunk",
+        "get_employee",
+        "get_indexed_document",
+        "get_search_query",
+        "get_source",
+        "list_access_grants_for_employee",
+        "list_access_grants_for_indexed_document",
+        "list_document_chunks_for_indexed_document",
+        "list_indexed_documents_for_source",
+        "list_search_queries_for_employee",
+        "search_access_grants",
+        "search_document_chunks",
+        "search_employees",
+        "search_indexed_documents",
+        "search_search_queries",
+        "search_sources",
+        "transition_access_grant_status",
+        "update_access_grant",
+        "update_document_chunk",
+        "update_employee",
+        "update_indexed_document",
+        "update_search_query",
+        "update_source"
+      ],
+      "tables": [
+        "access_grants",
+        "document_chunks",
+        "employees",
+        "indexed_documents",
+        "search_queries",
+        "sources"
+      ]
+    }
+  ]
+}

worlds/defensebench-100/schema.sql

@@ -0,0 +1,7 @@
+CREATE TABLE access_grants (grant_id INTEGER PRIMARY KEY AUTOINCREMENT, document_id INTEGER REFERENCES indexed_documents(document_id), employee_id INTEGER REFERENCES employees(employee_id), level TEXT DEFAULT 'read', status TEXT DEFAULT 'active', CHECK ("status" IS NULL OR "status" IN ('active', 'inactive', 'revoked', 'suspended')));
+CREATE TABLE document_chunks (chunk_id INTEGER PRIMARY KEY AUTOINCREMENT, document_id INTEGER REFERENCES indexed_documents(document_id), ordinal INTEGER DEFAULT 0, body TEXT);
+CREATE TABLE employees (employee_id INTEGER PRIMARY KEY AUTOINCREMENT, name TEXT NOT NULL, email TEXT, team TEXT);
+CREATE TABLE indexed_documents (document_id INTEGER PRIMARY KEY AUTOINCREMENT, source_id INTEGER REFERENCES sources(source_id), title TEXT NOT NULL, owner_email TEXT, updated_on TEXT, sensitivity TEXT DEFAULT 'internal', status TEXT DEFAULT 'indexed');
+CREATE TABLE search_queries (query_id INTEGER PRIMARY KEY AUTOINCREMENT, employee_id INTEGER REFERENCES employees(employee_id), query_text TEXT, ran_on TEXT, result_count INTEGER DEFAULT 0);
+CREATE TABLE sources (source_id INTEGER PRIMARY KEY AUTOINCREMENT, name TEXT NOT NULL, kind TEXT, connected INTEGER DEFAULT 1, last_synced_on TEXT);
+DELETE FROM "sqlite_sequence";

worlds/defensebench-100/seed.sql

@@ -0,0 +1,127 @@
+INSERT INTO "access_grants" VALUES(1,4,10,'read','revoked');
+INSERT INTO "access_grants" VALUES(2,3,5,'read','active');
+INSERT INTO "access_grants" VALUES(3,4,3,'read','suspended');
+INSERT INTO "access_grants" VALUES(4,4,2,'read','active');
+INSERT INTO "access_grants" VALUES(5,3,3,'read','revoked');
+INSERT INTO "access_grants" VALUES(6,3,2,'read','revoked');
+INSERT INTO "access_grants" VALUES(7,8,2,'read','inactive');
+INSERT INTO "access_grants" VALUES(8,6,3,'read','suspended');
+INSERT INTO "access_grants" VALUES(9,8,2,'read','inactive');
+INSERT INTO "access_grants" VALUES(10,3,10,'read','active');
+INSERT INTO "access_grants" VALUES(11,2,9,'read','suspended');
+INSERT INTO "access_grants" VALUES(12,10,2,'read','active');
+INSERT INTO "access_grants" VALUES(13,10,11,'read','active');
+INSERT INTO "access_grants" VALUES(14,1,13,'read','active');
+INSERT INTO "access_grants" VALUES(15,3,6,'read','suspended');
+INSERT INTO "access_grants" VALUES(16,2,2,'read','revoked');
+INSERT INTO "access_grants" VALUES(17,6,11,'read','suspended');
+INSERT INTO "access_grants" VALUES(9401,9201,9101,'read','active');
+INSERT INTO "access_grants" VALUES(9402,9201,9102,'read','active');
+INSERT INTO "access_grants" VALUES(9403,9201,9103,'read','active');
+INSERT INTO "access_grants" VALUES(9404,9201,9104,'read','revoked');
+INSERT INTO "access_grants" VALUES(9411,9203,9102,'read','active');
+INSERT INTO "document_chunks" VALUES(1,6,0,'{"checksum": "ea03d17c", "received_via": "manual_entry", "record_type": "document_chunk", "reference": "DC-000001", "source": "web_portal", "version": 3}');
+INSERT INTO "document_chunks" VALUES(2,4,0,'{"checksum": "0b938c21", "received_via": "manual_entry", "record_type": "document_chunk", "reference": "DC-000002", "source": "erp_sync", "version": 1}');
+INSERT INTO "document_chunks" VALUES(3,10,0,'{"checksum": "c817ccba", "received_via": "manual_entry", "record_type": "document_chunk", "reference": "DC-000003", "source": "edi_gateway", "version": 4}');
+INSERT INTO "document_chunks" VALUES(4,12,0,'{"checksum": "59359023", "received_via": "manual_entry", "record_type": "document_chunk", "reference": "DC-000004", "source": "shopee", "version": 3}');
+INSERT INTO "document_chunks" VALUES(5,7,0,'{"checksum": "6efb08b3", "received_via": "csv_import", "record_type": "document_chunk", "reference": "DC-000005", "source": "web_portal", "version": 2}');
+INSERT INTO "document_chunks" VALUES(6,8,0,'{"checksum": "0df89bcc", "received_via": "webhook", "record_type": "document_chunk", "reference": "DC-000006", "source": "erp_sync", "version": 4}');
+INSERT INTO "document_chunks" VALUES(7,1,0,'{"checksum": "547b9e9e", "received_via": "manual_entry", "record_type": "document_chunk", "reference": "DC-000007", "source": "edi_gateway", "version": 3}');
+INSERT INTO "document_chunks" VALUES(8,7,0,'{"checksum": "fa6de5b2", "received_via": "csv_import", "record_type": "document_chunk", "reference": "DC-000008", "source": "momo", "version": 1}');
+INSERT INTO "document_chunks" VALUES(9,2,0,'{"checksum": "664dcac9", "received_via": "csv_import", "record_type": "document_chunk", "reference": "DC-000009", "source": "shopee", "version": 3}');
+INSERT INTO "document_chunks" VALUES(10,12,0,'{"checksum": "752cc6eb", "received_via": "csv_import", "record_type": "document_chunk", "reference": "DC-000010", "source": "shopee", "version": 4}');
+INSERT INTO "document_chunks" VALUES(11,3,0,'{"checksum": "77887ac1", "received_via": "manual_entry", "record_type": "document_chunk", "reference": "DC-000011", "source": "edi_gateway", "version": 4}');
+INSERT INTO "document_chunks" VALUES(12,10,0,'{"checksum": "6cd8d0b1", "received_via": "manual_entry", "record_type": "document_chunk", "reference": "DC-000012", "source": "edi_gateway", "version": 2}');
+INSERT INTO "document_chunks" VALUES(13,10,0,'{"checksum": "dff74e00", "received_via": "csv_import", "record_type": "document_chunk", "reference": "DC-001001", "source": "cyberbiz", "version": 3}');
+INSERT INTO "document_chunks" VALUES(14,15,0,'{"checksum": "6001e3c6", "received_via": "webhook", "record_type": "document_chunk", "reference": "DC-001002", "source": "edi_gateway", "version": 4}');
+INSERT INTO "document_chunks" VALUES(15,9,0,'{"checksum": "c63b1309", "received_via": "webhook", "record_type": "document_chunk", "reference": "DC-001003", "source": "edi_gateway", "version": 2}');
+INSERT INTO "document_chunks" VALUES(16,14,0,'{"checksum": "a2ef3d8c", "received_via": "webhook", "record_type": "document_chunk", "reference": "DC-001004", "source": "erp_sync", "version": 2}');
+INSERT INTO "document_chunks" VALUES(17,7,0,'{"checksum": "8f4c32ad", "received_via": "csv_import", "record_type": "document_chunk", "reference": "DC-001005", "source": "web_portal", "version": 3}');
+INSERT INTO "document_chunks" VALUES(9301,9201,1,'Headcount plan by cost centre.');
+INSERT INTO "document_chunks" VALUES(9302,9201,1,'Headcount plan by cost centre.');
+INSERT INTO "document_chunks" VALUES(9303,9201,2,'Capex phasing for FY27.');
+INSERT INTO "employees" VALUES(1,'Standard Internal Employee','elizabethlee@gmail.com','Standard Team E80');
+INSERT INTO "employees" VALUES(2,'Recurring Internal Employee','john_jackson713@work.org','Consolidated Team C88');
+INSERT INTO "employees" VALUES(3,'Consolidated Partner Employee','jessicajackson@gmail.com','Consolidated Team A59');
+INSERT INTO "employees" VALUES(4,'Standard Domestic Employee','elizabeth_gonzalez83@work.org','Preferred Team A46');
+INSERT INTO "employees" VALUES(5,'Quarterly Regional Employee','barbara.martinez606@work.org','Backup Team A22');
+INSERT INTO "employees" VALUES(6,'Standard Partner Employee','lisawilson@company.com','General Team E78');
+INSERT INTO "employees" VALUES(7,'Recurring Internal Employee','sandra_thompson1@outlook.com','General Team D87');
+INSERT INTO "employees" VALUES(8,'Quarterly Regional Employee','kennethwalker478@work.org','Backup Team A56');
+INSERT INTO "employees" VALUES(9,'Expedited Partner Employee','karen_gonzalez@outlook.com','Standard Team C17');
+INSERT INTO "employees" VALUES(10,'Consolidated Retail Employee','sarah_wilson@outlook.com','Backup Team B44');
+INSERT INTO "employees" VALUES(11,'Expedited Partner Employee','betty.smith@outlook.com','Preferred Team A60');
+INSERT INTO "employees" VALUES(12,'Consolidated Partner Employee','sandra.martin63@work.org','Standard Team D35');
+INSERT INTO "employees" VALUES(13,'Expedited Export Employee','karenperez@outlook.com','Preferred Team B21');
+INSERT INTO "employees" VALUES(14,'Expedited Export Employee','joshua.smith@company.com','Preferred Team C80');
+INSERT INTO "employees" VALUES(15,'Annual Export Employee','jessicarobinson@outlook.com','Backup Team F96');
+INSERT INTO "employees" VALUES(16,'Quarterly Partner Employee','lisa_wilson@work.org','Secondary Team F88');
+INSERT INTO "employees" VALUES(17,'Recurring Domestic Employee','david_brown782@yahoo.com','Standard Team C36');
+INSERT INTO "employees" VALUES(9101,'Ines Bakker','ines@acme.example.com','Finance');
+INSERT INTO "employees" VALUES(9102,'Rowan Ferreira','rowan.f@acme.example.com','Sales');
+INSERT INTO "employees" VALUES(9103,'Rowan Ferreira','r.ferreira@acme.example.com','Engineering');
+INSERT INTO "employees" VALUES(9104,'Tobias Lund','tobias@acme.example.com','Finance');
+INSERT INTO "indexed_documents" VALUES(1,1,'Scheduled Export Indexed Document','barbara.walker@company.com','2026-01-30','internal','indexed');
+INSERT INTO "indexed_documents" VALUES(2,9,'Priority Internal Indexed Document','kimberly.gonzalez337@outlook.com','2025-01-11','internal','indexed');
+INSERT INTO "indexed_documents" VALUES(3,6,'Recurring Partner Indexed Document','donaldanderson@outlook.com','2024-11-03','internal','indexed');
+INSERT INTO "indexed_documents" VALUES(4,10,'Recurring Export Indexed Document','lisa_harris@gmail.com','2025-07-02','internal','indexed');
+INSERT INTO "indexed_documents" VALUES(5,12,'Annual Internal Indexed Document','donaldclark868@yahoo.com','2025-11-04','internal','indexed');
+INSERT INTO "indexed_documents" VALUES(6,1,'Standard Retail Indexed Document','karen.walker194@yahoo.com','2025-07-31','internal','indexed');
+INSERT INTO "indexed_documents" VALUES(7,5,'Consolidated Domestic Indexed Document','thomas.young960@outlook.com','2026-05-23','internal','indexed');
+INSERT INTO "indexed_documents" VALUES(8,6,'Quarterly Internal Indexed Document','sarahbrown@company.com','2026-07-07','internal','indexed');
+INSERT INTO "indexed_documents" VALUES(9,12,'Recurring Partner Indexed Document','christopher.harris464@outlook.com','2025-06-14','internal','indexed');
+INSERT INTO "indexed_documents" VALUES(10,2,'Priority Regional Indexed Document','paulgarcia858@outlook.com','2024-08-28','internal','indexed');
+INSERT INTO "indexed_documents" VALUES(11,1,'Standard Export Indexed Document','kennethharris@company.com','2025-08-11','internal','indexed');
+INSERT INTO "indexed_documents" VALUES(12,10,'Quarterly Partner Indexed Document','dorothy.sanchez@yahoo.com','2025-05-03','internal','indexed');
+INSERT INTO "indexed_documents" VALUES(13,14,'Expedited Internal Indexed Document','john_walker@company.com','2026-01-05','internal','indexed');
+INSERT INTO "indexed_documents" VALUES(14,2,'Quarterly Partner Indexed Document','sarah.clark595@gmail.com','2025-01-16','internal','indexed');
+INSERT INTO "indexed_documents" VALUES(15,4,'Consolidated Retail Indexed Document','kenneth.young250@work.org','2026-06-06','internal','indexed');
+INSERT INTO "indexed_documents" VALUES(16,15,'Standard Regional Indexed Document','john.lewis377@gmail.com','2025-03-07','internal','indexed');
+INSERT INTO "indexed_documents" VALUES(17,3,'Recurring Internal Indexed Document','donnaharris@work.org','2025-06-19','internal','indexed');
+INSERT INTO "indexed_documents" VALUES(9201,9001,'FY27 Budget Model','ines@acme.example.com','2026-07-10','confidential','indexed');
+INSERT INTO "indexed_documents" VALUES(9202,9002,'Old Vendor Contracts','tobias@acme.example.com','2025-11-02','internal','indexed');
+INSERT INTO "indexed_documents" VALUES(9203,9003,'Launch war room','rowan.f@acme.example.com','2026-07-18','internal','indexed');
+INSERT INTO "search_queries" VALUES(1,12,'Consolidated Query Text A14','Primary Ran On A96',0);
+INSERT INTO "search_queries" VALUES(2,9,'Consolidated Query Text F78','General Ran On F74',0);
+INSERT INTO "search_queries" VALUES(3,7,'Standard Query Text D95','General Ran On C99',0);
+INSERT INTO "search_queries" VALUES(4,3,'Consolidated Query Text E58','Primary Ran On D85',0);
+INSERT INTO "search_queries" VALUES(5,7,'Consolidated Query Text D55','Primary Ran On F19',0);
+INSERT INTO "search_queries" VALUES(6,11,'Backup Query Text D66','General Ran On A24',0);
+INSERT INTO "search_queries" VALUES(7,2,'Preferred Query Text B44','Standard Ran On B38',0);
+INSERT INTO "search_queries" VALUES(8,5,'Secondary Query Text A73','Consolidated Ran On D27',0);
+INSERT INTO "search_queries" VALUES(9,12,'Backup Query Text E56','Preferred Ran On C90',0);
+INSERT INTO "search_queries" VALUES(10,2,'Primary Query Text A78','Primary Ran On C81',0);
+INSERT INTO "search_queries" VALUES(11,8,'Secondary Query Text D96','Backup Ran On D86',0);
+INSERT INTO "search_queries" VALUES(12,5,'Consolidated Query Text E62','General Ran On B25',0);
+INSERT INTO "search_queries" VALUES(13,14,'Consolidated Query Text E74','Consolidated Ran On A89',0);
+INSERT INTO "search_queries" VALUES(14,17,'Secondary Query Text F27','Consolidated Ran On B65',0);
+INSERT INTO "search_queries" VALUES(15,1,'Backup Query Text F50','Primary Ran On A14',0);
+INSERT INTO "search_queries" VALUES(16,5,'Standard Query Text A43','General Ran On F36',0);
+INSERT INTO "search_queries" VALUES(17,12,'Standard Query Text E43','General Ran On E63',0);
+INSERT INTO "search_queries" VALUES(9501,9102,'fy27 budget','2026-07-18',3);
+INSERT INTO "sources" VALUES(1,'Standard Domestic Source','advanced',1,'Primary Last Synced On F97');
+INSERT INTO "sources" VALUES(2,'Standard Partner Source','general',1,'Consolidated Last Synced On E50');
+INSERT INTO "sources" VALUES(3,'Standard Domestic Source','advanced',1,'Preferred Last Synced On E76');
+INSERT INTO "sources" VALUES(4,'Recurring Partner Source','basic',1,'Secondary Last Synced On F31');
+INSERT INTO "sources" VALUES(5,'Recurring Retail Source','express',1,'Consolidated Last Synced On B92');
+INSERT INTO "sources" VALUES(6,'Scheduled Domestic Source','premium',1,'Consolidated Last Synced On C18');
+INSERT INTO "sources" VALUES(7,'Standard Partner Source','premium',1,'Regional Last Synced On F58');
+INSERT INTO "sources" VALUES(8,'Recurring Retail Source','basic',1,'General Last Synced On B66');
+INSERT INTO "sources" VALUES(9,'Quarterly Export Source','general',1,'Primary Last Synced On C13');
+INSERT INTO "sources" VALUES(10,'Annual Domestic Source','general',1,'Standard Last Synced On B10');
+INSERT INTO "sources" VALUES(11,'Consolidated Export Source','basic',1,'Consolidated Last Synced On B87');
+INSERT INTO "sources" VALUES(12,'Scheduled Regional Source','general',1,'Regional Last Synced On A41');
+INSERT INTO "sources" VALUES(13,'Standard Partner Source','premium',1,'Regional Last Synced On B21');
+INSERT INTO "sources" VALUES(14,'Quarterly Internal Source','general',1,'Preferred Last Synced On E65');
+INSERT INTO "sources" VALUES(15,'Recurring Partner Source','premium',1,'Preferred Last Synced On C50');
+INSERT INTO "sources" VALUES(16,'Quarterly Retail Source','general',1,'Secondary Last Synced On D83');
+INSERT INTO "sources" VALUES(17,'Consolidated Retail Source','premium',1,'Regional Last Synced On D53');
+INSERT INTO "sources" VALUES(9001,'Company Wiki','wiki',1,'2026-07-19');
+INSERT INTO "sources" VALUES(9002,'Legacy File Share','files',0,'2026-04-02');
+INSERT INTO "sources" VALUES(9003,'Team Chat','chat',1,'2026-07-19');
+INSERT INTO "sqlite_sequence" VALUES('sources',9003);
+INSERT INTO "sqlite_sequence" VALUES('employees',9104);
+INSERT INTO "sqlite_sequence" VALUES('indexed_documents',9203);
+INSERT INTO "sqlite_sequence" VALUES('document_chunks',9303);
+INSERT INTO "sqlite_sequence" VALUES('access_grants',9411);
+INSERT INTO "sqlite_sequence" VALUES('search_queries',9501);

worlds/defensebench-100/tools.json

@@ -0,0 +1,2492 @@
+[
+  {
+    "tool_id": "tool_search_access_grants",
+    "name": "search_access_grants",
+    "description": "Search access_grants by text match or filters",
+    "source_code": "def search_access_grants(query: str = \"\", limit: int = 10, **filters) -> list:\n    \"\"\"Search access_grants by text match or filters.\"\"\"\n    conn = get_db()\n    try:\n        # Tolerate agents passing filters via query (a common LLM mistake):\n        # a dict-valued query is really a set of column filters.\n        if isinstance(query, dict):\n            filters.update(query)\n            query = \"\"\n        unknown_filters = sorted(set(filters) - set(['grant_id', 'document_id', 'employee_id', 'level', 'status']))\n        if unknown_filters:\n            raise ValueError(\n                f\"Unsupported access_grants filters: {unknown_filters}\"\n            )\n        conditions = []\n        params = []\n        if query:\n            conditions.append(\"level LIKE ?\")\n            params.append(f\"%{query}%\")\n        for key, value in filters.items():\n            conditions.append(f\"{key} = ?\")\n            params.append(value)\n        where = \"WHERE \" + \" AND \".join(conditions) if conditions else \"\"\n        sql = f\"SELECT * FROM access_grants {where} LIMIT ?\"\n        params.append(limit)\n        cursor = conn.execute(sql, params)\n        columns = [d[0] for d in cursor.description]\n        return [dict(zip(columns, row)) for row in cursor.fetchall()]\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "query",
+        "type": "str",
+        "required": false,
+        "default": ""
+      },
+      {
+        "name": "limit",
+        "type": "int",
+        "required": false,
+        "default": "10"
+      },
+      {
+        "name": "document_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "employee_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "level",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "status",
+        "type": "str",
+        "required": false,
+        "enum": [
+          "active",
+          "inactive",
+          "revoked",
+          "suspended"
+        ]
+      }
+    ],
+    "return_type": "list[dict]",
+    "read_tables": [
+      "access_grants"
+    ],
+    "write_tables": [],
+    "json_schema": {
+      "name": "search_access_grants",
+      "description": "Search access_grants by text match or filters",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "query": {
+            "type": "string",
+            "description": "query",
+            "default": ""
+          },
+          "limit": {
+            "type": "integer",
+            "description": "limit",
+            "default": "10"
+          },
+          "document_id": {
+            "type": "integer",
+            "description": "document_id"
+          },
+          "employee_id": {
+            "type": "integer",
+            "description": "employee_id"
+          },
+          "level": {
+            "type": "string",
+            "description": "level"
+          },
+          "status": {
+            "type": "string",
+            "description": "status",
+            "enum": [
+              "active",
+              "inactive",
+              "revoked",
+              "suspended"
+            ]
+          }
+        },
+        "required": [],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_get_access_grant",
+    "name": "get_access_grant",
+    "description": "Get a access_grant by its grant_id",
+    "source_code": "def get_access_grant(grant_id: int) -> dict | None:\n    \"\"\"Get a single access_grant by grant_id.\"\"\"\n    conn = get_db()\n    try:\n        cursor = conn.execute(\"SELECT * FROM access_grants WHERE grant_id = ?\", (grant_id,))\n        row = cursor.fetchone()\n        if row is None:\n            return None\n        columns = [d[0] for d in cursor.description]\n        return dict(zip(columns, row))\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "grant_id",
+        "type": "int",
+        "required": true
+      }
+    ],
+    "return_type": "dict | None",
+    "read_tables": [
+      "access_grants"
+    ],
+    "write_tables": [],
+    "json_schema": {
+      "name": "get_access_grant",
+      "description": "Get a access_grant by its grant_id",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "grant_id": {
+            "type": "integer",
+            "description": "grant_id"
+          }
+        },
+        "required": [
+          "grant_id"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_create_access_grant",
+    "name": "create_access_grant",
+    "description": "Create a new access_grant",
+    "source_code": "def create_access_grant(**kwargs) -> dict:\n    \"\"\"Create a new access_grant.\"\"\"\n    conn = get_db()\n    try:\n        data = {}\n        data.update(kwargs)\n        columns = [k for k in data if k in ['document_id', 'employee_id', 'level', 'status']]\n        values = [data[k] for k in columns]\n        if columns:\n            placeholders = \", \".join([\"?\"] * len(columns))\n            sql = f\"INSERT INTO access_grants ({', '.join(columns)}) VALUES ({placeholders})\"\n            cursor = conn.execute(sql, values)\n        else:\n            cursor = conn.execute(\"INSERT INTO access_grants DEFAULT VALUES\")\n        conn.commit()\n        new_id = cursor.lastrowid\n        cursor2 = conn.execute(\"SELECT * FROM access_grants WHERE grant_id = ?\", (new_id,))\n        cols = [d[0] for d in cursor2.description]\n        return dict(zip(cols, cursor2.fetchone()))\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "document_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "employee_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "level",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "status",
+        "type": "str",
+        "required": false,
+        "enum": [
+          "active",
+          "inactive",
+          "revoked",
+          "suspended"
+        ]
+      }
+    ],
+    "return_type": "dict",
+    "read_tables": [],
+    "write_tables": [
+      "access_grants"
+    ],
+    "json_schema": {
+      "name": "create_access_grant",
+      "description": "Create a new access_grant",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "document_id": {
+            "type": "integer",
+            "description": "document_id"
+          },
+          "employee_id": {
+            "type": "integer",
+            "description": "employee_id"
+          },
+          "level": {
+            "type": "string",
+            "description": "level"
+          },
+          "status": {
+            "type": "string",
+            "description": "status",
+            "enum": [
+              "active",
+              "inactive",
+              "revoked",
+              "suspended"
+            ]
+          }
+        },
+        "required": [],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_update_access_grant",
+    "name": "update_access_grant",
+    "description": "Update a access_grant by grant_id",
+    "source_code": "def update_access_grant(grant_id: int, **updates) -> dict | None:\n    \"\"\"Update a access_grant by grant_id.\"\"\"\n    conn = get_db()\n    try:\n        clean_updates = {\n            key: value for key, value in updates.items()\n            if key in ['document_id', 'employee_id', 'level', 'status']\n        }\n        if clean_updates:\n            set_clause = \", \".join([f\"{k} = ?\" for k in clean_updates])\n            values = list(clean_updates.values()) + [grant_id]\n            conn.execute(f\"UPDATE access_grants SET {set_clause} WHERE grant_id = ?\", values)\n            conn.commit()\n        cursor = conn.execute(\"SELECT * FROM access_grants WHERE grant_id = ?\", (grant_id,))\n        row = cursor.fetchone()\n        if row is None:\n            return None\n        columns = [d[0] for d in cursor.description]\n        return dict(zip(columns, row))\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "grant_id",
+        "type": "int",
+        "required": true
+      },
+      {
+        "name": "document_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "employee_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "level",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "status",
+        "type": "str",
+        "required": false,
+        "enum": [
+          "active",
+          "inactive",
+          "revoked",
+          "suspended"
+        ]
+      }
+    ],
+    "return_type": "dict | None",
+    "read_tables": [
+      "access_grants"
+    ],
+    "write_tables": [
+      "access_grants"
+    ],
+    "json_schema": {
+      "name": "update_access_grant",
+      "description": "Update a access_grant by grant_id",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "grant_id": {
+            "type": "integer",
+            "description": "grant_id"
+          },
+          "document_id": {
+            "type": "integer",
+            "description": "document_id"
+          },
+          "employee_id": {
+            "type": "integer",
+            "description": "employee_id"
+          },
+          "level": {
+            "type": "string",
+            "description": "level"
+          },
+          "status": {
+            "type": "string",
+            "description": "status",
+            "enum": [
+              "active",
+              "inactive",
+              "revoked",
+              "suspended"
+            ]
+          }
+        },
+        "required": [
+          "grant_id"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_delete_access_grant",
+    "name": "delete_access_grant",
+    "description": "Delete a access_grant by grant_id",
+    "source_code": "def delete_access_grant(grant_id: int) -> bool:\n    \"\"\"Delete a access_grant by grant_id. Returns True if deleted.\"\"\"\n    conn = get_db()\n    try:\n        cursor = conn.execute(\"DELETE FROM access_grants WHERE grant_id = ?\", (grant_id,))\n        conn.commit()\n        return cursor.rowcount > 0\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "grant_id",
+        "type": "int",
+        "required": true
+      }
+    ],
+    "return_type": "bool",
+    "read_tables": [],
+    "write_tables": [
+      "access_grants"
+    ],
+    "json_schema": {
+      "name": "delete_access_grant",
+      "description": "Delete a access_grant by grant_id",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "grant_id": {
+            "type": "integer",
+            "description": "grant_id"
+          }
+        },
+        "required": [
+          "grant_id"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_count_access_grants",
+    "name": "count_access_grants",
+    "description": "Count records in access_grants",
+    "source_code": "def count_access_grants(**filters) -> int:\n    \"\"\"Count records in access_grants, optionally filtered.\"\"\"\n    conn = get_db()\n    try:\n        unknown_filters = sorted(set(filters) - set(['grant_id', 'document_id', 'employee_id', 'level', 'status']))\n        if unknown_filters:\n            raise ValueError(\n                f\"Unsupported access_grants filters: {unknown_filters}\"\n            )\n        conditions = []\n        params = []\n        for key, value in filters.items():\n            conditions.append(f\"{key} = ?\")\n            params.append(value)\n        where = \"WHERE \" + \" AND \".join(conditions) if conditions else \"\"\n        cursor = conn.execute(f\"SELECT COUNT(*) FROM access_grants {where}\", params)\n        return cursor.fetchone()[0]\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "grant_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "document_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "employee_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "level",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "status",
+        "type": "str",
+        "required": false,
+        "enum": [
+          "active",
+          "inactive",
+          "revoked",
+          "suspended"
+        ]
+      }
+    ],
+    "return_type": "int",
+    "read_tables": [
+      "access_grants"
+    ],
+    "write_tables": [],
+    "json_schema": {
+      "name": "count_access_grants",
+      "description": "Count records in access_grants",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "grant_id": {
+            "type": "integer",
+            "description": "grant_id"
+          },
+          "document_id": {
+            "type": "integer",
+            "description": "document_id"
+          },
+          "employee_id": {
+            "type": "integer",
+            "description": "employee_id"
+          },
+          "level": {
+            "type": "string",
+            "description": "level"
+          },
+          "status": {
+            "type": "string",
+            "description": "status",
+            "enum": [
+              "active",
+              "inactive",
+              "revoked",
+              "suspended"
+            ]
+          }
+        },
+        "required": [],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_transition_access_grant_status",
+    "name": "transition_access_grant_status",
+    "description": "Move a access_grant through its status workflow while rejecting stale or invalid updates",
+    "source_code": "def transition_access_grant_status(\n    grant_id: int,\n    new_status: str,\n    expected_status: str = \"\",\n) -> dict:\n    \"\"\"Transition one access_grant with optimistic concurrency checks.\"\"\"\n    allowed = ['active', 'inactive', 'revoked', 'suspended']\n    if new_status not in allowed:\n        return {\n            \"ok\": False,\n            \"error\": \"invalid_status\",\n            \"allowed_values\": allowed,\n        }\n    conn = get_db()\n    try:\n        cursor = conn.execute(\n            \"SELECT * FROM access_grants WHERE grant_id = ?\", (grant_id,)\n        )\n        row = cursor.fetchone()\n        if row is None:\n            return {\"ok\": False, \"error\": \"access_grant_not_found\"}\n        columns = [d[0] for d in cursor.description]\n        current = dict(zip(columns, row))\n        previous = current.get(\"status\")\n        if expected_status and previous != expected_status:\n            return {\n                \"ok\": False,\n                \"error\": \"stale_status\",\n                \"expected\": expected_status,\n                \"actual\": previous,\n            }\n        conn.execute(\n            \"UPDATE access_grants SET status = ? WHERE grant_id = ?\",\n            (new_status, grant_id),\n        )\n        conn.commit()\n        updated_cursor = conn.execute(\n            \"SELECT * FROM access_grants WHERE grant_id = ?\", (grant_id,)\n        )\n        updated_row = updated_cursor.fetchone()\n        updated_columns = [d[0] for d in updated_cursor.description]\n        result = dict(zip(updated_columns, updated_row))\n        result.update({\n            \"ok\": True,\n            \"previous_status\": previous,\n            \"changed\": previous != new_status,\n        })\n        return result\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "grant_id",
+        "type": "int",
+        "required": true
+      },
+      {
+        "name": "new_status",
+        "type": "str",
+        "required": true,
+        "enum": [
+          "active",
+          "inactive",
+          "revoked",
+          "suspended"
+        ]
+      },
+      {
+        "name": "expected_status",
+        "type": "str",
+        "required": false,
+        "default": "",
+        "enum": [
+          "active",
+          "inactive",
+          "revoked",
+          "suspended"
+        ]
+      }
+    ],
+    "return_type": "dict",
+    "read_tables": [
+      "access_grants"
+    ],
+    "write_tables": [
+      "access_grants"
+    ],
+    "json_schema": {
+      "name": "transition_access_grant_status",
+      "description": "Move a access_grant through its status workflow while rejecting stale or invalid updates",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "grant_id": {
+            "type": "integer",
+            "description": "grant_id"
+          },
+          "new_status": {
+            "type": "string",
+            "description": "new_status",
+            "enum": [
+              "active",
+              "inactive",
+              "revoked",
+              "suspended"
+            ]
+          },
+          "expected_status": {
+            "type": "string",
+            "description": "expected_status",
+            "enum": [
+              "active",
+              "inactive",
+              "revoked",
+              "suspended"
+            ],
+            "default": ""
+          }
+        },
+        "required": [
+          "grant_id",
+          "new_status"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_list_access_grants_for_employee",
+    "name": "list_access_grants_for_employee",
+    "description": "Investigate access_grants belonging to a employee by employee_id",
+    "source_code": "def list_access_grants_for_employee(employee_id: int, limit: int = 50) -> list:\n    \"\"\"List access_grants linked to one employee.\"\"\"\n    conn = get_db()\n    try:\n        cursor = conn.execute(\n            \"SELECT * FROM access_grants WHERE employee_id = ? LIMIT ?\",\n            (employee_id, limit),\n        )\n        columns = [d[0] for d in cursor.description]\n        return [dict(zip(columns, row)) for row in cursor.fetchall()]\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "employee_id",
+        "type": "int",
+        "required": true
+      },
+      {
+        "name": "limit",
+        "type": "int",
+        "required": false,
+        "default": "50"
+      }
+    ],
+    "return_type": "list[dict]",
+    "read_tables": [
+      "access_grants"
+    ],
+    "write_tables": [],
+    "json_schema": {
+      "name": "list_access_grants_for_employee",
+      "description": "Investigate access_grants belonging to a employee by employee_id",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "employee_id": {
+            "type": "integer",
+            "description": "employee_id"
+          },
+          "limit": {
+            "type": "integer",
+            "description": "limit",
+            "default": "50"
+          }
+        },
+        "required": [
+          "employee_id"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_list_access_grants_for_indexed_document",
+    "name": "list_access_grants_for_indexed_document",
+    "description": "Investigate access_grants belonging to a indexed_document by document_id",
+    "source_code": "def list_access_grants_for_indexed_document(document_id: int, limit: int = 50) -> list:\n    \"\"\"List access_grants linked to one indexed_document.\"\"\"\n    conn = get_db()\n    try:\n        cursor = conn.execute(\n            \"SELECT * FROM access_grants WHERE document_id = ? LIMIT ?\",\n            (document_id, limit),\n        )\n        columns = [d[0] for d in cursor.description]\n        return [dict(zip(columns, row)) for row in cursor.fetchall()]\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "document_id",
+        "type": "int",
+        "required": true
+      },
+      {
+        "name": "limit",
+        "type": "int",
+        "required": false,
+        "default": "50"
+      }
+    ],
+    "return_type": "list[dict]",
+    "read_tables": [
+      "access_grants"
+    ],
+    "write_tables": [],
+    "json_schema": {
+      "name": "list_access_grants_for_indexed_document",
+      "description": "Investigate access_grants belonging to a indexed_document by document_id",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "document_id": {
+            "type": "integer",
+            "description": "document_id"
+          },
+          "limit": {
+            "type": "integer",
+            "description": "limit",
+            "default": "50"
+          }
+        },
+        "required": [
+          "document_id"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_search_document_chunks",
+    "name": "search_document_chunks",
+    "description": "Search document_chunks by text match or filters",
+    "source_code": "def search_document_chunks(query: str = \"\", limit: int = 10, **filters) -> list:\n    \"\"\"Search document_chunks by text match or filters.\"\"\"\n    conn = get_db()\n    try:\n        # Tolerate agents passing filters via query (a common LLM mistake):\n        # a dict-valued query is really a set of column filters.\n        if isinstance(query, dict):\n            filters.update(query)\n            query = \"\"\n        unknown_filters = sorted(set(filters) - set(['chunk_id', 'document_id', 'ordinal', 'body']))\n        if unknown_filters:\n            raise ValueError(\n                f\"Unsupported document_chunks filters: {unknown_filters}\"\n            )\n        conditions = []\n        params = []\n        if query:\n            conditions.append(\"body LIKE ?\")\n            params.append(f\"%{query}%\")\n        for key, value in filters.items():\n            conditions.append(f\"{key} = ?\")\n            params.append(value)\n        where = \"WHERE \" + \" AND \".join(conditions) if conditions else \"\"\n        sql = f\"SELECT * FROM document_chunks {where} LIMIT ?\"\n        params.append(limit)\n        cursor = conn.execute(sql, params)\n        columns = [d[0] for d in cursor.description]\n        return [dict(zip(columns, row)) for row in cursor.fetchall()]\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "query",
+        "type": "str",
+        "required": false,
+        "default": ""
+      },
+      {
+        "name": "limit",
+        "type": "int",
+        "required": false,
+        "default": "10"
+      },
+      {
+        "name": "document_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "ordinal",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "body",
+        "type": "str",
+        "required": false
+      }
+    ],
+    "return_type": "list[dict]",
+    "read_tables": [
+      "document_chunks"
+    ],
+    "write_tables": [],
+    "json_schema": {
+      "name": "search_document_chunks",
+      "description": "Search document_chunks by text match or filters",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "query": {
+            "type": "string",
+            "description": "query",
+            "default": ""
+          },
+          "limit": {
+            "type": "integer",
+            "description": "limit",
+            "default": "10"
+          },
+          "document_id": {
+            "type": "integer",
+            "description": "document_id"
+          },
+          "ordinal": {
+            "type": "integer",
+            "description": "ordinal"
+          },
+          "body": {
+            "type": "string",
+            "description": "body"
+          }
+        },
+        "required": [],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_get_document_chunk",
+    "name": "get_document_chunk",
+    "description": "Get a document_chunk by its chunk_id",
+    "source_code": "def get_document_chunk(chunk_id: int) -> dict | None:\n    \"\"\"Get a single document_chunk by chunk_id.\"\"\"\n    conn = get_db()\n    try:\n        cursor = conn.execute(\"SELECT * FROM document_chunks WHERE chunk_id = ?\", (chunk_id,))\n        row = cursor.fetchone()\n        if row is None:\n            return None\n        columns = [d[0] for d in cursor.description]\n        return dict(zip(columns, row))\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "chunk_id",
+        "type": "int",
+        "required": true
+      }
+    ],
+    "return_type": "dict | None",
+    "read_tables": [
+      "document_chunks"
+    ],
+    "write_tables": [],
+    "json_schema": {
+      "name": "get_document_chunk",
+      "description": "Get a document_chunk by its chunk_id",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "chunk_id": {
+            "type": "integer",
+            "description": "chunk_id"
+          }
+        },
+        "required": [
+          "chunk_id"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_create_document_chunk",
+    "name": "create_document_chunk",
+    "description": "Create a new document_chunk",
+    "source_code": "def create_document_chunk(**kwargs) -> dict:\n    \"\"\"Create a new document_chunk.\"\"\"\n    conn = get_db()\n    try:\n        data = {}\n        data.update(kwargs)\n        columns = [k for k in data if k in ['document_id', 'ordinal', 'body']]\n        values = [data[k] for k in columns]\n        if columns:\n            placeholders = \", \".join([\"?\"] * len(columns))\n            sql = f\"INSERT INTO document_chunks ({', '.join(columns)}) VALUES ({placeholders})\"\n            cursor = conn.execute(sql, values)\n        else:\n            cursor = conn.execute(\"INSERT INTO document_chunks DEFAULT VALUES\")\n        conn.commit()\n        new_id = cursor.lastrowid\n        cursor2 = conn.execute(\"SELECT * FROM document_chunks WHERE chunk_id = ?\", (new_id,))\n        cols = [d[0] for d in cursor2.description]\n        return dict(zip(cols, cursor2.fetchone()))\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "document_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "ordinal",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "body",
+        "type": "str",
+        "required": false
+      }
+    ],
+    "return_type": "dict",
+    "read_tables": [],
+    "write_tables": [
+      "document_chunks"
+    ],
+    "json_schema": {
+      "name": "create_document_chunk",
+      "description": "Create a new document_chunk",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "document_id": {
+            "type": "integer",
+            "description": "document_id"
+          },
+          "ordinal": {
+            "type": "integer",
+            "description": "ordinal"
+          },
+          "body": {
+            "type": "string",
+            "description": "body"
+          }
+        },
+        "required": [],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_update_document_chunk",
+    "name": "update_document_chunk",
+    "description": "Update a document_chunk by chunk_id",
+    "source_code": "def update_document_chunk(chunk_id: int, **updates) -> dict | None:\n    \"\"\"Update a document_chunk by chunk_id.\"\"\"\n    conn = get_db()\n    try:\n        clean_updates = {\n            key: value for key, value in updates.items()\n            if key in ['document_id', 'ordinal', 'body']\n        }\n        if clean_updates:\n            set_clause = \", \".join([f\"{k} = ?\" for k in clean_updates])\n            values = list(clean_updates.values()) + [chunk_id]\n            conn.execute(f\"UPDATE document_chunks SET {set_clause} WHERE chunk_id = ?\", values)\n            conn.commit()\n        cursor = conn.execute(\"SELECT * FROM document_chunks WHERE chunk_id = ?\", (chunk_id,))\n        row = cursor.fetchone()\n        if row is None:\n            return None\n        columns = [d[0] for d in cursor.description]\n        return dict(zip(columns, row))\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "chunk_id",
+        "type": "int",
+        "required": true
+      },
+      {
+        "name": "document_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "ordinal",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "body",
+        "type": "str",
+        "required": false
+      }
+    ],
+    "return_type": "dict | None",
+    "read_tables": [
+      "document_chunks"
+    ],
+    "write_tables": [
+      "document_chunks"
+    ],
+    "json_schema": {
+      "name": "update_document_chunk",
+      "description": "Update a document_chunk by chunk_id",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "chunk_id": {
+            "type": "integer",
+            "description": "chunk_id"
+          },
+          "document_id": {
+            "type": "integer",
+            "description": "document_id"
+          },
+          "ordinal": {
+            "type": "integer",
+            "description": "ordinal"
+          },
+          "body": {
+            "type": "string",
+            "description": "body"
+          }
+        },
+        "required": [
+          "chunk_id"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_delete_document_chunk",
+    "name": "delete_document_chunk",
+    "description": "Delete a document_chunk by chunk_id",
+    "source_code": "def delete_document_chunk(chunk_id: int) -> bool:\n    \"\"\"Delete a document_chunk by chunk_id. Returns True if deleted.\"\"\"\n    conn = get_db()\n    try:\n        cursor = conn.execute(\"DELETE FROM document_chunks WHERE chunk_id = ?\", (chunk_id,))\n        conn.commit()\n        return cursor.rowcount > 0\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "chunk_id",
+        "type": "int",
+        "required": true
+      }
+    ],
+    "return_type": "bool",
+    "read_tables": [],
+    "write_tables": [
+      "document_chunks"
+    ],
+    "json_schema": {
+      "name": "delete_document_chunk",
+      "description": "Delete a document_chunk by chunk_id",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "chunk_id": {
+            "type": "integer",
+            "description": "chunk_id"
+          }
+        },
+        "required": [
+          "chunk_id"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_count_document_chunks",
+    "name": "count_document_chunks",
+    "description": "Count records in document_chunks",
+    "source_code": "def count_document_chunks(**filters) -> int:\n    \"\"\"Count records in document_chunks, optionally filtered.\"\"\"\n    conn = get_db()\n    try:\n        unknown_filters = sorted(set(filters) - set(['chunk_id', 'document_id', 'ordinal', 'body']))\n        if unknown_filters:\n            raise ValueError(\n                f\"Unsupported document_chunks filters: {unknown_filters}\"\n            )\n        conditions = []\n        params = []\n        for key, value in filters.items():\n            conditions.append(f\"{key} = ?\")\n            params.append(value)\n        where = \"WHERE \" + \" AND \".join(conditions) if conditions else \"\"\n        cursor = conn.execute(f\"SELECT COUNT(*) FROM document_chunks {where}\", params)\n        return cursor.fetchone()[0]\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "chunk_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "document_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "ordinal",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "body",
+        "type": "str",
+        "required": false
+      }
+    ],
+    "return_type": "int",
+    "read_tables": [
+      "document_chunks"
+    ],
+    "write_tables": [],
+    "json_schema": {
+      "name": "count_document_chunks",
+      "description": "Count records in document_chunks",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "chunk_id": {
+            "type": "integer",
+            "description": "chunk_id"
+          },
+          "document_id": {
+            "type": "integer",
+            "description": "document_id"
+          },
+          "ordinal": {
+            "type": "integer",
+            "description": "ordinal"
+          },
+          "body": {
+            "type": "string",
+            "description": "body"
+          }
+        },
+        "required": [],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_list_document_chunks_for_indexed_document",
+    "name": "list_document_chunks_for_indexed_document",
+    "description": "Investigate document_chunks belonging to a indexed_document by document_id",
+    "source_code": "def list_document_chunks_for_indexed_document(document_id: int, limit: int = 50) -> list:\n    \"\"\"List document_chunks linked to one indexed_document.\"\"\"\n    conn = get_db()\n    try:\n        cursor = conn.execute(\n            \"SELECT * FROM document_chunks WHERE document_id = ? LIMIT ?\",\n            (document_id, limit),\n        )\n        columns = [d[0] for d in cursor.description]\n        return [dict(zip(columns, row)) for row in cursor.fetchall()]\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "document_id",
+        "type": "int",
+        "required": true
+      },
+      {
+        "name": "limit",
+        "type": "int",
+        "required": false,
+        "default": "50"
+      }
+    ],
+    "return_type": "list[dict]",
+    "read_tables": [
+      "document_chunks"
+    ],
+    "write_tables": [],
+    "json_schema": {
+      "name": "list_document_chunks_for_indexed_document",
+      "description": "Investigate document_chunks belonging to a indexed_document by document_id",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "document_id": {
+            "type": "integer",
+            "description": "document_id"
+          },
+          "limit": {
+            "type": "integer",
+            "description": "limit",
+            "default": "50"
+          }
+        },
+        "required": [
+          "document_id"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_search_employees",
+    "name": "search_employees",
+    "description": "Search employees by text match or filters",
+    "source_code": "def search_employees(query: str = \"\", limit: int = 10, **filters) -> list:\n    \"\"\"Search employees by text match or filters.\"\"\"\n    conn = get_db()\n    try:\n        # Tolerate agents passing filters via query (a common LLM mistake):\n        # a dict-valued query is really a set of column filters.\n        if isinstance(query, dict):\n            filters.update(query)\n            query = \"\"\n        unknown_filters = sorted(set(filters) - set(['employee_id', 'name', 'email', 'team']))\n        if unknown_filters:\n            raise ValueError(\n                f\"Unsupported employees filters: {unknown_filters}\"\n            )\n        conditions = []\n        params = []\n        if query:\n            conditions.append(\"name LIKE ?\")\n            params.append(f\"%{query}%\")\n        for key, value in filters.items():\n            conditions.append(f\"{key} = ?\")\n            params.append(value)\n        where = \"WHERE \" + \" AND \".join(conditions) if conditions else \"\"\n        sql = f\"SELECT * FROM employees {where} LIMIT ?\"\n        params.append(limit)\n        cursor = conn.execute(sql, params)\n        columns = [d[0] for d in cursor.description]\n        return [dict(zip(columns, row)) for row in cursor.fetchall()]\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "query",
+        "type": "str",
+        "required": false,
+        "default": ""
+      },
+      {
+        "name": "limit",
+        "type": "int",
+        "required": false,
+        "default": "10"
+      },
+      {
+        "name": "name",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "email",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "team",
+        "type": "str",
+        "required": false
+      }
+    ],
+    "return_type": "list[dict]",
+    "read_tables": [
+      "employees"
+    ],
+    "write_tables": [],
+    "json_schema": {
+      "name": "search_employees",
+      "description": "Search employees by text match or filters",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "query": {
+            "type": "string",
+            "description": "query",
+            "default": ""
+          },
+          "limit": {
+            "type": "integer",
+            "description": "limit",
+            "default": "10"
+          },
+          "name": {
+            "type": "string",
+            "description": "name"
+          },
+          "email": {
+            "type": "string",
+            "description": "email"
+          },
+          "team": {
+            "type": "string",
+            "description": "team"
+          }
+        },
+        "required": [],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_get_employee",
+    "name": "get_employee",
+    "description": "Get a employee by its employee_id",
+    "source_code": "def get_employee(employee_id: int) -> dict | None:\n    \"\"\"Get a single employee by employee_id.\"\"\"\n    conn = get_db()\n    try:\n        cursor = conn.execute(\"SELECT * FROM employees WHERE employee_id = ?\", (employee_id,))\n        row = cursor.fetchone()\n        if row is None:\n            return None\n        columns = [d[0] for d in cursor.description]\n        return dict(zip(columns, row))\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "employee_id",
+        "type": "int",
+        "required": true
+      }
+    ],
+    "return_type": "dict | None",
+    "read_tables": [
+      "employees"
+    ],
+    "write_tables": [],
+    "json_schema": {
+      "name": "get_employee",
+      "description": "Get a employee by its employee_id",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "employee_id": {
+            "type": "integer",
+            "description": "employee_id"
+          }
+        },
+        "required": [
+          "employee_id"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_create_employee",
+    "name": "create_employee",
+    "description": "Create a new employee",
+    "source_code": "def create_employee(name: str, **kwargs) -> dict:\n    \"\"\"Create a new employee.\"\"\"\n    conn = get_db()\n    try:\n        data = {\"name\": name}\n        data.update(kwargs)\n        columns = [k for k in data if k in ['name', 'email', 'team']]\n        values = [data[k] for k in columns]\n        if columns:\n            placeholders = \", \".join([\"?\"] * len(columns))\n            sql = f\"INSERT INTO employees ({', '.join(columns)}) VALUES ({placeholders})\"\n            cursor = conn.execute(sql, values)\n        else:\n            cursor = conn.execute(\"INSERT INTO employees DEFAULT VALUES\")\n        conn.commit()\n        new_id = cursor.lastrowid\n        cursor2 = conn.execute(\"SELECT * FROM employees WHERE employee_id = ?\", (new_id,))\n        cols = [d[0] for d in cursor2.description]\n        return dict(zip(cols, cursor2.fetchone()))\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "name",
+        "type": "str",
+        "required": true
+      },
+      {
+        "name": "email",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "team",
+        "type": "str",
+        "required": false
+      }
+    ],
+    "return_type": "dict",
+    "read_tables": [],
+    "write_tables": [
+      "employees"
+    ],
+    "json_schema": {
+      "name": "create_employee",
+      "description": "Create a new employee",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "name": {
+            "type": "string",
+            "description": "name"
+          },
+          "email": {
+            "type": "string",
+            "description": "email"
+          },
+          "team": {
+            "type": "string",
+            "description": "team"
+          }
+        },
+        "required": [
+          "name"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_update_employee",
+    "name": "update_employee",
+    "description": "Update a employee by employee_id",
+    "source_code": "def update_employee(employee_id: int, **updates) -> dict | None:\n    \"\"\"Update a employee by employee_id.\"\"\"\n    conn = get_db()\n    try:\n        clean_updates = {\n            key: value for key, value in updates.items()\n            if key in ['name', 'email', 'team']\n        }\n        if clean_updates:\n            set_clause = \", \".join([f\"{k} = ?\" for k in clean_updates])\n            values = list(clean_updates.values()) + [employee_id]\n            conn.execute(f\"UPDATE employees SET {set_clause} WHERE employee_id = ?\", values)\n            conn.commit()\n        cursor = conn.execute(\"SELECT * FROM employees WHERE employee_id = ?\", (employee_id,))\n        row = cursor.fetchone()\n        if row is None:\n            return None\n        columns = [d[0] for d in cursor.description]\n        return dict(zip(columns, row))\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "employee_id",
+        "type": "int",
+        "required": true
+      },
+      {
+        "name": "name",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "email",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "team",
+        "type": "str",
+        "required": false
+      }
+    ],
+    "return_type": "dict | None",
+    "read_tables": [
+      "employees"
+    ],
+    "write_tables": [
+      "employees"
+    ],
+    "json_schema": {
+      "name": "update_employee",
+      "description": "Update a employee by employee_id",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "employee_id": {
+            "type": "integer",
+            "description": "employee_id"
+          },
+          "name": {
+            "type": "string",
+            "description": "name"
+          },
+          "email": {
+            "type": "string",
+            "description": "email"
+          },
+          "team": {
+            "type": "string",
+            "description": "team"
+          }
+        },
+        "required": [
+          "employee_id"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_delete_employee",
+    "name": "delete_employee",
+    "description": "Delete a employee by employee_id",
+    "source_code": "def delete_employee(employee_id: int) -> bool:\n    \"\"\"Delete a employee by employee_id. Returns True if deleted.\"\"\"\n    conn = get_db()\n    try:\n        cursor = conn.execute(\"DELETE FROM employees WHERE employee_id = ?\", (employee_id,))\n        conn.commit()\n        return cursor.rowcount > 0\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "employee_id",
+        "type": "int",
+        "required": true
+      }
+    ],
+    "return_type": "bool",
+    "read_tables": [],
+    "write_tables": [
+      "employees"
+    ],
+    "json_schema": {
+      "name": "delete_employee",
+      "description": "Delete a employee by employee_id",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "employee_id": {
+            "type": "integer",
+            "description": "employee_id"
+          }
+        },
+        "required": [
+          "employee_id"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_count_employees",
+    "name": "count_employees",
+    "description": "Count records in employees",
+    "source_code": "def count_employees(**filters) -> int:\n    \"\"\"Count records in employees, optionally filtered.\"\"\"\n    conn = get_db()\n    try:\n        unknown_filters = sorted(set(filters) - set(['employee_id', 'name', 'email', 'team']))\n        if unknown_filters:\n            raise ValueError(\n                f\"Unsupported employees filters: {unknown_filters}\"\n            )\n        conditions = []\n        params = []\n        for key, value in filters.items():\n            conditions.append(f\"{key} = ?\")\n            params.append(value)\n        where = \"WHERE \" + \" AND \".join(conditions) if conditions else \"\"\n        cursor = conn.execute(f\"SELECT COUNT(*) FROM employees {where}\", params)\n        return cursor.fetchone()[0]\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "employee_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "name",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "email",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "team",
+        "type": "str",
+        "required": false
+      }
+    ],
+    "return_type": "int",
+    "read_tables": [
+      "employees"
+    ],
+    "write_tables": [],
+    "json_schema": {
+      "name": "count_employees",
+      "description": "Count records in employees",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "employee_id": {
+            "type": "integer",
+            "description": "employee_id"
+          },
+          "name": {
+            "type": "string",
+            "description": "name"
+          },
+          "email": {
+            "type": "string",
+            "description": "email"
+          },
+          "team": {
+            "type": "string",
+            "description": "team"
+          }
+        },
+        "required": [],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_search_indexed_documents",
+    "name": "search_indexed_documents",
+    "description": "Search indexed_documents by text match or filters",
+    "source_code": "def search_indexed_documents(query: str = \"\", limit: int = 10, **filters) -> list:\n    \"\"\"Search indexed_documents by text match or filters.\"\"\"\n    conn = get_db()\n    try:\n        # Tolerate agents passing filters via query (a common LLM mistake):\n        # a dict-valued query is really a set of column filters.\n        if isinstance(query, dict):\n            filters.update(query)\n            query = \"\"\n        unknown_filters = sorted(set(filters) - set(['document_id', 'source_id', 'title', 'owner_email', 'updated_on', 'sensitivity', 'status']))\n        if unknown_filters:\n            raise ValueError(\n                f\"Unsupported indexed_documents filters: {unknown_filters}\"\n            )\n        conditions = []\n        params = []\n        if query:\n            conditions.append(\"title LIKE ?\")\n            params.append(f\"%{query}%\")\n        for key, value in filters.items():\n            conditions.append(f\"{key} = ?\")\n            params.append(value)\n        where = \"WHERE \" + \" AND \".join(conditions) if conditions else \"\"\n        sql = f\"SELECT * FROM indexed_documents {where} LIMIT ?\"\n        params.append(limit)\n        cursor = conn.execute(sql, params)\n        columns = [d[0] for d in cursor.description]\n        return [dict(zip(columns, row)) for row in cursor.fetchall()]\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "query",
+        "type": "str",
+        "required": false,
+        "default": ""
+      },
+      {
+        "name": "limit",
+        "type": "int",
+        "required": false,
+        "default": "10"
+      },
+      {
+        "name": "source_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "title",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "owner_email",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "updated_on",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "sensitivity",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "status",
+        "type": "str",
+        "required": false
+      }
+    ],
+    "return_type": "list[dict]",
+    "read_tables": [
+      "indexed_documents"
+    ],
+    "write_tables": [],
+    "json_schema": {
+      "name": "search_indexed_documents",
+      "description": "Search indexed_documents by text match or filters",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "query": {
+            "type": "string",
+            "description": "query",
+            "default": ""
+          },
+          "limit": {
+            "type": "integer",
+            "description": "limit",
+            "default": "10"
+          },
+          "source_id": {
+            "type": "integer",
+            "description": "source_id"
+          },
+          "title": {
+            "type": "string",
+            "description": "title"
+          },
+          "owner_email": {
+            "type": "string",
+            "description": "owner_email"
+          },
+          "updated_on": {
+            "type": "string",
+            "description": "updated_on"
+          },
+          "sensitivity": {
+            "type": "string",
+            "description": "sensitivity"
+          },
+          "status": {
+            "type": "string",
+            "description": "status"
+          }
+        },
+        "required": [],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_get_indexed_document",
+    "name": "get_indexed_document",
+    "description": "Get a indexed_document by its document_id",
+    "source_code": "def get_indexed_document(document_id: int) -> dict | None:\n    \"\"\"Get a single indexed_document by document_id.\"\"\"\n    conn = get_db()\n    try:\n        cursor = conn.execute(\"SELECT * FROM indexed_documents WHERE document_id = ?\", (document_id,))\n        row = cursor.fetchone()\n        if row is None:\n            return None\n        columns = [d[0] for d in cursor.description]\n        return dict(zip(columns, row))\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "document_id",
+        "type": "int",
+        "required": true
+      }
+    ],
+    "return_type": "dict | None",
+    "read_tables": [
+      "indexed_documents"
+    ],
+    "write_tables": [],
+    "json_schema": {
+      "name": "get_indexed_document",
+      "description": "Get a indexed_document by its document_id",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "document_id": {
+            "type": "integer",
+            "description": "document_id"
+          }
+        },
+        "required": [
+          "document_id"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_create_indexed_document",
+    "name": "create_indexed_document",
+    "description": "Create a new indexed_document",
+    "source_code": "def create_indexed_document(title: str, **kwargs) -> dict:\n    \"\"\"Create a new indexed_document.\"\"\"\n    conn = get_db()\n    try:\n        data = {\"title\": title}\n        data.update(kwargs)\n        columns = [k for k in data if k in ['source_id', 'title', 'owner_email', 'updated_on', 'sensitivity', 'status']]\n        values = [data[k] for k in columns]\n        if columns:\n            placeholders = \", \".join([\"?\"] * len(columns))\n            sql = f\"INSERT INTO indexed_documents ({', '.join(columns)}) VALUES ({placeholders})\"\n            cursor = conn.execute(sql, values)\n        else:\n            cursor = conn.execute(\"INSERT INTO indexed_documents DEFAULT VALUES\")\n        conn.commit()\n        new_id = cursor.lastrowid\n        cursor2 = conn.execute(\"SELECT * FROM indexed_documents WHERE document_id = ?\", (new_id,))\n        cols = [d[0] for d in cursor2.description]\n        return dict(zip(cols, cursor2.fetchone()))\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "source_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "title",
+        "type": "str",
+        "required": true
+      },
+      {
+        "name": "owner_email",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "updated_on",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "sensitivity",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "status",
+        "type": "str",
+        "required": false
+      }
+    ],
+    "return_type": "dict",
+    "read_tables": [],
+    "write_tables": [
+      "indexed_documents"
+    ],
+    "json_schema": {
+      "name": "create_indexed_document",
+      "description": "Create a new indexed_document",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "source_id": {
+            "type": "integer",
+            "description": "source_id"
+          },
+          "title": {
+            "type": "string",
+            "description": "title"
+          },
+          "owner_email": {
+            "type": "string",
+            "description": "owner_email"
+          },
+          "updated_on": {
+            "type": "string",
+            "description": "updated_on"
+          },
+          "sensitivity": {
+            "type": "string",
+            "description": "sensitivity"
+          },
+          "status": {
+            "type": "string",
+            "description": "status"
+          }
+        },
+        "required": [
+          "title"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_update_indexed_document",
+    "name": "update_indexed_document",
+    "description": "Update a indexed_document by document_id",
+    "source_code": "def update_indexed_document(document_id: int, **updates) -> dict | None:\n    \"\"\"Update a indexed_document by document_id.\"\"\"\n    conn = get_db()\n    try:\n        clean_updates = {\n            key: value for key, value in updates.items()\n            if key in ['source_id', 'title', 'owner_email', 'updated_on', 'sensitivity', 'status']\n        }\n        if clean_updates:\n            set_clause = \", \".join([f\"{k} = ?\" for k in clean_updates])\n            values = list(clean_updates.values()) + [document_id]\n            conn.execute(f\"UPDATE indexed_documents SET {set_clause} WHERE document_id = ?\", values)\n            conn.commit()\n        cursor = conn.execute(\"SELECT * FROM indexed_documents WHERE document_id = ?\", (document_id,))\n        row = cursor.fetchone()\n        if row is None:\n            return None\n        columns = [d[0] for d in cursor.description]\n        return dict(zip(columns, row))\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "document_id",
+        "type": "int",
+        "required": true
+      },
+      {
+        "name": "source_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "title",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "owner_email",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "updated_on",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "sensitivity",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "status",
+        "type": "str",
+        "required": false
+      }
+    ],
+    "return_type": "dict | None",
+    "read_tables": [
+      "indexed_documents"
+    ],
+    "write_tables": [
+      "indexed_documents"
+    ],
+    "json_schema": {
+      "name": "update_indexed_document",
+      "description": "Update a indexed_document by document_id",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "document_id": {
+            "type": "integer",
+            "description": "document_id"
+          },
+          "source_id": {
+            "type": "integer",
+            "description": "source_id"
+          },
+          "title": {
+            "type": "string",
+            "description": "title"
+          },
+          "owner_email": {
+            "type": "string",
+            "description": "owner_email"
+          },
+          "updated_on": {
+            "type": "string",
+            "description": "updated_on"
+          },
+          "sensitivity": {
+            "type": "string",
+            "description": "sensitivity"
+          },
+          "status": {
+            "type": "string",
+            "description": "status"
+          }
+        },
+        "required": [
+          "document_id"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_delete_indexed_document",
+    "name": "delete_indexed_document",
+    "description": "Delete a indexed_document by document_id",
+    "source_code": "def delete_indexed_document(document_id: int) -> bool:\n    \"\"\"Delete a indexed_document by document_id. Returns True if deleted.\"\"\"\n    conn = get_db()\n    try:\n        cursor = conn.execute(\"DELETE FROM indexed_documents WHERE document_id = ?\", (document_id,))\n        conn.commit()\n        return cursor.rowcount > 0\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "document_id",
+        "type": "int",
+        "required": true
+      }
+    ],
+    "return_type": "bool",
+    "read_tables": [],
+    "write_tables": [
+      "indexed_documents"
+    ],
+    "json_schema": {
+      "name": "delete_indexed_document",
+      "description": "Delete a indexed_document by document_id",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "document_id": {
+            "type": "integer",
+            "description": "document_id"
+          }
+        },
+        "required": [
+          "document_id"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_count_indexed_documents",
+    "name": "count_indexed_documents",
+    "description": "Count records in indexed_documents",
+    "source_code": "def count_indexed_documents(**filters) -> int:\n    \"\"\"Count records in indexed_documents, optionally filtered.\"\"\"\n    conn = get_db()\n    try:\n        unknown_filters = sorted(set(filters) - set(['document_id', 'source_id', 'title', 'owner_email', 'updated_on', 'sensitivity', 'status']))\n        if unknown_filters:\n            raise ValueError(\n                f\"Unsupported indexed_documents filters: {unknown_filters}\"\n            )\n        conditions = []\n        params = []\n        for key, value in filters.items():\n            conditions.append(f\"{key} = ?\")\n            params.append(value)\n        where = \"WHERE \" + \" AND \".join(conditions) if conditions else \"\"\n        cursor = conn.execute(f\"SELECT COUNT(*) FROM indexed_documents {where}\", params)\n        return cursor.fetchone()[0]\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "document_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "source_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "title",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "owner_email",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "updated_on",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "sensitivity",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "status",
+        "type": "str",
+        "required": false
+      }
+    ],
+    "return_type": "int",
+    "read_tables": [
+      "indexed_documents"
+    ],
+    "write_tables": [],
+    "json_schema": {
+      "name": "count_indexed_documents",
+      "description": "Count records in indexed_documents",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "document_id": {
+            "type": "integer",
+            "description": "document_id"
+          },
+          "source_id": {
+            "type": "integer",
+            "description": "source_id"
+          },
+          "title": {
+            "type": "string",
+            "description": "title"
+          },
+          "owner_email": {
+            "type": "string",
+            "description": "owner_email"
+          },
+          "updated_on": {
+            "type": "string",
+            "description": "updated_on"
+          },
+          "sensitivity": {
+            "type": "string",
+            "description": "sensitivity"
+          },
+          "status": {
+            "type": "string",
+            "description": "status"
+          }
+        },
+        "required": [],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_list_indexed_documents_for_source",
+    "name": "list_indexed_documents_for_source",
+    "description": "Investigate indexed_documents belonging to a source by source_id",
+    "source_code": "def list_indexed_documents_for_source(source_id: int, limit: int = 50) -> list:\n    \"\"\"List indexed_documents linked to one source.\"\"\"\n    conn = get_db()\n    try:\n        cursor = conn.execute(\n            \"SELECT * FROM indexed_documents WHERE source_id = ? LIMIT ?\",\n            (source_id, limit),\n        )\n        columns = [d[0] for d in cursor.description]\n        return [dict(zip(columns, row)) for row in cursor.fetchall()]\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "source_id",
+        "type": "int",
+        "required": true
+      },
+      {
+        "name": "limit",
+        "type": "int",
+        "required": false,
+        "default": "50"
+      }
+    ],
+    "return_type": "list[dict]",
+    "read_tables": [
+      "indexed_documents"
+    ],
+    "write_tables": [],
+    "json_schema": {
+      "name": "list_indexed_documents_for_source",
+      "description": "Investigate indexed_documents belonging to a source by source_id",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "source_id": {
+            "type": "integer",
+            "description": "source_id"
+          },
+          "limit": {
+            "type": "integer",
+            "description": "limit",
+            "default": "50"
+          }
+        },
+        "required": [
+          "source_id"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_search_search_queries",
+    "name": "search_search_queries",
+    "description": "Search search_queries by text match or filters",
+    "source_code": "def search_search_queries(query: str = \"\", limit: int = 10, **filters) -> list:\n    \"\"\"Search search_queries by text match or filters.\"\"\"\n    conn = get_db()\n    try:\n        # Tolerate agents passing filters via query (a common LLM mistake):\n        # a dict-valued query is really a set of column filters.\n        if isinstance(query, dict):\n            filters.update(query)\n            query = \"\"\n        unknown_filters = sorted(set(filters) - set(['query_id', 'employee_id', 'query_text', 'ran_on', 'result_count']))\n        if unknown_filters:\n            raise ValueError(\n                f\"Unsupported search_queries filters: {unknown_filters}\"\n            )\n        conditions = []\n        params = []\n        if query:\n            conditions.append(\"query_text LIKE ?\")\n            params.append(f\"%{query}%\")\n        for key, value in filters.items():\n            conditions.append(f\"{key} = ?\")\n            params.append(value)\n        where = \"WHERE \" + \" AND \".join(conditions) if conditions else \"\"\n        sql = f\"SELECT * FROM search_queries {where} LIMIT ?\"\n        params.append(limit)\n        cursor = conn.execute(sql, params)\n        columns = [d[0] for d in cursor.description]\n        return [dict(zip(columns, row)) for row in cursor.fetchall()]\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "query",
+        "type": "str",
+        "required": false,
+        "default": ""
+      },
+      {
+        "name": "limit",
+        "type": "int",
+        "required": false,
+        "default": "10"
+      },
+      {
+        "name": "employee_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "query_text",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "ran_on",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "result_count",
+        "type": "int",
+        "required": false
+      }
+    ],
+    "return_type": "list[dict]",
+    "read_tables": [
+      "search_queries"
+    ],
+    "write_tables": [],
+    "json_schema": {
+      "name": "search_search_queries",
+      "description": "Search search_queries by text match or filters",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "query": {
+            "type": "string",
+            "description": "query",
+            "default": ""
+          },
+          "limit": {
+            "type": "integer",
+            "description": "limit",
+            "default": "10"
+          },
+          "employee_id": {
+            "type": "integer",
+            "description": "employee_id"
+          },
+          "query_text": {
+            "type": "string",
+            "description": "query_text"
+          },
+          "ran_on": {
+            "type": "string",
+            "description": "ran_on"
+          },
+          "result_count": {
+            "type": "integer",
+            "description": "result_count"
+          }
+        },
+        "required": [],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_get_search_query",
+    "name": "get_search_query",
+    "description": "Get a search_query by its query_id",
+    "source_code": "def get_search_query(query_id: int) -> dict | None:\n    \"\"\"Get a single search_query by query_id.\"\"\"\n    conn = get_db()\n    try:\n        cursor = conn.execute(\"SELECT * FROM search_queries WHERE query_id = ?\", (query_id,))\n        row = cursor.fetchone()\n        if row is None:\n            return None\n        columns = [d[0] for d in cursor.description]\n        return dict(zip(columns, row))\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "query_id",
+        "type": "int",
+        "required": true
+      }
+    ],
+    "return_type": "dict | None",
+    "read_tables": [
+      "search_queries"
+    ],
+    "write_tables": [],
+    "json_schema": {
+      "name": "get_search_query",
+      "description": "Get a search_query by its query_id",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "query_id": {
+            "type": "integer",
+            "description": "query_id"
+          }
+        },
+        "required": [
+          "query_id"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_create_search_query",
+    "name": "create_search_query",
+    "description": "Create a new search_query",
+    "source_code": "def create_search_query(**kwargs) -> dict:\n    \"\"\"Create a new search_query.\"\"\"\n    conn = get_db()\n    try:\n        data = {}\n        data.update(kwargs)\n        columns = [k for k in data if k in ['employee_id', 'query_text', 'ran_on', 'result_count']]\n        values = [data[k] for k in columns]\n        if columns:\n            placeholders = \", \".join([\"?\"] * len(columns))\n            sql = f\"INSERT INTO search_queries ({', '.join(columns)}) VALUES ({placeholders})\"\n            cursor = conn.execute(sql, values)\n        else:\n            cursor = conn.execute(\"INSERT INTO search_queries DEFAULT VALUES\")\n        conn.commit()\n        new_id = cursor.lastrowid\n        cursor2 = conn.execute(\"SELECT * FROM search_queries WHERE query_id = ?\", (new_id,))\n        cols = [d[0] for d in cursor2.description]\n        return dict(zip(cols, cursor2.fetchone()))\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "employee_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "query_text",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "ran_on",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "result_count",
+        "type": "int",
+        "required": false
+      }
+    ],
+    "return_type": "dict",
+    "read_tables": [],
+    "write_tables": [
+      "search_queries"
+    ],
+    "json_schema": {
+      "name": "create_search_query",
+      "description": "Create a new search_query",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "employee_id": {
+            "type": "integer",
+            "description": "employee_id"
+          },
+          "query_text": {
+            "type": "string",
+            "description": "query_text"
+          },
+          "ran_on": {
+            "type": "string",
+            "description": "ran_on"
+          },
+          "result_count": {
+            "type": "integer",
+            "description": "result_count"
+          }
+        },
+        "required": [],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_update_search_query",
+    "name": "update_search_query",
+    "description": "Update a search_query by query_id",
+    "source_code": "def update_search_query(query_id: int, **updates) -> dict | None:\n    \"\"\"Update a search_query by query_id.\"\"\"\n    conn = get_db()\n    try:\n        clean_updates = {\n            key: value for key, value in updates.items()\n            if key in ['employee_id', 'query_text', 'ran_on', 'result_count']\n        }\n        if clean_updates:\n            set_clause = \", \".join([f\"{k} = ?\" for k in clean_updates])\n            values = list(clean_updates.values()) + [query_id]\n            conn.execute(f\"UPDATE search_queries SET {set_clause} WHERE query_id = ?\", values)\n            conn.commit()\n        cursor = conn.execute(\"SELECT * FROM search_queries WHERE query_id = ?\", (query_id,))\n        row = cursor.fetchone()\n        if row is None:\n            return None\n        columns = [d[0] for d in cursor.description]\n        return dict(zip(columns, row))\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "query_id",
+        "type": "int",
+        "required": true
+      },
+      {
+        "name": "employee_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "query_text",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "ran_on",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "result_count",
+        "type": "int",
+        "required": false
+      }
+    ],
+    "return_type": "dict | None",
+    "read_tables": [
+      "search_queries"
+    ],
+    "write_tables": [
+      "search_queries"
+    ],
+    "json_schema": {
+      "name": "update_search_query",
+      "description": "Update a search_query by query_id",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "query_id": {
+            "type": "integer",
+            "description": "query_id"
+          },
+          "employee_id": {
+            "type": "integer",
+            "description": "employee_id"
+          },
+          "query_text": {
+            "type": "string",
+            "description": "query_text"
+          },
+          "ran_on": {
+            "type": "string",
+            "description": "ran_on"
+          },
+          "result_count": {
+            "type": "integer",
+            "description": "result_count"
+          }
+        },
+        "required": [
+          "query_id"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_delete_search_query",
+    "name": "delete_search_query",
+    "description": "Delete a search_query by query_id",
+    "source_code": "def delete_search_query(query_id: int) -> bool:\n    \"\"\"Delete a search_query by query_id. Returns True if deleted.\"\"\"\n    conn = get_db()\n    try:\n        cursor = conn.execute(\"DELETE FROM search_queries WHERE query_id = ?\", (query_id,))\n        conn.commit()\n        return cursor.rowcount > 0\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "query_id",
+        "type": "int",
+        "required": true
+      }
+    ],
+    "return_type": "bool",
+    "read_tables": [],
+    "write_tables": [
+      "search_queries"
+    ],
+    "json_schema": {
+      "name": "delete_search_query",
+      "description": "Delete a search_query by query_id",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "query_id": {
+            "type": "integer",
+            "description": "query_id"
+          }
+        },
+        "required": [
+          "query_id"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_count_search_queries",
+    "name": "count_search_queries",
+    "description": "Count records in search_queries",
+    "source_code": "def count_search_queries(**filters) -> int:\n    \"\"\"Count records in search_queries, optionally filtered.\"\"\"\n    conn = get_db()\n    try:\n        unknown_filters = sorted(set(filters) - set(['query_id', 'employee_id', 'query_text', 'ran_on', 'result_count']))\n        if unknown_filters:\n            raise ValueError(\n                f\"Unsupported search_queries filters: {unknown_filters}\"\n            )\n        conditions = []\n        params = []\n        for key, value in filters.items():\n            conditions.append(f\"{key} = ?\")\n            params.append(value)\n        where = \"WHERE \" + \" AND \".join(conditions) if conditions else \"\"\n        cursor = conn.execute(f\"SELECT COUNT(*) FROM search_queries {where}\", params)\n        return cursor.fetchone()[0]\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "query_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "employee_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "query_text",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "ran_on",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "result_count",
+        "type": "int",
+        "required": false
+      }
+    ],
+    "return_type": "int",
+    "read_tables": [
+      "search_queries"
+    ],
+    "write_tables": [],
+    "json_schema": {
+      "name": "count_search_queries",
+      "description": "Count records in search_queries",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "query_id": {
+            "type": "integer",
+            "description": "query_id"
+          },
+          "employee_id": {
+            "type": "integer",
+            "description": "employee_id"
+          },
+          "query_text": {
+            "type": "string",
+            "description": "query_text"
+          },
+          "ran_on": {
+            "type": "string",
+            "description": "ran_on"
+          },
+          "result_count": {
+            "type": "integer",
+            "description": "result_count"
+          }
+        },
+        "required": [],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_list_search_queries_for_employee",
+    "name": "list_search_queries_for_employee",
+    "description": "Investigate search_queries belonging to a employee by employee_id",
+    "source_code": "def list_search_queries_for_employee(employee_id: int, limit: int = 50) -> list:\n    \"\"\"List search_queries linked to one employee.\"\"\"\n    conn = get_db()\n    try:\n        cursor = conn.execute(\n            \"SELECT * FROM search_queries WHERE employee_id = ? LIMIT ?\",\n            (employee_id, limit),\n        )\n        columns = [d[0] for d in cursor.description]\n        return [dict(zip(columns, row)) for row in cursor.fetchall()]\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "employee_id",
+        "type": "int",
+        "required": true
+      },
+      {
+        "name": "limit",
+        "type": "int",
+        "required": false,
+        "default": "50"
+      }
+    ],
+    "return_type": "list[dict]",
+    "read_tables": [
+      "search_queries"
+    ],
+    "write_tables": [],
+    "json_schema": {
+      "name": "list_search_queries_for_employee",
+      "description": "Investigate search_queries belonging to a employee by employee_id",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "employee_id": {
+            "type": "integer",
+            "description": "employee_id"
+          },
+          "limit": {
+            "type": "integer",
+            "description": "limit",
+            "default": "50"
+          }
+        },
+        "required": [
+          "employee_id"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_search_sources",
+    "name": "search_sources",
+    "description": "Search sources by text match or filters",
+    "source_code": "def search_sources(query: str = \"\", limit: int = 10, **filters) -> list:\n    \"\"\"Search sources by text match or filters.\"\"\"\n    conn = get_db()\n    try:\n        # Tolerate agents passing filters via query (a common LLM mistake):\n        # a dict-valued query is really a set of column filters.\n        if isinstance(query, dict):\n            filters.update(query)\n            query = \"\"\n        unknown_filters = sorted(set(filters) - set(['source_id', 'name', 'kind', 'connected', 'last_synced_on']))\n        if unknown_filters:\n            raise ValueError(\n                f\"Unsupported sources filters: {unknown_filters}\"\n            )\n        conditions = []\n        params = []\n        if query:\n            conditions.append(\"name LIKE ?\")\n            params.append(f\"%{query}%\")\n        for key, value in filters.items():\n            conditions.append(f\"{key} = ?\")\n            params.append(value)\n        where = \"WHERE \" + \" AND \".join(conditions) if conditions else \"\"\n        sql = f\"SELECT * FROM sources {where} LIMIT ?\"\n        params.append(limit)\n        cursor = conn.execute(sql, params)\n        columns = [d[0] for d in cursor.description]\n        return [dict(zip(columns, row)) for row in cursor.fetchall()]\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "query",
+        "type": "str",
+        "required": false,
+        "default": ""
+      },
+      {
+        "name": "limit",
+        "type": "int",
+        "required": false,
+        "default": "10"
+      },
+      {
+        "name": "name",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "kind",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "connected",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "last_synced_on",
+        "type": "str",
+        "required": false
+      }
+    ],
+    "return_type": "list[dict]",
+    "read_tables": [
+      "sources"
+    ],
+    "write_tables": [],
+    "json_schema": {
+      "name": "search_sources",
+      "description": "Search sources by text match or filters",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "query": {
+            "type": "string",
+            "description": "query",
+            "default": ""
+          },
+          "limit": {
+            "type": "integer",
+            "description": "limit",
+            "default": "10"
+          },
+          "name": {
+            "type": "string",
+            "description": "name"
+          },
+          "kind": {
+            "type": "string",
+            "description": "kind"
+          },
+          "connected": {
+            "type": "integer",
+            "description": "connected"
+          },
+          "last_synced_on": {
+            "type": "string",
+            "description": "last_synced_on"
+          }
+        },
+        "required": [],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_get_source",
+    "name": "get_source",
+    "description": "Get a source by its source_id",
+    "source_code": "def get_source(source_id: int) -> dict | None:\n    \"\"\"Get a single source by source_id.\"\"\"\n    conn = get_db()\n    try:\n        cursor = conn.execute(\"SELECT * FROM sources WHERE source_id = ?\", (source_id,))\n        row = cursor.fetchone()\n        if row is None:\n            return None\n        columns = [d[0] for d in cursor.description]\n        return dict(zip(columns, row))\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "source_id",
+        "type": "int",
+        "required": true
+      }
+    ],
+    "return_type": "dict | None",
+    "read_tables": [
+      "sources"
+    ],
+    "write_tables": [],
+    "json_schema": {
+      "name": "get_source",
+      "description": "Get a source by its source_id",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "source_id": {
+            "type": "integer",
+            "description": "source_id"
+          }
+        },
+        "required": [
+          "source_id"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_create_source",
+    "name": "create_source",
+    "description": "Create a new source",
+    "source_code": "def create_source(name: str, **kwargs) -> dict:\n    \"\"\"Create a new source.\"\"\"\n    conn = get_db()\n    try:\n        data = {\"name\": name}\n        data.update(kwargs)\n        columns = [k for k in data if k in ['name', 'kind', 'connected', 'last_synced_on']]\n        values = [data[k] for k in columns]\n        if columns:\n            placeholders = \", \".join([\"?\"] * len(columns))\n            sql = f\"INSERT INTO sources ({', '.join(columns)}) VALUES ({placeholders})\"\n            cursor = conn.execute(sql, values)\n        else:\n            cursor = conn.execute(\"INSERT INTO sources DEFAULT VALUES\")\n        conn.commit()\n        new_id = cursor.lastrowid\n        cursor2 = conn.execute(\"SELECT * FROM sources WHERE source_id = ?\", (new_id,))\n        cols = [d[0] for d in cursor2.description]\n        return dict(zip(cols, cursor2.fetchone()))\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "name",
+        "type": "str",
+        "required": true
+      },
+      {
+        "name": "kind",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "connected",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "last_synced_on",
+        "type": "str",
+        "required": false
+      }
+    ],
+    "return_type": "dict",
+    "read_tables": [],
+    "write_tables": [
+      "sources"
+    ],
+    "json_schema": {
+      "name": "create_source",
+      "description": "Create a new source",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "name": {
+            "type": "string",
+            "description": "name"
+          },
+          "kind": {
+            "type": "string",
+            "description": "kind"
+          },
+          "connected": {
+            "type": "integer",
+            "description": "connected"
+          },
+          "last_synced_on": {
+            "type": "string",
+            "description": "last_synced_on"
+          }
+        },
+        "required": [
+          "name"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_update_source",
+    "name": "update_source",
+    "description": "Update a source by source_id",
+    "source_code": "def update_source(source_id: int, **updates) -> dict | None:\n    \"\"\"Update a source by source_id.\"\"\"\n    conn = get_db()\n    try:\n        clean_updates = {\n            key: value for key, value in updates.items()\n            if key in ['name', 'kind', 'connected', 'last_synced_on']\n        }\n        if clean_updates:\n            set_clause = \", \".join([f\"{k} = ?\" for k in clean_updates])\n            values = list(clean_updates.values()) + [source_id]\n            conn.execute(f\"UPDATE sources SET {set_clause} WHERE source_id = ?\", values)\n            conn.commit()\n        cursor = conn.execute(\"SELECT * FROM sources WHERE source_id = ?\", (source_id,))\n        row = cursor.fetchone()\n        if row is None:\n            return None\n        columns = [d[0] for d in cursor.description]\n        return dict(zip(columns, row))\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "source_id",
+        "type": "int",
+        "required": true
+      },
+      {
+        "name": "name",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "kind",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "connected",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "last_synced_on",
+        "type": "str",
+        "required": false
+      }
+    ],
+    "return_type": "dict | None",
+    "read_tables": [
+      "sources"
+    ],
+    "write_tables": [
+      "sources"
+    ],
+    "json_schema": {
+      "name": "update_source",
+      "description": "Update a source by source_id",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "source_id": {
+            "type": "integer",
+            "description": "source_id"
+          },
+          "name": {
+            "type": "string",
+            "description": "name"
+          },
+          "kind": {
+            "type": "string",
+            "description": "kind"
+          },
+          "connected": {
+            "type": "integer",
+            "description": "connected"
+          },
+          "last_synced_on": {
+            "type": "string",
+            "description": "last_synced_on"
+          }
+        },
+        "required": [
+          "source_id"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_delete_source",
+    "name": "delete_source",
+    "description": "Delete a source by source_id",
+    "source_code": "def delete_source(source_id: int) -> bool:\n    \"\"\"Delete a source by source_id. Returns True if deleted.\"\"\"\n    conn = get_db()\n    try:\n        cursor = conn.execute(\"DELETE FROM sources WHERE source_id = ?\", (source_id,))\n        conn.commit()\n        return cursor.rowcount > 0\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "source_id",
+        "type": "int",
+        "required": true
+      }
+    ],
+    "return_type": "bool",
+    "read_tables": [],
+    "write_tables": [
+      "sources"
+    ],
+    "json_schema": {
+      "name": "delete_source",
+      "description": "Delete a source by source_id",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "source_id": {
+            "type": "integer",
+            "description": "source_id"
+          }
+        },
+        "required": [
+          "source_id"
+        ],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "tool_id": "tool_count_sources",
+    "name": "count_sources",
+    "description": "Count records in sources",
+    "source_code": "def count_sources(**filters) -> int:\n    \"\"\"Count records in sources, optionally filtered.\"\"\"\n    conn = get_db()\n    try:\n        unknown_filters = sorted(set(filters) - set(['source_id', 'name', 'kind', 'connected', 'last_synced_on']))\n        if unknown_filters:\n            raise ValueError(\n                f\"Unsupported sources filters: {unknown_filters}\"\n            )\n        conditions = []\n        params = []\n        for key, value in filters.items():\n            conditions.append(f\"{key} = ?\")\n            params.append(value)\n        where = \"WHERE \" + \" AND \".join(conditions) if conditions else \"\"\n        cursor = conn.execute(f\"SELECT COUNT(*) FROM sources {where}\", params)\n        return cursor.fetchone()[0]\n    finally:\n        conn.close()\n",
+    "parameters": [
+      {
+        "name": "source_id",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "name",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "kind",
+        "type": "str",
+        "required": false
+      },
+      {
+        "name": "connected",
+        "type": "int",
+        "required": false
+      },
+      {
+        "name": "last_synced_on",
+        "type": "str",
+        "required": false
+      }
+    ],
+    "return_type": "int",
+    "read_tables": [
+      "sources"
+    ],
+    "write_tables": [],
+    "json_schema": {
+      "name": "count_sources",
+      "description": "Count records in sources",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "source_id": {
+            "type": "integer",
+            "description": "source_id"
+          },
+          "name": {
+            "type": "string",
+            "description": "name"
+          },
+          "kind": {
+            "type": "string",
+            "description": "kind"
+          },
+          "connected": {
+            "type": "integer",
+            "description": "connected"
+          },
+          "last_synced_on": {
+            "type": "string",
+            "description": "last_synced_on"
+          }
+        },
+        "required": [],
+        "additionalProperties": false
+      }
+    }
+  }
+]

worlds/defensebench-100/world.json

@@ -0,0 +1,88 @@
+{
+  "world_id": "env_blobfish_benchmarks_defensebench_access__d2a0878c",
+  "blobfish_version": "0.1.0",
+  "vertical": "enterprise_search",
+  "tenant": "blobfish-benchmarks",
+  "brief": "",
+  "topic": "enterprise search index sources documents access grants queries",
+  "domain": "enterprise_search",
+  "engine": "builtin",
+  "seed": 20260830,
+  "created_at": 1787733440.8542368,
+  "world_digest": "e1df820514e93081e497aa1aa2fd50423e3dbf0cd766e4850838096a1bc641f6",
+  "entities": [
+    "sources",
+    "employees",
+    "indexed_documents",
+    "document_chunks",
+    "access_grants",
+    "search_queries"
+  ],
+  "tables_without_tools": [],
+  "counts": {
+    "tables": 6,
+    "rows": 121,
+    "tools": 42,
+    "tasks": 4,
+    "tasks_rejected": 0
+  },
+  "difficulty": {
+    "hard": 1,
+    "medium": 2,
+    "easy": 1
+  },
+  "splits": {
+    "train": [
+      "es_lock_down_confidential_document",
+      "es_mark_disconnected_source_stale",
+      "es_duplicate_person_name"
+    ],
+    "heldout": [
+      "es_dedupe_document_chunk"
+    ]
+  },
+  "task_creation": {
+    "schema_version": 1,
+    "status": "PASS",
+    "candidate_tasks": 4,
+    "accepted_tasks": 4,
+    "rejected_tasks": 0,
+    "rejections_by_gate": {},
+    "solvability": {
+      "oracle_replay_passed": 4,
+      "oracle_replay_failed": 0,
+      "all_admitted_tasks_oracle_solvable": true,
+      "impossible_candidates_removed_before_packaging": 0
+    },
+    "verifier_attribution": {
+      "all_admitted_tasks_attribution_safe": true,
+      "null_rollout_and_reference_replay_required": true,
+      "full_logical_state_protected_for_mutations": true
+    },
+    "workflow_coherence": {
+      "graph_candidates": 0,
+      "dependency_backed_candidates": 0,
+      "admitted_graph_tasks": 0,
+      "all_admitted_graph_tasks_dependency_backed": true,
+      "independent_connectivity_edges_allowed": false
+    },
+    "instruction_grounding": {
+      "graph_candidates": 0,
+      "fully_grounded_candidates": 0,
+      "admitted_graph_tasks": 0,
+      "all_admitted_graph_tasks_fully_determined_by_prompt": true,
+      "oracle_only_hidden_plans_allowed": false
+    },
+    "difficulty_calibration": {
+      "status": "NOT_MEASURED",
+      "claim_allowed": false,
+      "reason": "oracle success and verifier discrimination prove solvability, not model difficulty; a provider-identified stronger/less-capable model pair is required for a capability-separation claim"
+    }
+  },
+  "personas": [
+    "471317fa90287c5b",
+    "cd182d8c59453f30",
+    "e706287f6711fe7e",
+    "9a20f0dce752a1dd"
+  ]
+}

worlds/discoverybench-102/mcp/manifest.json

@@ -0,0 +1,69 @@
+{
+  "schema_version": "blobfish.mcp-service-layout.v1",
+  "world_id": "env_blobfish_benchmarks_discoverybench_scien_f7e1dbce",
+  "grouping": "tool.forge_service provenance only; no name/description inference (spec invariant 10)",
+  "systems": [],
+  "server_count": 1,
+  "tool_count": 43,
+  "servers": [
+    {
+      "name": "core",
+      "kind": "core",
+      "home": "core/",
+      "tool_count": 43,
+      "tools": [
+        "count_assay_results",
+        "count_assay_runs",
+        "count_assays",
+        "count_compounds",
+        "count_literature_refs",
+        "count_targets",
+        "create_assay",
+        "create_assay_result",
+        "create_assay_run",
+        "create_compound",
+        "create_literature_ref",
+        "create_target",
+        "delete_assay",
+        "delete_assay_result",
+        "delete_assay_run",
+        "delete_compound",
+        "delete_literature_ref",
+        "delete_target",
+        "get_assay",
+        "get_assay_result",
+        "get_assay_run",
+        "get_compound",
+        "get_literature_ref",
+        "get_target",
+        "list_assay_results_for_assay_run",
+        "list_assay_runs_for_assay",
+        "list_assay_runs_for_compound",
+        "list_assays_for_target",
+        "list_literature_refs_for_target",
+        "search_assay_results",
+        "search_assay_runs",
+        "search_assays",
+        "search_compounds",
+        "search_literature_refs",
+        "search_targets",
+        "transition_assay_run_status",
+        "transition_assay_status",
+        "update_assay",
+        "update_assay_result",
+        "update_assay_run",
+        "update_compound",
+        "update_literature_ref",
+        "update_target"
+      ],
+      "tables": [
+        "assay_results",
+        "assay_runs",
+        "assays",
+        "compounds",
+        "literature_refs",
+        "targets"
+      ]
+    }
+  ]
+}

worlds/discoverybench-102/schema.sql

@@ -0,0 +1,7 @@
+CREATE TABLE assay_results (result_id INTEGER PRIMARY KEY AUTOINCREMENT, run_id INTEGER REFERENCES assay_runs(run_id), replicate INTEGER DEFAULT 1, ic50_nm REAL DEFAULT 0, percent_inhibition REAL DEFAULT 0, qc_flag TEXT DEFAULT 'pass');
+CREATE TABLE assay_runs (run_id INTEGER PRIMARY KEY AUTOINCREMENT, assay_id INTEGER REFERENCES assays(assay_id), compound_id INTEGER REFERENCES compounds(compound_id), run_on TEXT, plate_label TEXT, status TEXT DEFAULT 'pending', CHECK ("status" IS NULL OR "status" IN ('approved', 'complete', 'completed', 'invalid', 'pending', 'rejected')));
+CREATE TABLE assays (assay_id INTEGER PRIMARY KEY AUTOINCREMENT, target_id INTEGER REFERENCES targets(target_id), name TEXT NOT NULL, assay_type TEXT, readout TEXT, status TEXT DEFAULT 'active', CHECK ("status" IS NULL OR "status" IN ('active', 'inactive', 'suspended')));
+CREATE TABLE compounds (compound_id INTEGER PRIMARY KEY AUTOINCREMENT, code TEXT NOT NULL, name TEXT, molecular_weight REAL DEFAULT 0, status TEXT DEFAULT 'screening', screening_note TEXT);
+CREATE TABLE literature_refs (ref_id INTEGER PRIMARY KEY AUTOINCREMENT, target_id INTEGER REFERENCES targets(target_id), citation TEXT, doi TEXT, relevance TEXT DEFAULT 'medium', CHECK ("relevance" IS NULL OR "relevance" IN ('low', 'medium', 'high', 'urgent')));
+CREATE TABLE targets (target_id INTEGER PRIMARY KEY AUTOINCREMENT, gene_symbol TEXT NOT NULL, protein_name TEXT, organism TEXT DEFAULT 'human', pathway TEXT);
+DELETE FROM "sqlite_sequence";

worlds/discoverybench-102/seed.sql

@@ -0,0 +1,125 @@
+INSERT INTO "assay_results" VALUES(1,2,1,404.29,25.5,'pass');
+INSERT INTO "assay_results" VALUES(2,12,1,432.55,30.8,'pass');
+INSERT INTO "assay_results" VALUES(3,5,1,89.73,12.6,'pass');
+INSERT INTO "assay_results" VALUES(4,10,1,214.47,46.6,'pass');
+INSERT INTO "assay_results" VALUES(5,11,1,343.63,34.2,'pass');
+INSERT INTO "assay_results" VALUES(6,3,1,485.11,47.0,'pass');
+INSERT INTO "assay_results" VALUES(7,4,1,133.91,43.1,'pass');
+INSERT INTO "assay_results" VALUES(8,6,1,224.05,32.6,'pass');
+INSERT INTO "assay_results" VALUES(9,11,1,123.47,25.5,'pass');
+INSERT INTO "assay_results" VALUES(10,5,1,349.85,42.3,'pass');
+INSERT INTO "assay_results" VALUES(11,3,1,27.56,5.6,'pass');
+INSERT INTO "assay_results" VALUES(12,1,1,10.63,39.4,'pass');
+INSERT INTO "assay_results" VALUES(13,17,1,453.73,48.1,'pass');
+INSERT INTO "assay_results" VALUES(14,3,1,461.03,41.7,'pass');
+INSERT INTO "assay_results" VALUES(15,15,1,70.31,21.0,'pass');
+INSERT INTO "assay_results" VALUES(16,4,1,279.98,43.7,'pass');
+INSERT INTO "assay_results" VALUES(17,4,1,237.27,35.2,'pass');
+INSERT INTO "assay_results" VALUES(9401,9301,1,20.0,91.0,'pass');
+INSERT INTO "assay_results" VALUES(9402,9301,2,24.0,89.0,'pass');
+INSERT INTO "assay_results" VALUES(9403,9301,3,22.0,90.0,'pass');
+INSERT INTO "assay_results" VALUES(9404,9301,4,400.0,11.0,'fail');
+INSERT INTO "assay_results" VALUES(9411,9302,1,0.0,4.0,'fail');
+INSERT INTO "assay_results" VALUES(9412,9302,1,0.0,4.0,'fail');
+INSERT INTO "assay_results" VALUES(9413,9302,2,0.0,6.0,'fail');
+INSERT INTO "assay_runs" VALUES(1,9,4,'Primary Run On B29','Consolidated Retail Assay Run','complete');
+INSERT INTO "assay_runs" VALUES(2,9,4,'Secondary Run On F64','Scheduled Export Assay Run','rejected');
+INSERT INTO "assay_runs" VALUES(3,11,2,'Standard Run On F78','Annual Export Assay Run','completed');
+INSERT INTO "assay_runs" VALUES(4,5,10,'Regional Run On A39','Expedited Regional Assay Run','rejected');
+INSERT INTO "assay_runs" VALUES(5,1,4,'General Run On E57','Standard Export Assay Run','completed');
+INSERT INTO "assay_runs" VALUES(6,3,8,'Preferred Run On C45','Standard Internal Assay Run','complete');
+INSERT INTO "assay_runs" VALUES(7,12,11,'Secondary Run On D89','Scheduled Regional Assay Run','invalid');
+INSERT INTO "assay_runs" VALUES(8,1,6,'Primary Run On D93','Consolidated Export Assay Run','complete');
+INSERT INTO "assay_runs" VALUES(9,1,2,'Consolidated Run On F73','Consolidated Domestic Assay Run','rejected');
+INSERT INTO "assay_runs" VALUES(10,6,8,'Regional Run On B36','Expedited Export Assay Run','pending');
+INSERT INTO "assay_runs" VALUES(11,8,5,'General Run On A16','Expedited Retail Assay Run','completed');
+INSERT INTO "assay_runs" VALUES(12,4,5,'Secondary Run On A92','Priority Regional Assay Run','approved');
+INSERT INTO "assay_runs" VALUES(13,2,3,'Preferred Run On E87','Consolidated Export Assay Run','complete');
+INSERT INTO "assay_runs" VALUES(14,8,17,'Secondary Run On A20','Annual Internal Assay Run','completed');
+INSERT INTO "assay_runs" VALUES(15,13,15,'Preferred Run On A11','Recurring Domestic Assay Run','pending');
+INSERT INTO "assay_runs" VALUES(16,2,11,'Regional Run On F79','Quarterly Retail Assay Run','approved');
+INSERT INTO "assay_runs" VALUES(17,8,2,'Primary Run On C92','Standard Partner Assay Run','approved');
+INSERT INTO "assay_runs" VALUES(9301,9201,9101,'2026-07-14','PLT-0091','pending');
+INSERT INTO "assay_runs" VALUES(9302,9202,9103,'2026-07-15','PLT-0092','pending');
+INSERT INTO "assays" VALUES(1,11,'Priority Retail Assay','standard','Primary Readout D56','suspended');
+INSERT INTO "assays" VALUES(2,10,'Standard Retail Assay','standard','Secondary Readout F80','inactive');
+INSERT INTO "assays" VALUES(3,2,'Priority Retail Assay','premium','General Readout B61','suspended');
+INSERT INTO "assays" VALUES(4,7,'Consolidated Regional Assay','advanced','Consolidated Readout E61','suspended');
+INSERT INTO "assays" VALUES(5,5,'Quarterly Internal Assay','basic','Backup Readout F73','inactive');
+INSERT INTO "assays" VALUES(6,4,'Standard Partner Assay','standard','Standard Readout B36','suspended');
+INSERT INTO "assays" VALUES(7,12,'Consolidated Regional Assay','standard','Consolidated Readout D41','inactive');
+INSERT INTO "assays" VALUES(8,9,'Expedited Export Assay','standard','Primary Readout C32','active');
+INSERT INTO "assays" VALUES(9,8,'Expedited Partner Assay','advanced','Preferred Readout E58','suspended');
+INSERT INTO "assays" VALUES(10,8,'Scheduled Regional Assay','general','General Readout E36','inactive');
+INSERT INTO "assays" VALUES(11,7,'Annual Domestic Assay','basic','Standard Readout D36','active');
+INSERT INTO "assays" VALUES(12,6,'Standard Internal Assay','standard','Preferred Readout E58','suspended');
+INSERT INTO "assays" VALUES(13,6,'Recurring Internal Assay','premium','Preferred Readout A86','inactive');
+INSERT INTO "assays" VALUES(14,9,'Quarterly Regional Assay','advanced','General Readout B86','active');
+INSERT INTO "assays" VALUES(15,3,'Standard Internal Assay','general','Primary Readout D81','active');
+INSERT INTO "assays" VALUES(16,1,'Recurring Regional Assay','basic','Backup Readout A54','active');
+INSERT INTO "assays" VALUES(17,8,'Consolidated Internal Assay','premium','General Readout D84','inactive');
+INSERT INTO "assays" VALUES(9201,9001,'KRAS G12C biochemical','biochemical','ic50','active');
+INSERT INTO "assays" VALUES(9202,9002,'PARP1 cellular','cellular','percent_inhibition','active');
+INSERT INTO "compounds" VALUES(1,'LR05A61G','Scheduled Partner Compound',178.4,'screening','Pending review');
+INSERT INTO "compounds" VALUES(2,'ISYHI88G','Expedited Regional Compound',16.11,'screening','Updated compound per request');
+INSERT INTO "compounds" VALUES(3,'3SC4LJVN','Scheduled Retail Compound',173.22,'screening','Follow-up required');
+INSERT INTO "compounds" VALUES(4,'0IG7LVRV','Recurring Partner Compound',248.81,'screening','Routine compound entry');
+INSERT INTO "compounds" VALUES(5,'N8385WIQ','Recurring Internal Compound',196.38,'screening','Updated compound per request');
+INSERT INTO "compounds" VALUES(6,'WRDG9XD1','Recurring Retail Compound',366.89,'screening','Pending review');
+INSERT INTO "compounds" VALUES(7,'D60P8SHU','Quarterly Retail Compound',217.18,'screening','Verified and approved');
+INSERT INTO "compounds" VALUES(8,'N0T73H0S','Expedited Internal Compound',208.69,'screening','Follow-up required');
+INSERT INTO "compounds" VALUES(9,'FMUC1K4V','Expedited Regional Compound',260.98,'screening','Pending review');
+INSERT INTO "compounds" VALUES(10,'ZK0LBM1J','Annual Retail Compound',273.72,'screening','Updated compound per request');
+INSERT INTO "compounds" VALUES(11,'LMC7QZ9N','Consolidated Export Compound',294.27,'screening','Escalated for attention');
+INSERT INTO "compounds" VALUES(12,'6J5POE83','Standard Retail Compound',151.7,'screening','Follow-up required');
+INSERT INTO "compounds" VALUES(13,'KPI64D8H','Recurring Export Compound',199.18,'screening','Pending review');
+INSERT INTO "compounds" VALUES(14,'T3D613DV','Priority Domestic Compound',265.52,'screening','Escalated for attention');
+INSERT INTO "compounds" VALUES(15,'FJ1MQFJ0','Expedited Partner Compound',458.94,'screening','Updated compound per request');
+INSERT INTO "compounds" VALUES(16,'2IG31ZFZ','Consolidated Export Compound',458.07,'screening','Routine compound entry');
+INSERT INTO "compounds" VALUES(17,'M3C4M7BT','Expedited Retail Compound',142.42,'screening','Follow-up required');
+INSERT INTO "compounds" VALUES(9101,'BF-1042','Velatinib',421.5,'screening','');
+INSERT INTO "compounds" VALUES(9102,'BF-2077','Velatinib',448.2,'screening','');
+INSERT INTO "compounds" VALUES(9103,'BF-3311','Corabinib',388.9,'screening','');
+INSERT INTO "literature_refs" VALUES(1,8,'Secondary Citation B43','Preferred Doi E49','urgent');
+INSERT INTO "literature_refs" VALUES(2,3,'Primary Citation B86','Regional Doi C15','high');
+INSERT INTO "literature_refs" VALUES(3,5,'Primary Citation A60','Regional Doi B73','urgent');
+INSERT INTO "literature_refs" VALUES(4,11,'Consolidated Citation A19','Backup Doi B69','low');
+INSERT INTO "literature_refs" VALUES(5,3,'Primary Citation F18','Backup Doi A37','medium');
+INSERT INTO "literature_refs" VALUES(6,9,'Primary Citation E39','Regional Doi C31','medium');
+INSERT INTO "literature_refs" VALUES(7,1,'Consolidated Citation A16','Secondary Doi A20','high');
+INSERT INTO "literature_refs" VALUES(8,12,'Preferred Citation E94','General Doi F54','low');
+INSERT INTO "literature_refs" VALUES(9,8,'General Citation A46','Secondary Doi B97','urgent');
+INSERT INTO "literature_refs" VALUES(10,11,'Consolidated Citation F77','Consolidated Doi C49','high');
+INSERT INTO "literature_refs" VALUES(11,9,'Primary Citation E78','General Doi E90','urgent');
+INSERT INTO "literature_refs" VALUES(12,1,'Regional Citation B45','Primary Doi E52','medium');
+INSERT INTO "literature_refs" VALUES(13,8,'Backup Citation A35','Regional Doi B58','high');
+INSERT INTO "literature_refs" VALUES(14,5,'General Citation E59','Regional Doi B76','low');
+INSERT INTO "literature_refs" VALUES(15,13,'Secondary Citation F28','Consolidated Doi C56','low');
+INSERT INTO "literature_refs" VALUES(16,15,'Preferred Citation F30','Consolidated Doi E40','medium');
+INSERT INTO "literature_refs" VALUES(17,4,'Backup Citation F63','Consolidated Doi A22','high');
+INSERT INTO "literature_refs" VALUES(9501,9001,'Covalent KRAS G12C inhibition, 2024','10.1000/kras.2024','high');
+INSERT INTO "targets" VALUES(1,'Backup Gene Symbol C94','Priority Regional Target','human','Secondary Pathway F13');
+INSERT INTO "targets" VALUES(2,'Standard Gene Symbol A86','Recurring Retail Target','human','Secondary Pathway E42');
+INSERT INTO "targets" VALUES(3,'Secondary Gene Symbol A82','Standard Domestic Target','human','Standard Pathway C43');
+INSERT INTO "targets" VALUES(4,'General Gene Symbol C41','Quarterly Domestic Target','human','Preferred Pathway E55');
+INSERT INTO "targets" VALUES(5,'Preferred Gene Symbol D91','Consolidated Domestic Target','human','Backup Pathway A67');
+INSERT INTO "targets" VALUES(6,'Consolidated Gene Symbol A59','Priority Regional Target','human','Primary Pathway A24');
+INSERT INTO "targets" VALUES(7,'General Gene Symbol D78','Consolidated Regional Target','human','General Pathway A31');
+INSERT INTO "targets" VALUES(8,'Preferred Gene Symbol C10','Annual Internal Target','human','General Pathway F78');
+INSERT INTO "targets" VALUES(9,'Standard Gene Symbol D63','Expedited Retail Target','human','Regional Pathway B79');
+INSERT INTO "targets" VALUES(10,'Consolidated Gene Symbol C84','Standard Domestic Target','human','Secondary Pathway C76');
+INSERT INTO "targets" VALUES(11,'Consolidated Gene Symbol F17','Quarterly Export Target','human','Consolidated Pathway E52');
+INSERT INTO "targets" VALUES(12,'General Gene Symbol B36','Priority Domestic Target','human','Backup Pathway D51');
+INSERT INTO "targets" VALUES(13,'Standard Gene Symbol A44','Priority Export Target','human','Consolidated Pathway D58');
+INSERT INTO "targets" VALUES(14,'Secondary Gene Symbol C16','Expedited Partner Target','human','Primary Pathway D33');
+INSERT INTO "targets" VALUES(15,'Primary Gene Symbol E27','Quarterly Export Target','human','Regional Pathway F29');
+INSERT INTO "targets" VALUES(16,'Regional Gene Symbol A59','Scheduled Regional Target','human','Primary Pathway E12');
+INSERT INTO "targets" VALUES(17,'Consolidated Gene Symbol C94','Scheduled Internal Target','human','Standard Pathway F30');
+INSERT INTO "targets" VALUES(9001,'KRAS','GTPase KRas','human','MAPK');
+INSERT INTO "targets" VALUES(9002,'PARP1','Poly ADP-ribose polymerase 1','human','DNA repair');
+INSERT INTO "sqlite_sequence" VALUES('targets',9002);
+INSERT INTO "sqlite_sequence" VALUES('compounds',9103);
+INSERT INTO "sqlite_sequence" VALUES('assays',9202);
+INSERT INTO "sqlite_sequence" VALUES('assay_runs',9302);
+INSERT INTO "sqlite_sequence" VALUES('assay_results',9413);
+INSERT INTO "sqlite_sequence" VALUES('literature_refs',9501);

worlds/discoverybench-102/world.json

@@ -0,0 +1,88 @@
+{
+  "world_id": "env_blobfish_benchmarks_discoverybench_scien_f7e1dbce",
+  "blobfish_version": "0.1.0",
+  "vertical": "bio_research",
+  "tenant": "blobfish-benchmarks",
+  "brief": "",
+  "topic": "preclinical research targets compounds assays screening qc",
+  "domain": "bio_research",
+  "engine": "builtin",
+  "seed": 20260829,
+  "created_at": 1787733344.29895,
+  "world_digest": "475b17df4a7efe8265400a4d47889288764829150a4dc63e9b1188a06948cd29",
+  "entities": [
+    "targets",
+    "compounds",
+    "assays",
+    "assay_runs",
+    "assay_results",
+    "literature_refs"
+  ],
+  "tables_without_tools": [],
+  "counts": {
+    "tables": 6,
+    "rows": 119,
+    "tools": 43,
+    "tasks": 4,
+    "tasks_rejected": 0
+  },
+  "difficulty": {
+    "hard": 1,
+    "medium": 2,
+    "easy": 1
+  },
+  "splits": {
+    "train": [
+      "bio_promote_hit_compound",
+      "bio_invalidate_failed_run",
+      "bio_duplicate_compound_name"
+    ],
+    "heldout": [
+      "bio_dedupe_assay_result"
+    ]
+  },
+  "task_creation": {
+    "schema_version": 1,
+    "status": "PASS",
+    "candidate_tasks": 4,
+    "accepted_tasks": 4,
+    "rejected_tasks": 0,
+    "rejections_by_gate": {},
+    "solvability": {
+      "oracle_replay_passed": 4,
+      "oracle_replay_failed": 0,
+      "all_admitted_tasks_oracle_solvable": true,
+      "impossible_candidates_removed_before_packaging": 0
+    },
+    "verifier_attribution": {
+      "all_admitted_tasks_attribution_safe": true,
+      "null_rollout_and_reference_replay_required": true,
+      "full_logical_state_protected_for_mutations": true
+    },
+    "workflow_coherence": {
+      "graph_candidates": 0,
+      "dependency_backed_candidates": 0,
+      "admitted_graph_tasks": 0,
+      "all_admitted_graph_tasks_dependency_backed": true,
+      "independent_connectivity_edges_allowed": false
+    },
+    "instruction_grounding": {
+      "graph_candidates": 0,
+      "fully_grounded_candidates": 0,
+      "admitted_graph_tasks": 0,
+      "all_admitted_graph_tasks_fully_determined_by_prompt": true,
+      "oracle_only_hidden_plans_allowed": false
+    },
+    "difficulty_calibration": {
+      "status": "NOT_MEASURED",
+      "claim_allowed": false,
+      "reason": "oracle success and verifier discrimination prove solvability, not model difficulty; a provider-identified stronger/less-capable model pair is required for a capability-separation claim"
+    }
+  },
+  "personas": [
+    "9ac570ba275cbd24",
+    "9c415c70fb19634e",
+    "1c0ac25cdffc4b65",
+    "cc99d5ca833239f1"
+  ]
+}

worlds/filingbench-50/schema.sql

@@ -0,0 +1,7 @@
+CREATE TABLE dcf_assumptions (assumption_id INTEGER PRIMARY KEY AUTOINCREMENT, model_id INTEGER REFERENCES dcf_models(model_id), name TEXT, value REAL DEFAULT 0, source TEXT, status TEXT DEFAULT 'unverified');
+CREATE TABLE dcf_models (model_id INTEGER PRIMARY KEY AUTOINCREMENT, company_id INTEGER REFERENCES fa_companies(company_id), name TEXT NOT NULL, discount_rate REAL DEFAULT 0, terminal_growth REAL DEFAULT 0, status TEXT DEFAULT 'draft', enterprise_value REAL DEFAULT 0, CHECK ("status" IS NULL OR "status" IN ('draft', 'submitted', 'approved', 'final')));
+CREATE TABLE fa_companies (company_id INTEGER PRIMARY KEY AUTOINCREMENT, name TEXT NOT NULL, ticker TEXT, sector TEXT);
+CREATE TABLE financial_periods (period_id INTEGER PRIMARY KEY AUTOINCREMENT, company_id INTEGER REFERENCES fa_companies(company_id), period_label TEXT, revenue REAL DEFAULT 0, ebitda REAL DEFAULT 0, capex REAL DEFAULT 0, free_cash_flow REAL DEFAULT 0);
+CREATE TABLE lbo_models (lbo_id INTEGER PRIMARY KEY AUTOINCREMENT, company_id INTEGER REFERENCES fa_companies(company_id), entry_multiple REAL DEFAULT 0, exit_multiple REAL DEFAULT 0, leverage_turns REAL DEFAULT 0, irr_pct REAL DEFAULT 0, status TEXT DEFAULT 'draft', CHECK ("status" IS NULL OR "status" IN ('draft', 'submitted', 'approved', 'final')));
+CREATE TABLE review_notes (note_id INTEGER PRIMARY KEY AUTOINCREMENT, model_id INTEGER REFERENCES dcf_models(model_id), reviewer_name TEXT, body TEXT, status TEXT DEFAULT 'open', CHECK ("status" IS NULL OR "status" IN ('open', 'closed', 'on_hold')));
+DELETE FROM "sqlite_sequence";

worlds/filingbench-50/seed.sql

@@ -0,0 +1,122 @@
+INSERT INTO "dcf_assumptions" VALUES(1,12,'Priority Partner Dcf Assumption',270.05,'Backup Source F70','unverified');
+INSERT INTO "dcf_assumptions" VALUES(2,3,'Consolidated Internal Dcf Assumption',404.67,'Primary Source B93','unverified');
+INSERT INTO "dcf_assumptions" VALUES(3,7,'Consolidated Regional Dcf Assumption',139.9,'Standard Source D40','unverified');
+INSERT INTO "dcf_assumptions" VALUES(4,10,'Quarterly Regional Dcf Assumption',183.47,'Secondary Source D36','unverified');
+INSERT INTO "dcf_assumptions" VALUES(5,8,'Annual Domestic Dcf Assumption',101.91,'Preferred Source F38','unverified');
+INSERT INTO "dcf_assumptions" VALUES(6,1,'Quarterly Internal Dcf Assumption',198.88,'Regional Source D14','unverified');
+INSERT INTO "dcf_assumptions" VALUES(7,6,'Quarterly Regional Dcf Assumption',83.46,'General Source F92','unverified');
+INSERT INTO "dcf_assumptions" VALUES(8,2,'Annual Export Dcf Assumption',61.46,'Preferred Source C85','unverified');
+INSERT INTO "dcf_assumptions" VALUES(9,7,'Recurring Internal Dcf Assumption',306.69,'Standard Source C83','unverified');
+INSERT INTO "dcf_assumptions" VALUES(10,9,'Standard Partner Dcf Assumption',433.6,'Preferred Source E60','unverified');
+INSERT INTO "dcf_assumptions" VALUES(11,5,'Consolidated Retail Dcf Assumption',342.09,'Secondary Source C34','unverified');
+INSERT INTO "dcf_assumptions" VALUES(12,10,'Consolidated Retail Dcf Assumption',17.48,'Primary Source F51','unverified');
+INSERT INTO "dcf_assumptions" VALUES(13,8,'Quarterly Partner Dcf Assumption',199.47,'Regional Source F88','unverified');
+INSERT INTO "dcf_assumptions" VALUES(14,7,'Recurring Retail Dcf Assumption',161.6,'Regional Source B75','unverified');
+INSERT INTO "dcf_assumptions" VALUES(15,13,'Annual Partner Dcf Assumption',415.76,'Regional Source D16','unverified');
+INSERT INTO "dcf_assumptions" VALUES(16,9,'Scheduled Regional Dcf Assumption',169.31,'Backup Source A65','unverified');
+INSERT INTO "dcf_assumptions" VALUES(17,14,'Expedited Internal Dcf Assumption',217.1,'Preferred Source D15','unverified');
+INSERT INTO "dcf_assumptions" VALUES(9301,9201,'WACC',0.09,'Capital IQ','unverified');
+INSERT INTO "dcf_assumptions" VALUES(9311,9202,'Terminal growth',0.02,'House view','unverified');
+INSERT INTO "dcf_assumptions" VALUES(9312,9202,'Terminal growth',0.02,'House view','unverified');
+INSERT INTO "dcf_models" VALUES(1,7,'Priority Export Dcf Model',419.28,334.88,'draft',184.89);
+INSERT INTO "dcf_models" VALUES(2,5,'Recurring Retail Dcf Model',409.23,284.39,'submitted',351.35);
+INSERT INTO "dcf_models" VALUES(3,3,'Expedited Internal Dcf Model',263.7,442.2,'submitted',308.05);
+INSERT INTO "dcf_models" VALUES(4,9,'Expedited Partner Dcf Model',93.2,207.57,'final',184.21);
+INSERT INTO "dcf_models" VALUES(5,10,'Recurring Domestic Dcf Model',310.85,77.41,'draft',397.0);
+INSERT INTO "dcf_models" VALUES(6,12,'Expedited Domestic Dcf Model',142.96,475.7,'approved',291.46);
+INSERT INTO "dcf_models" VALUES(7,2,'Consolidated Domestic Dcf Model',409.91,319.89,'final',78.65);
+INSERT INTO "dcf_models" VALUES(8,4,'Expedited Partner Dcf Model',276.69,162.67,'approved',335.81);
+INSERT INTO "dcf_models" VALUES(9,11,'Standard Export Dcf Model',55.27,464.98,'approved',220.84);
+INSERT INTO "dcf_models" VALUES(10,3,'Consolidated Partner Dcf Model',418.9,439.72,'draft',497.88);
+INSERT INTO "dcf_models" VALUES(11,6,'Expedited Domestic Dcf Model',185.53,316.62,'submitted',19.15);
+INSERT INTO "dcf_models" VALUES(12,4,'Recurring Export Dcf Model',407.56,324.17,'draft',10.32);
+INSERT INTO "dcf_models" VALUES(13,2,'Annual Internal Dcf Model',66.64,193.28,'draft',439.29);
+INSERT INTO "dcf_models" VALUES(14,17,'Expedited Partner Dcf Model',212.69,144.0,'approved',338.0);
+INSERT INTO "dcf_models" VALUES(15,12,'Recurring Partner Dcf Model',384.01,347.04,'draft',413.38);
+INSERT INTO "dcf_models" VALUES(16,17,'Scheduled Export Dcf Model',70.8,240.79,'draft',460.29);
+INSERT INTO "dcf_models" VALUES(17,11,'Recurring Domestic Dcf Model',61.83,268.48,'submitted',23.58);
+INSERT INTO "dcf_models" VALUES(9201,9001,'Solstice base case',0.09,0.02,'draft',0.0);
+INSERT INTO "dcf_models" VALUES(9202,9003,'Kestrel quick cut',0.0,0.02,'final',640.0);
+INSERT INTO "fa_companies" VALUES(1,'Thomas Global Industries','Standard Ticker E95','Backup Sector A61');
+INSERT INTO "fa_companies" VALUES(2,'Lopez Allied Group','Secondary Ticker A95','Regional Sector A83');
+INSERT INTO "fa_companies" VALUES(3,'Gonzalez Supply Co','Backup Ticker E92','Backup Sector A48');
+INSERT INTO "fa_companies" VALUES(4,'Moore United Inc','Standard Ticker C33','Backup Sector E83');
+INSERT INTO "fa_companies" VALUES(5,'Thompson Global Supply Co','Primary Ticker F86','General Sector D39');
+INSERT INTO "fa_companies" VALUES(6,'Brown Meridian Trading Co','Backup Ticker B71','Backup Sector D63');
+INSERT INTO "fa_companies" VALUES(7,'Thomas Corp','Preferred Ticker B99','Consolidated Sector E69');
+INSERT INTO "fa_companies" VALUES(8,'Gonzalez Continental Logistics','General Ticker F38','General Sector A66');
+INSERT INTO "fa_companies" VALUES(9,'Young Meridian Enterprises','General Ticker E35','Consolidated Sector F24');
+INSERT INTO "fa_companies" VALUES(10,'Walker Holdings','Standard Ticker A14','Consolidated Sector B29');
+INSERT INTO "fa_companies" VALUES(11,'Harris Precision LLC','Secondary Ticker F44','Backup Sector B68');
+INSERT INTO "fa_companies" VALUES(12,'Lee United LLC','Standard Ticker C22','Standard Sector D98');
+INSERT INTO "fa_companies" VALUES(13,'Jones Supply Co','Backup Ticker B79','Secondary Sector F43');
+INSERT INTO "fa_companies" VALUES(14,'Smith Meridian Logistics','Preferred Ticker A46','Standard Sector A84');
+INSERT INTO "fa_companies" VALUES(15,'Lewis Allied Industries','Consolidated Ticker A92','General Sector F99');
+INSERT INTO "fa_companies" VALUES(16,'King Northern Solutions','Secondary Ticker D55','Backup Sector B35');
+INSERT INTO "fa_companies" VALUES(17,'Ramirez Precision Enterprises','Regional Ticker B51','Backup Sector B99');
+INSERT INTO "fa_companies" VALUES(9001,'Solstice Energy','SLE','energy');
+INSERT INTO "fa_companies" VALUES(9002,'Solstice Energy','SLE-B','energy');
+INSERT INTO "fa_companies" VALUES(9003,'Kestrel Retail','KSR','consumer');
+INSERT INTO "financial_periods" VALUES(1,8,'Consolidated Partner Financial Period',379.26,457.25,340.27,379.57);
+INSERT INTO "financial_periods" VALUES(2,12,'Annual Retail Financial Period',370.26,269.49,53.71,189.24);
+INSERT INTO "financial_periods" VALUES(3,4,'Recurring Internal Financial Period',215.32,293.39,122.68,117.32);
+INSERT INTO "financial_periods" VALUES(4,2,'Annual Retail Financial Period',320.36,321.49,57.42,355.51);
+INSERT INTO "financial_periods" VALUES(5,6,'Standard Internal Financial Period',430.38,32.19,305.19,303.13);
+INSERT INTO "financial_periods" VALUES(6,7,'Annual Regional Financial Period',425.26,14.34,88.47,17.57);
+INSERT INTO "financial_periods" VALUES(7,12,'Priority Export Financial Period',455.6,96.69,348.18,357.82);
+INSERT INTO "financial_periods" VALUES(8,1,'Standard Retail Financial Period',256.23,273.99,444.19,309.48);
+INSERT INTO "financial_periods" VALUES(9,4,'Annual Retail Financial Period',492.16,150.68,90.31,284.24);
+INSERT INTO "financial_periods" VALUES(10,5,'Recurring Domestic Financial Period',380.99,223.94,493.04,352.72);
+INSERT INTO "financial_periods" VALUES(11,6,'Consolidated Domestic Financial Period',347.01,244.41,252.43,380.21);
+INSERT INTO "financial_periods" VALUES(12,5,'Annual Internal Financial Period',50.77,386.54,53.33,309.6);
+INSERT INTO "financial_periods" VALUES(13,6,'Priority Retail Financial Period',390.07,457.97,204.22,44.53);
+INSERT INTO "financial_periods" VALUES(14,9,'Quarterly Partner Financial Period',189.77,83.85,319.46,225.1);
+INSERT INTO "financial_periods" VALUES(15,10,'Recurring Internal Financial Period',298.75,76.22,15.09,122.31);
+INSERT INTO "financial_periods" VALUES(16,15,'Annual Regional Financial Period',160.24,108.71,361.66,156.11);
+INSERT INTO "financial_periods" VALUES(17,1,'Consolidated Retail Financial Period',17.26,477.49,325.07,476.72);
+INSERT INTO "financial_periods" VALUES(9101,9001,'2024',800.0,200.0,60.0,100.0);
+INSERT INTO "financial_periods" VALUES(9102,9001,'2025',900.0,230.0,65.0,120.0);
+INSERT INTO "financial_periods" VALUES(9103,9001,'2026',1000.0,260.0,70.0,140.0);
+INSERT INTO "financial_periods" VALUES(9111,9003,'2026',500.0,90.0,25.0,45.0);
+INSERT INTO "lbo_models" VALUES(1,6,466.55,491.87,319.16,166.15,'draft');
+INSERT INTO "lbo_models" VALUES(2,10,260.39,224.04,476.01,255.65,'approved');
+INSERT INTO "lbo_models" VALUES(3,5,10.62,44.85,474.78,418.23,'final');
+INSERT INTO "lbo_models" VALUES(4,10,468.69,188.71,317.79,78.29,'submitted');
+INSERT INTO "lbo_models" VALUES(5,2,495.44,269.53,198.77,246.93,'draft');
+INSERT INTO "lbo_models" VALUES(6,3,185.59,229.48,491.37,179.55,'final');
+INSERT INTO "lbo_models" VALUES(7,6,476.94,210.77,41.87,226.67,'draft');
+INSERT INTO "lbo_models" VALUES(8,7,171.97,272.36,340.15,326.52,'submitted');
+INSERT INTO "lbo_models" VALUES(9,1,25.51,93.21,235.93,362.09,'approved');
+INSERT INTO "lbo_models" VALUES(10,9,290.39,498.97,412.45,292.96,'draft');
+INSERT INTO "lbo_models" VALUES(11,11,117.87,381.73,53.27,308.8,'final');
+INSERT INTO "lbo_models" VALUES(12,4,161.34,70.37,369.84,345.68,'submitted');
+INSERT INTO "lbo_models" VALUES(13,12,56.61,386.86,404.51,220.36,'final');
+INSERT INTO "lbo_models" VALUES(14,15,432.42,298.39,238.92,171.4,'approved');
+INSERT INTO "lbo_models" VALUES(15,6,103.87,131.76,327.78,63.33,'draft');
+INSERT INTO "lbo_models" VALUES(16,15,220.69,56.81,68.3,156.03,'submitted');
+INSERT INTO "lbo_models" VALUES(17,15,54.91,402.97,468.67,472.96,'submitted');
+INSERT INTO "lbo_models" VALUES(9401,9003,8.0,9.0,4.5,21.5,'draft');
+INSERT INTO "review_notes" VALUES(1,11,'Consolidated Export Review Note','{"checksum": "6cd0f00d", "received_via": "csv_import", "record_type": "review_note", "reference": "RN-000001", "source": "web_portal", "version": 1}','on_hold');
+INSERT INTO "review_notes" VALUES(2,8,'Standard Retail Review Note','{"checksum": "92b660b0", "received_via": "webhook", "record_type": "review_note", "reference": "RN-000002", "source": "shopee", "version": 1}','closed');
+INSERT INTO "review_notes" VALUES(3,3,'Consolidated Export Review Note','{"checksum": "6beb5423", "received_via": "csv_import", "record_type": "review_note", "reference": "RN-000003", "source": "edi_gateway", "version": 4}','on_hold');
+INSERT INTO "review_notes" VALUES(4,9,'Standard Internal Review Note','{"checksum": "0da727cf", "received_via": "csv_import", "record_type": "review_note", "reference": "RN-000004", "source": "web_portal", "version": 2}','open');
+INSERT INTO "review_notes" VALUES(5,9,'Expedited Retail Review Note','{"checksum": "16c41be5", "received_via": "webhook", "record_type": "review_note", "reference": "RN-000005", "source": "edi_gateway", "version": 1}','closed');
+INSERT INTO "review_notes" VALUES(6,11,'Priority Retail Review Note','{"checksum": "72f4c4b8", "received_via": "api", "record_type": "review_note", "reference": "RN-000006", "source": "shopee", "version": 2}','closed');
+INSERT INTO "review_notes" VALUES(7,10,'Priority Partner Review Note','{"checksum": "d4901bce", "received_via": "webhook", "record_type": "review_note", "reference": "RN-000007", "source": "web_portal", "version": 3}','open');
+INSERT INTO "review_notes" VALUES(8,2,'Recurring Retail Review Note','{"checksum": "681af84e", "received_via": "manual_entry", "record_type": "review_note", "reference": "RN-000008", "source": "cyberbiz", "version": 4}','open');
+INSERT INTO "review_notes" VALUES(9,3,'Consolidated Partner Review Note','{"checksum": "6d5b4a56", "received_via": "csv_import", "record_type": "review_note", "reference": "RN-000009", "source": "cyberbiz", "version": 4}','closed');
+INSERT INTO "review_notes" VALUES(10,6,'Expedited Internal Review Note','{"checksum": "a450995c", "received_via": "webhook", "record_type": "review_note", "reference": "RN-000010", "source": "cyberbiz", "version": 3}','open');
+INSERT INTO "review_notes" VALUES(11,8,'Expedited Regional Review Note','{"checksum": "be53910d", "received_via": "manual_entry", "record_type": "review_note", "reference": "RN-000011", "source": "shopee", "version": 2}','closed');
+INSERT INTO "review_notes" VALUES(12,11,'Quarterly Retail Review Note','{"checksum": "f61a1a0a", "received_via": "csv_import", "record_type": "review_note", "reference": "RN-000012", "source": "shopee", "version": 3}','open');
+INSERT INTO "review_notes" VALUES(13,17,'Consolidated Domestic Review Note','{"checksum": "2fed51c3", "received_via": "manual_entry", "record_type": "review_note", "reference": "RN-001001", "source": "momo", "version": 4}','closed');
+INSERT INTO "review_notes" VALUES(14,15,'Annual Partner Review Note','{"checksum": "f2e73995", "received_via": "manual_entry", "record_type": "review_note", "reference": "RN-001002", "source": "erp_sync", "version": 2}','closed');
+INSERT INTO "review_notes" VALUES(15,1,'Recurring Export Review Note','{"checksum": "32caee16", "received_via": "manual_entry", "record_type": "review_note", "reference": "RN-001003", "source": "cyberbiz", "version": 4}','on_hold');
+INSERT INTO "review_notes" VALUES(16,10,'Standard Internal Review Note','{"checksum": "042d186d", "received_via": "manual_entry", "record_type": "review_note", "reference": "RN-001004", "source": "web_portal", "version": 1}','closed');
+INSERT INTO "review_notes" VALUES(17,6,'Standard Domestic Review Note','{"checksum": "d5be494e", "received_via": "csv_import", "record_type": "review_note", "reference": "RN-001005", "source": "cyberbiz", "version": 2}','open');
+INSERT INTO "review_notes" VALUES(9501,9202,'Head of Research','Rebuild before circulating.','open');
+INSERT INTO "sqlite_sequence" VALUES('fa_companies',9003);
+INSERT INTO "sqlite_sequence" VALUES('financial_periods',9111);
+INSERT INTO "sqlite_sequence" VALUES('dcf_models',9202);
+INSERT INTO "sqlite_sequence" VALUES('dcf_assumptions',9312);
+INSERT INTO "sqlite_sequence" VALUES('lbo_models',9401);
+INSERT INTO "sqlite_sequence" VALUES('review_notes',9501);

worlds/filingbench-50/world.json

@@ -0,0 +1,88 @@
+{
+  "world_id": "env_blobfish_benchmarks_filingbench_financia_599d3443",
+  "blobfish_version": "0.1.0",
+  "vertical": "financial_analysis",
+  "tenant": "blobfish-benchmarks",
+  "brief": "",
+  "topic": "financial analysis dcf model assumptions lbo review",
+  "domain": "financial_analysis",
+  "engine": "builtin",
+  "seed": 20260827,
+  "created_at": 1787733344.49571,
+  "world_digest": "cd3085c1e21b8793bd4ba01fe48bd2a711e793f8233d931b4944fa1bafadaf41",
+  "entities": [
+    "fa_companies",
+    "financial_periods",
+    "dcf_models",
+    "dcf_assumptions",
+    "lbo_models",
+    "review_notes"
+  ],
+  "tables_without_tools": [],
+  "counts": {
+    "tables": 6,
+    "rows": 116,
+    "tools": 44,
+    "tasks": 4,
+    "tasks_rejected": 0
+  },
+  "difficulty": {
+    "hard": 1,
+    "medium": 2,
+    "easy": 1
+  },
+  "splits": {
+    "train": [
+      "fa_finalize_solstice_dcf",
+      "fa_revert_model_without_discount_rate",
+      "fa_duplicate_company_listing"
+    ],
+    "heldout": [
+      "fa_dedupe_dcf_assumption"
+    ]
+  },
+  "task_creation": {
+    "schema_version": 1,
+    "status": "PASS",
+    "candidate_tasks": 4,
+    "accepted_tasks": 4,
+    "rejected_tasks": 0,
+    "rejections_by_gate": {},
+    "solvability": {
+      "oracle_replay_passed": 4,
+      "oracle_replay_failed": 0,
+      "all_admitted_tasks_oracle_solvable": true,
+      "impossible_candidates_removed_before_packaging": 0
+    },
+    "verifier_attribution": {
+      "all_admitted_tasks_attribution_safe": true,
+      "null_rollout_and_reference_replay_required": true,
+      "full_logical_state_protected_for_mutations": true
+    },
+    "workflow_coherence": {
+      "graph_candidates": 0,
+      "dependency_backed_candidates": 0,
+      "admitted_graph_tasks": 0,
+      "all_admitted_graph_tasks_dependency_backed": true,
+      "independent_connectivity_edges_allowed": false
+    },
+    "instruction_grounding": {
+      "graph_candidates": 0,
+      "fully_grounded_candidates": 0,
+      "admitted_graph_tasks": 0,
+      "all_admitted_graph_tasks_fully_determined_by_prompt": true,
+      "oracle_only_hidden_plans_allowed": false
+    },
+    "difficulty_calibration": {
+      "status": "NOT_MEASURED",
+      "claim_allowed": false,
+      "reason": "oracle success and verifier discrimination prove solvability, not model difficulty; a provider-identified stronger/less-capable model pair is required for a capability-separation claim"
+    }
+  },
+  "personas": [
+    "dbc93e793a83fd97",
+    "99bb97c76e367c3b",
+    "16e78d2fe0263069",
+    "fb74c0a1bc263f77"
+  ]
+}

worlds/opsbench-14/mcp/manifest.json

@@ -0,0 +1,68 @@
+{
+  "schema_version": "blobfish.mcp-service-layout.v1",
+  "world_id": "env_blobfish_benchmarks_opsbench_enterprise__8377a7f8",
+  "grouping": "tool.forge_service provenance only; no name/description inference (spec invariant 10)",
+  "systems": [],
+  "server_count": 1,
+  "tool_count": 42,
+  "servers": [
+    {
+      "name": "core",
+      "kind": "core",
+      "home": "core/",
+      "tool_count": 42,
+      "tools": [
+        "count_change_requests",
+        "count_compliance_controls",
+        "count_process_docs",
+        "count_process_steps",
+        "count_vendor_reviews",
+        "count_vendors",
+        "create_change_request",
+        "create_compliance_control",
+        "create_process_doc",
+        "create_process_step",
+        "create_vendor",
+        "create_vendor_review",
+        "delete_change_request",
+        "delete_compliance_control",
+        "delete_process_doc",
+        "delete_process_step",
+        "delete_vendor",
+        "delete_vendor_review",
+        "get_change_request",
+        "get_compliance_control",
+        "get_process_doc",
+        "get_process_step",
+        "get_vendor",
+        "get_vendor_review",
+        "list_change_requests_for_process_doc",
+        "list_process_steps_for_process_doc",
+        "list_vendor_reviews_for_vendor",
+        "search_change_requests",
+        "search_compliance_controls",
+        "search_process_docs",
+        "search_process_steps",
+        "search_vendor_reviews",
+        "search_vendors",
+        "transition_change_request_status",
+        "transition_vendor_review_status",
+        "transition_vendor_status",
+        "update_change_request",
+        "update_compliance_control",
+        "update_process_doc",
+        "update_process_step",
+        "update_vendor",
+        "update_vendor_review"
+      ],
+      "tables": [
+        "change_requests",
+        "compliance_controls",
+        "process_docs",
+        "process_steps",
+        "vendor_reviews",
+        "vendors"
+      ]
+    }
+  ]
+}

worlds/opsbench-14/schema.sql

@@ -0,0 +1,7 @@
+CREATE TABLE change_requests (request_id INTEGER PRIMARY KEY AUTOINCREMENT, doc_id INTEGER REFERENCES process_docs(doc_id), title TEXT NOT NULL, risk_level TEXT DEFAULT 'low', status TEXT DEFAULT 'submitted', requested_by TEXT, CHECK ("status" IS NULL OR "status" IN ('graded', 'rejected', 'returned', 'reviewing', 'submitted')));
+CREATE TABLE compliance_controls (control_id INTEGER PRIMARY KEY AUTOINCREMENT, name TEXT NOT NULL, framework TEXT, status TEXT DEFAULT 'untested', last_tested_on TEXT);
+CREATE TABLE process_docs (doc_id INTEGER PRIMARY KEY AUTOINCREMENT, name TEXT NOT NULL, owner_name TEXT, status TEXT DEFAULT 'published', version INTEGER DEFAULT 1);
+CREATE TABLE process_steps (step_id INTEGER PRIMARY KEY AUTOINCREMENT, doc_id INTEGER REFERENCES process_docs(doc_id), step_number INTEGER DEFAULT 1, description TEXT, control_ref TEXT);
+CREATE TABLE vendor_reviews (review_id INTEGER PRIMARY KEY AUTOINCREMENT, vendor_id INTEGER REFERENCES vendors(vendor_id), reviewer_name TEXT, score REAL DEFAULT 0, status TEXT DEFAULT 'open', reviewed_on TEXT, CHECK ("status" IS NULL OR "status" IN ('open', 'closed', 'on_hold')));
+CREATE TABLE vendors (vendor_id INTEGER PRIMARY KEY AUTOINCREMENT, name TEXT NOT NULL, category TEXT, tier TEXT DEFAULT 'tier3', status TEXT DEFAULT 'active', renewal_on TEXT, CHECK ("status" IS NULL OR "status" IN ('active', 'inactive', 'renewal_overdue', 'suspended')));
+DELETE FROM "sqlite_sequence";

worlds/opsbench-14/seed.sql

@@ -0,0 +1,122 @@
+INSERT INTO "change_requests" VALUES(1,7,'Quarterly Regional Change Request','low','submitted','Betty Brown');
+INSERT INTO "change_requests" VALUES(2,9,'Expedited Regional Change Request','low','graded','Patricia Smith');
+INSERT INTO "change_requests" VALUES(3,7,'Scheduled Export Change Request','low','reviewing','John Moore');
+INSERT INTO "change_requests" VALUES(4,2,'Priority Internal Change Request','low','returned','Susan Clark');
+INSERT INTO "change_requests" VALUES(5,5,'Recurring Regional Change Request','low','returned','Mary Hernandez');
+INSERT INTO "change_requests" VALUES(6,3,'Annual Export Change Request','low','submitted','Jessica Smith');
+INSERT INTO "change_requests" VALUES(7,3,'Quarterly Partner Change Request','low','reviewing','Andrew White');
+INSERT INTO "change_requests" VALUES(8,2,'Annual Domestic Change Request','low','reviewing','Elizabeth Moore');
+INSERT INTO "change_requests" VALUES(9,8,'Scheduled Retail Change Request','low','graded','Patricia Harris');
+INSERT INTO "change_requests" VALUES(10,9,'Consolidated Export Change Request','low','reviewing','Elizabeth Smith');
+INSERT INTO "change_requests" VALUES(11,5,'Expedited Retail Change Request','low','rejected','Kimberly Wilson');
+INSERT INTO "change_requests" VALUES(12,11,'Quarterly Regional Change Request','low','reviewing','Linda Williams');
+INSERT INTO "change_requests" VALUES(13,3,'Standard Export Change Request','low','rejected','Kenneth Moore');
+INSERT INTO "change_requests" VALUES(14,10,'Consolidated Regional Change Request','low','graded','Joshua Taylor');
+INSERT INTO "change_requests" VALUES(15,14,'Consolidated Regional Change Request','low','returned','Sarah Smith');
+INSERT INTO "change_requests" VALUES(16,12,'Scheduled Retail Change Request','low','rejected','Robert Thompson');
+INSERT INTO "change_requests" VALUES(17,8,'Annual Export Change Request','low','submitted','Steven Clark');
+INSERT INTO "change_requests" VALUES(9301,9101,'Skip sanctions screening for tier3','high','submitted','Riley Garcia');
+INSERT INTO "change_requests" VALUES(9302,9102,'Add secondary pager','low','submitted','Quinn Okafor');
+INSERT INTO "compliance_controls" VALUES(1,'Priority Domestic Compliance Control','Backup Framework B54','untested','General Last Tested On B15');
+INSERT INTO "compliance_controls" VALUES(2,'Standard Domestic Compliance Control','Primary Framework C98','untested','Regional Last Tested On C52');
+INSERT INTO "compliance_controls" VALUES(3,'Scheduled Domestic Compliance Control','Backup Framework A77','untested','Standard Last Tested On C99');
+INSERT INTO "compliance_controls" VALUES(4,'Consolidated Partner Compliance Control','Backup Framework B40','untested','Regional Last Tested On B48');
+INSERT INTO "compliance_controls" VALUES(5,'Scheduled Internal Compliance Control','General Framework A86','untested','Secondary Last Tested On D67');
+INSERT INTO "compliance_controls" VALUES(6,'Scheduled Internal Compliance Control','Backup Framework D23','untested','General Last Tested On F93');
+INSERT INTO "compliance_controls" VALUES(7,'Annual Export Compliance Control','Preferred Framework F59','untested','Consolidated Last Tested On A93');
+INSERT INTO "compliance_controls" VALUES(8,'Annual Domestic Compliance Control','Consolidated Framework C53','untested','Backup Last Tested On C61');
+INSERT INTO "compliance_controls" VALUES(9,'Recurring Regional Compliance Control','Standard Framework C89','untested','Standard Last Tested On B50');
+INSERT INTO "compliance_controls" VALUES(10,'Standard Domestic Compliance Control','Regional Framework F67','untested','General Last Tested On D12');
+INSERT INTO "compliance_controls" VALUES(11,'Consolidated Regional Compliance Control','General Framework C36','untested','Primary Last Tested On E84');
+INSERT INTO "compliance_controls" VALUES(12,'Recurring Internal Compliance Control','Consolidated Framework F38','untested','Preferred Last Tested On C66');
+INSERT INTO "compliance_controls" VALUES(13,'Recurring Partner Compliance Control','Secondary Framework D50','untested','Primary Last Tested On F98');
+INSERT INTO "compliance_controls" VALUES(14,'Consolidated Partner Compliance Control','Consolidated Framework B51','untested','Backup Last Tested On F21');
+INSERT INTO "compliance_controls" VALUES(15,'Expedited Partner Compliance Control','Primary Framework C76','untested','Primary Last Tested On D24');
+INSERT INTO "compliance_controls" VALUES(16,'Consolidated Retail Compliance Control','Secondary Framework D73','untested','General Last Tested On D40');
+INSERT INTO "compliance_controls" VALUES(17,'Quarterly Partner Compliance Control','Primary Framework C71','untested','Consolidated Last Tested On E45');
+INSERT INTO "compliance_controls" VALUES(9401,'Sanctions screening','SOC2','passing','2026-06-30');
+INSERT INTO "process_docs" VALUES(1,'Annual Retail Process Doc','Quarterly Retail Process Doc','published',1);
+INSERT INTO "process_docs" VALUES(2,'Scheduled Export Process Doc','Annual Internal Process Doc','published',1);
+INSERT INTO "process_docs" VALUES(3,'Quarterly Internal Process Doc','Consolidated Export Process Doc','published',1);
+INSERT INTO "process_docs" VALUES(4,'Scheduled Partner Process Doc','Scheduled Domestic Process Doc','published',1);
+INSERT INTO "process_docs" VALUES(5,'Annual Internal Process Doc','Priority Regional Process Doc','published',1);
+INSERT INTO "process_docs" VALUES(6,'Expedited Regional Process Doc','Annual Internal Process Doc','published',1);
+INSERT INTO "process_docs" VALUES(7,'Expedited Export Process Doc','Standard Retail Process Doc','published',1);
+INSERT INTO "process_docs" VALUES(8,'Recurring Export Process Doc','Standard Regional Process Doc','published',1);
+INSERT INTO "process_docs" VALUES(9,'Standard Internal Process Doc','Recurring Internal Process Doc','published',1);
+INSERT INTO "process_docs" VALUES(10,'Priority Partner Process Doc','Recurring Internal Process Doc','published',1);
+INSERT INTO "process_docs" VALUES(11,'Scheduled Retail Process Doc','Consolidated Partner Process Doc','published',1);
+INSERT INTO "process_docs" VALUES(12,'Expedited Export Process Doc','Annual Domestic Process Doc','published',1);
+INSERT INTO "process_docs" VALUES(13,'Quarterly Retail Process Doc','Annual Regional Process Doc','published',1);
+INSERT INTO "process_docs" VALUES(14,'Recurring Retail Process Doc','Expedited Retail Process Doc','published',1);
+INSERT INTO "process_docs" VALUES(15,'Standard Regional Process Doc','Consolidated Regional Process Doc','published',1);
+INSERT INTO "process_docs" VALUES(16,'Consolidated Retail Process Doc','Scheduled Regional Process Doc','published',1);
+INSERT INTO "process_docs" VALUES(17,'Scheduled Domestic Process Doc','Expedited Partner Process Doc','published',1);
+INSERT INTO "process_docs" VALUES(9101,'Vendor Onboarding','Jordan Patel','published',3);
+INSERT INTO "process_docs" VALUES(9102,'Incident Escalation','Jordan Patel','published',2);
+INSERT INTO "process_steps" VALUES(1,7,1,'Verified and approved','CON-550001');
+INSERT INTO "process_steps" VALUES(2,2,1,'No issues found','CON-770002');
+INSERT INTO "process_steps" VALUES(3,4,1,'Verified and approved','CON-250003');
+INSERT INTO "process_steps" VALUES(4,4,1,'Updated process step per request','CON-190004');
+INSERT INTO "process_steps" VALUES(5,4,1,'Updated process step per request','CON-770005');
+INSERT INTO "process_steps" VALUES(6,12,1,'Pending review','CON-640006');
+INSERT INTO "process_steps" VALUES(7,12,1,'Follow-up required','CON-210007');
+INSERT INTO "process_steps" VALUES(8,9,1,'Follow-up required','CON-450008');
+INSERT INTO "process_steps" VALUES(9,4,1,'Updated process step per request','CON-470009');
+INSERT INTO "process_steps" VALUES(10,8,1,'No issues found','CON-390010');
+INSERT INTO "process_steps" VALUES(11,4,1,'Escalated for attention','CON-570011');
+INSERT INTO "process_steps" VALUES(12,6,1,'No issues found','CON-290012');
+INSERT INTO "process_steps" VALUES(13,4,1,'No issues found','CON-391001');
+INSERT INTO "process_steps" VALUES(14,12,1,'No issues found','CON-441002');
+INSERT INTO "process_steps" VALUES(15,10,1,'Verified and approved','CON-761003');
+INSERT INTO "process_steps" VALUES(16,14,1,'Verified and approved','CON-181004');
+INSERT INTO "process_steps" VALUES(17,17,1,'Routine process step entry','CON-911005');
+INSERT INTO "process_steps" VALUES(9201,9101,1,'Collect vendor tax documentation','CTRL-014');
+INSERT INTO "process_steps" VALUES(9202,9101,1,'Collect vendor tax documentation','CTRL-014');
+INSERT INTO "process_steps" VALUES(9203,9101,2,'Run sanctions screening','CTRL-021');
+INSERT INTO "process_steps" VALUES(9204,9101,3,'Grant system access','');
+INSERT INTO "process_steps" VALUES(9211,9102,1,'Page the on-call owner','CTRL-033');
+INSERT INTO "vendor_reviews" VALUES(1,7,'Anderson Corp',370.68,'on_hold','General Reviewed On E80');
+INSERT INTO "vendor_reviews" VALUES(2,11,'Clark Continental LLC',130.2,'on_hold','Consolidated Reviewed On C44');
+INSERT INTO "vendor_reviews" VALUES(3,8,'Perez LLC',269.6,'open','General Reviewed On F36');
+INSERT INTO "vendor_reviews" VALUES(4,3,'Lewis Enterprises',194.16,'on_hold','Secondary Reviewed On F10');
+INSERT INTO "vendor_reviews" VALUES(5,6,'Clark Group',265.74,'open','Standard Reviewed On B50');
+INSERT INTO "vendor_reviews" VALUES(6,12,'White Logistics',266.62,'open','Preferred Reviewed On C84');
+INSERT INTO "vendor_reviews" VALUES(7,6,'Jones Holdings',29.09,'open','General Reviewed On E71');
+INSERT INTO "vendor_reviews" VALUES(8,1,'Lewis Industrial Group',93.51,'closed','Consolidated Reviewed On C64');
+INSERT INTO "vendor_reviews" VALUES(9,6,'Jackson Pacific Trading Co',79.64,'closed','Backup Reviewed On F72');
+INSERT INTO "vendor_reviews" VALUES(10,5,'Williams LLC',442.94,'closed','Regional Reviewed On A28');
+INSERT INTO "vendor_reviews" VALUES(11,7,'Young Allied Partners',445.58,'open','Backup Reviewed On D78');
+INSERT INTO "vendor_reviews" VALUES(12,6,'Anderson Industries',363.24,'closed','Regional Reviewed On F58');
+INSERT INTO "vendor_reviews" VALUES(13,9,'Gonzalez Trading Co',356.07,'on_hold','Preferred Reviewed On D40');
+INSERT INTO "vendor_reviews" VALUES(14,17,'Jones Industrial Enterprises',368.77,'closed','Consolidated Reviewed On A92');
+INSERT INTO "vendor_reviews" VALUES(15,6,'Taylor Solutions',467.97,'on_hold','Backup Reviewed On E76');
+INSERT INTO "vendor_reviews" VALUES(16,15,'Sanchez Northern Solutions',161.63,'open','Consolidated Reviewed On F91');
+INSERT INTO "vendor_reviews" VALUES(17,8,'White Precision Supply Co',323.65,'open','Primary Reviewed On F26');
+INSERT INTO "vendor_reviews" VALUES(9501,9003,'Jordan Patel',4.5,'closed','2026-05-20');
+INSERT INTO "vendors" VALUES(1,'Gonzalez Enterprises','Safety Equipment','tier3','active','General Renewal On E97');
+INSERT INTO "vendors" VALUES(2,'Garcia United Partners','Consumables','tier3','suspended','Consolidated Renewal On F75');
+INSERT INTO "vendors" VALUES(3,'Harris Corp','Packaging','tier3','inactive','Standard Renewal On B84');
+INSERT INTO "vendors" VALUES(4,'Davis Northern Trading Co','Tools & Hardware','tier3','renewal_overdue','Backup Renewal On E53');
+INSERT INTO "vendors" VALUES(5,'Wilson Trading Co','Tools & Hardware','tier3','renewal_overdue','Regional Renewal On F89');
+INSERT INTO "vendors" VALUES(6,'Garcia Allied Logistics','Packaging','tier3','renewal_overdue','Preferred Renewal On A13');
+INSERT INTO "vendors" VALUES(7,'Allen United Inc','Office Supplies','tier3','inactive','General Renewal On E94');
+INSERT INTO "vendors" VALUES(8,'Martin Enterprises','Tools & Hardware','tier3','renewal_overdue','General Renewal On E18');
+INSERT INTO "vendors" VALUES(9,'Ramirez Logistics','Electronics','tier3','active','Standard Renewal On D13');
+INSERT INTO "vendors" VALUES(10,'Anderson Enterprises','Logistics Equipment','tier3','suspended','Consolidated Renewal On A87');
+INSERT INTO "vendors" VALUES(11,'Perez Meridian Solutions','Office Supplies','tier3','inactive','Primary Renewal On D83');
+INSERT INTO "vendors" VALUES(12,'Gonzalez Pacific Corp','Consumables','tier3','renewal_overdue','Secondary Renewal On A34');
+INSERT INTO "vendors" VALUES(13,'Gonzalez Coastal Corp','Packaging','tier3','renewal_overdue','Standard Renewal On F42');
+INSERT INTO "vendors" VALUES(14,'Rodriguez Logistics','Packaging','tier3','inactive','General Renewal On B78');
+INSERT INTO "vendors" VALUES(15,'Young Meridian Logistics','Packaging','tier3','suspended','General Renewal On F64');
+INSERT INTO "vendors" VALUES(16,'Young Global Partners','Electronics','tier3','active','Consolidated Renewal On E54');
+INSERT INTO "vendors" VALUES(17,'Lee Supply Co','Safety Equipment','tier3','active','General Renewal On D84');
+INSERT INTO "vendors" VALUES(9001,'Cascade Facilities Ltd','facilities','tier2','active','2026-07-10');
+INSERT INTO "vendors" VALUES(9002,'Cascade Facilities Ltd','facilities','tier3','active','2026-11-30');
+INSERT INTO "vendors" VALUES(9003,'Orbit Freight','logistics','tier1','active','2027-01-15');
+INSERT INTO "sqlite_sequence" VALUES('vendors',9003);
+INSERT INTO "sqlite_sequence" VALUES('process_docs',9102);
+INSERT INTO "sqlite_sequence" VALUES('process_steps',9211);
+INSERT INTO "sqlite_sequence" VALUES('change_requests',9302);
+INSERT INTO "sqlite_sequence" VALUES('compliance_controls',9401);
+INSERT INTO "sqlite_sequence" VALUES('vendor_reviews',9501);

worlds/opsbench-14/world.json

@@ -0,0 +1,88 @@
+{
+  "world_id": "env_blobfish_benchmarks_opsbench_enterprise__8377a7f8",
+  "blobfish_version": "0.1.0",
+  "vertical": "operations",
+  "tenant": "blobfish-benchmarks",
+  "brief": "",
+  "topic": "operations vendor management process documentation change management compliance",
+  "domain": "operations",
+  "engine": "builtin",
+  "seed": 20260831,
+  "created_at": 1787733366.84014,
+  "world_digest": "39937737e9b493a2e1a4a10ba85780b15de76a222cf180538e40969cb6e3f1a8",
+  "entities": [
+    "vendors",
+    "process_docs",
+    "process_steps",
+    "change_requests",
+    "compliance_controls",
+    "vendor_reviews"
+  ],
+  "tables_without_tools": [],
+  "counts": {
+    "tables": 6,
+    "rows": 116,
+    "tools": 42,
+    "tasks": 4,
+    "tasks_rejected": 0
+  },
+  "difficulty": {
+    "hard": 1,
+    "medium": 2,
+    "easy": 1
+  },
+  "splits": {
+    "train": [
+      "ops_flag_overdue_vendor_renewal",
+      "ops_dedupe_process_step"
+    ],
+    "heldout": [
+      "ops_reject_uncontrolled_change",
+      "ops_duplicate_vendor_name"
+    ]
+  },
+  "task_creation": {
+    "schema_version": 1,
+    "status": "PASS",
+    "candidate_tasks": 4,
+    "accepted_tasks": 4,
+    "rejected_tasks": 0,
+    "rejections_by_gate": {},
+    "solvability": {
+      "oracle_replay_passed": 4,
+      "oracle_replay_failed": 0,
+      "all_admitted_tasks_oracle_solvable": true,
+      "impossible_candidates_removed_before_packaging": 0
+    },
+    "verifier_attribution": {
+      "all_admitted_tasks_attribution_safe": true,
+      "null_rollout_and_reference_replay_required": true,
+      "full_logical_state_protected_for_mutations": true
+    },
+    "workflow_coherence": {
+      "graph_candidates": 0,
+      "dependency_backed_candidates": 0,
+      "admitted_graph_tasks": 0,
+      "all_admitted_graph_tasks_dependency_backed": true,
+      "independent_connectivity_edges_allowed": false
+    },
+    "instruction_grounding": {
+      "graph_candidates": 0,
+      "fully_grounded_candidates": 0,
+      "admitted_graph_tasks": 0,
+      "all_admitted_graph_tasks_fully_determined_by_prompt": true,
+      "oracle_only_hidden_plans_allowed": false
+    },
+    "difficulty_calibration": {
+      "status": "NOT_MEASURED",
+      "claim_allowed": false,
+      "reason": "oracle success and verifier discrimination prove solvability, not model difficulty; a provider-identified stronger/less-capable model pair is required for a capability-separation claim"
+    }
+  },
+  "personas": [
+    "5330eed2b6315565",
+    "a9dcdac4e6c90e48",
+    "947e295f76ab223b",
+    "9184840c3734c10e"
+  ]
+}

worlds/repobench-500/schema.sql

@@ -0,0 +1,7 @@
+CREATE TABLE adr_records (adr_id INTEGER PRIMARY KEY AUTOINCREMENT, service_id INTEGER REFERENCES services(service_id), title TEXT NOT NULL, decision TEXT, status TEXT DEFAULT 'proposed', decided_on TEXT, CHECK ("status" IS NULL OR "status" IN ('accepted', 'interview', 'offered', 'placed', 'proposed', 'rejected', 'screening')));
+CREATE TABLE incidents (incident_id INTEGER PRIMARY KEY AUTOINCREMENT, service_id INTEGER REFERENCES services(service_id), summary TEXT, severity TEXT DEFAULT 'sev3', status TEXT DEFAULT 'open', detected_minute INTEGER DEFAULT 0, resolved_minute INTEGER, CHECK ("status" IS NULL OR "status" IN ('closed', 'on_hold', 'open', 'resolved')));
+CREATE TABLE pull_requests (pr_id INTEGER PRIMARY KEY AUTOINCREMENT, repo_id INTEGER REFERENCES repositories(repo_id), title TEXT NOT NULL, author TEXT, status TEXT DEFAULT 'open', approvals INTEGER DEFAULT 0, checks_status TEXT DEFAULT 'pending', CHECK ("status" IS NULL OR "status" IN ('approved', 'closed', 'on_hold', 'open')), CHECK ("checks_status" IS NULL OR "checks_status" IN ('approved', 'completed', 'failing', 'passing', 'pending', 'rejected')));
+CREATE TABLE repositories (repo_id INTEGER PRIMARY KEY AUTOINCREMENT, service_id INTEGER REFERENCES services(service_id), name TEXT NOT NULL, default_branch TEXT DEFAULT 'main', visibility TEXT DEFAULT 'internal');
+CREATE TABLE services (service_id INTEGER PRIMARY KEY AUTOINCREMENT, name TEXT NOT NULL, tier TEXT DEFAULT 'tier2', owner_team TEXT, oncall_name TEXT);
+CREATE TABLE standup_notes (note_id INTEGER PRIMARY KEY AUTOINCREMENT, service_id INTEGER REFERENCES services(service_id), author TEXT, day_label TEXT, body TEXT);
+DELETE FROM "sqlite_sequence";

worlds/repobench-500/seed.sql

@@ -0,0 +1,121 @@
+INSERT INTO "adr_records" VALUES(1,6,'Consolidated Regional Adr Record','Secondary Decision E10','placed','Regional Decided On E18');
+INSERT INTO "adr_records" VALUES(2,12,'Expedited Export Adr Record','Regional Decision A32','rejected','Primary Decided On D24');
+INSERT INTO "adr_records" VALUES(3,6,'Standard Partner Adr Record','Backup Decision E75','accepted','Regional Decided On C71');
+INSERT INTO "adr_records" VALUES(4,4,'Recurring Partner Adr Record','General Decision A85','proposed','Primary Decided On E18');
+INSERT INTO "adr_records" VALUES(5,7,'Scheduled Regional Adr Record','General Decision B14','accepted','Preferred Decided On F67');
+INSERT INTO "adr_records" VALUES(6,2,'Consolidated Partner Adr Record','Backup Decision E15','accepted','General Decided On D24');
+INSERT INTO "adr_records" VALUES(7,11,'Standard Internal Adr Record','Consolidated Decision C90','proposed','General Decided On E83');
+INSERT INTO "adr_records" VALUES(8,8,'Annual Retail Adr Record','Preferred Decision C27','offered','Regional Decided On E80');
+INSERT INTO "adr_records" VALUES(9,9,'Annual Domestic Adr Record','General Decision F67','interview','Regional Decided On E94');
+INSERT INTO "adr_records" VALUES(10,4,'Consolidated Domestic Adr Record','Backup Decision D87','accepted','Preferred Decided On D59');
+INSERT INTO "adr_records" VALUES(11,2,'Quarterly Partner Adr Record','General Decision C96','offered','Regional Decided On D35');
+INSERT INTO "adr_records" VALUES(12,4,'Annual Internal Adr Record','Primary Decision F40','accepted','Consolidated Decided On D26');
+INSERT INTO "adr_records" VALUES(13,4,'Expedited Regional Adr Record','Secondary Decision D98','accepted','Primary Decided On C78');
+INSERT INTO "adr_records" VALUES(14,17,'Standard Retail Adr Record','Secondary Decision E69','accepted','Backup Decided On E14');
+INSERT INTO "adr_records" VALUES(15,17,'Recurring Partner Adr Record','Primary Decision C56','interview','Preferred Decided On B65');
+INSERT INTO "adr_records" VALUES(16,13,'Priority Retail Adr Record','Standard Decision B52','placed','Primary Decided On D63');
+INSERT INTO "adr_records" VALUES(17,4,'Annual Export Adr Record','Regional Decision E22','rejected','Backup Decided On F57');
+INSERT INTO "adr_records" VALUES(9301,9002,'Idempotent capture keys','Use request-scoped keys.','accepted','2026-05-12');
+INSERT INTO "incidents" VALUES(1,9,'Escalated for attention','sev3','open',0,53);
+INSERT INTO "incidents" VALUES(2,2,'No issues found','sev3','closed',0,63);
+INSERT INTO "incidents" VALUES(3,10,'Escalated for attention','sev3','open',0,35);
+INSERT INTO "incidents" VALUES(4,11,'Verified and approved','sev3','open',0,8);
+INSERT INTO "incidents" VALUES(5,5,'Verified and approved','sev3','closed',0,24);
+INSERT INTO "incidents" VALUES(6,11,'No issues found','sev3','open',0,20);
+INSERT INTO "incidents" VALUES(7,10,'No issues found','sev3','resolved',0,89);
+INSERT INTO "incidents" VALUES(8,8,'Updated incident per request','sev3','closed',0,99);
+INSERT INTO "incidents" VALUES(9,12,'Updated incident per request','sev3','on_hold',0,75);
+INSERT INTO "incidents" VALUES(10,8,'Verified and approved','sev3','closed',0,89);
+INSERT INTO "incidents" VALUES(11,1,'Pending review','sev3','resolved',0,94);
+INSERT INTO "incidents" VALUES(12,6,'Verified and approved','sev3','resolved',0,45);
+INSERT INTO "incidents" VALUES(13,7,'Verified and approved','sev3','closed',0,60);
+INSERT INTO "incidents" VALUES(14,5,'Routine incident entry','sev3','closed',0,42);
+INSERT INTO "incidents" VALUES(15,15,'No issues found','sev3','resolved',0,58);
+INSERT INTO "incidents" VALUES(16,15,'Escalated for attention','sev3','open',0,83);
+INSERT INTO "incidents" VALUES(17,4,'Updated incident per request','sev3','open',0,61);
+INSERT INTO "incidents" VALUES(9401,9001,'Ingest queue saturated, consumers stalled','sev1','open',100,NULL);
+INSERT INTO "incidents" VALUES(9402,9002,'Elevated capture latency','sev3','resolved',300,330);
+INSERT INTO "pull_requests" VALUES(1,8,'Priority Regional Pull Request','Preferred Author D52','on_hold',0,'completed');
+INSERT INTO "pull_requests" VALUES(2,9,'Recurring Regional Pull Request','Primary Author C18','approved',0,'pending');
+INSERT INTO "pull_requests" VALUES(3,9,'Scheduled Retail Pull Request','General Author B16','open',0,'approved');
+INSERT INTO "pull_requests" VALUES(4,11,'Recurring Domestic Pull Request','Backup Author B37','closed',0,'rejected');
+INSERT INTO "pull_requests" VALUES(5,10,'Expedited Export Pull Request','Backup Author B86','on_hold',0,'failing');
+INSERT INTO "pull_requests" VALUES(6,7,'Recurring Partner Pull Request','Standard Author A47','open',0,'pending');
+INSERT INTO "pull_requests" VALUES(7,6,'Scheduled Internal Pull Request','Consolidated Author E65','on_hold',0,'completed');
+INSERT INTO "pull_requests" VALUES(8,2,'Priority Retail Pull Request','Regional Author D39','closed',0,'approved');
+INSERT INTO "pull_requests" VALUES(9,7,'Standard Partner Pull Request','Regional Author B42','closed',0,'passing');
+INSERT INTO "pull_requests" VALUES(10,5,'Priority Retail Pull Request','General Author C80','approved',0,'pending');
+INSERT INTO "pull_requests" VALUES(11,12,'Recurring Domestic Pull Request','Regional Author D21','approved',0,'passing');
+INSERT INTO "pull_requests" VALUES(12,10,'Annual Export Pull Request','Consolidated Author F50','open',0,'passing');
+INSERT INTO "pull_requests" VALUES(13,3,'Recurring Export Pull Request','Standard Author E44','approved',0,'approved');
+INSERT INTO "pull_requests" VALUES(14,5,'Standard Regional Pull Request','Secondary Author F23','open',0,'pending');
+INSERT INTO "pull_requests" VALUES(15,11,'Annual Export Pull Request','Secondary Author C63','on_hold',0,'passing');
+INSERT INTO "pull_requests" VALUES(16,2,'Recurring Regional Pull Request','Primary Author F41','on_hold',0,'completed');
+INSERT INTO "pull_requests" VALUES(17,4,'Recurring Retail Pull Request','Backup Author B68','open',0,'pending');
+INSERT INTO "pull_requests" VALUES(9201,9101,'Add ingest backpressure','Rowan Novak','approved',2,'failing');
+INSERT INTO "pull_requests" VALUES(9202,9103,'Retry failed captures','Emerson Sato','approved',2,'passing');
+INSERT INTO "pull_requests" VALUES(9203,9101,'Bump client SDK','Sage Ivanov','open',0,'pending');
+INSERT INTO "repositories" VALUES(1,10,'Priority Retail Repository','main','internal');
+INSERT INTO "repositories" VALUES(2,2,'Annual Export Repository','main','internal');
+INSERT INTO "repositories" VALUES(3,8,'Annual Internal Repository','main','internal');
+INSERT INTO "repositories" VALUES(4,4,'Recurring Internal Repository','main','internal');
+INSERT INTO "repositories" VALUES(5,8,'Standard Retail Repository','main','internal');
+INSERT INTO "repositories" VALUES(6,9,'Standard Partner Repository','main','internal');
+INSERT INTO "repositories" VALUES(7,3,'Standard Domestic Repository','main','internal');
+INSERT INTO "repositories" VALUES(8,8,'Expedited Export Repository','main','internal');
+INSERT INTO "repositories" VALUES(9,7,'Recurring Export Repository','main','internal');
+INSERT INTO "repositories" VALUES(10,4,'Consolidated Retail Repository','main','internal');
+INSERT INTO "repositories" VALUES(11,9,'Quarterly Domestic Repository','main','internal');
+INSERT INTO "repositories" VALUES(12,5,'Expedited Partner Repository','main','internal');
+INSERT INTO "repositories" VALUES(13,10,'Standard Retail Repository','main','internal');
+INSERT INTO "repositories" VALUES(14,16,'Standard Export Repository','main','internal');
+INSERT INTO "repositories" VALUES(15,5,'Recurring Internal Repository','main','internal');
+INSERT INTO "repositories" VALUES(16,12,'Consolidated Export Repository','main','internal');
+INSERT INTO "repositories" VALUES(17,10,'Priority Partner Repository','main','internal');
+INSERT INTO "repositories" VALUES(9101,9001,'ingest-worker','main','internal');
+INSERT INTO "repositories" VALUES(9102,9002,'ingest-worker','main','internal');
+INSERT INTO "repositories" VALUES(9103,9002,'billing-api','main','internal');
+INSERT INTO "services" VALUES(1,'Annual Regional Service','tier2','Consolidated Owner Team E50','Recurring Internal Service');
+INSERT INTO "services" VALUES(2,'Scheduled Export Service','tier2','Preferred Owner Team C42','Quarterly Export Service');
+INSERT INTO "services" VALUES(3,'Quarterly Partner Service','tier2','Primary Owner Team E98','Annual Regional Service');
+INSERT INTO "services" VALUES(4,'Priority Partner Service','tier2','Backup Owner Team E10','Priority Internal Service');
+INSERT INTO "services" VALUES(5,'Standard Partner Service','tier2','Backup Owner Team D54','Expedited Export Service');
+INSERT INTO "services" VALUES(6,'Quarterly Domestic Service','tier2','Primary Owner Team C27','Priority Internal Service');
+INSERT INTO "services" VALUES(7,'Annual Regional Service','tier2','Backup Owner Team B43','Quarterly Internal Service');
+INSERT INTO "services" VALUES(8,'Expedited Export Service','tier2','Backup Owner Team B80','Consolidated Export Service');
+INSERT INTO "services" VALUES(9,'Annual Export Service','tier2','Preferred Owner Team D42','Priority Domestic Service');
+INSERT INTO "services" VALUES(10,'Expedited Export Service','tier2','Standard Owner Team E51','Quarterly Retail Service');
+INSERT INTO "services" VALUES(11,'Expedited Partner Service','tier2','Secondary Owner Team C59','Expedited Domestic Service');
+INSERT INTO "services" VALUES(12,'Annual Internal Service','tier2','Backup Owner Team E40','Standard Domestic Service');
+INSERT INTO "services" VALUES(13,'Consolidated Internal Service','tier2','Standard Owner Team E20','Quarterly Domestic Service');
+INSERT INTO "services" VALUES(14,'Annual Retail Service','tier2','Regional Owner Team B81','Priority Domestic Service');
+INSERT INTO "services" VALUES(15,'Recurring Internal Service','tier2','Backup Owner Team B17','Consolidated Internal Service');
+INSERT INTO "services" VALUES(16,'Scheduled Regional Service','tier2','Secondary Owner Team E13','Recurring Partner Service');
+INSERT INTO "services" VALUES(17,'Expedited Retail Service','tier2','General Owner Team C37','Expedited Internal Service');
+INSERT INTO "services" VALUES(9001,'ingest','tier1','Data Platform','Rowan Novak');
+INSERT INTO "services" VALUES(9002,'billing','tier1','Payments','Emerson Sato');
+INSERT INTO "standup_notes" VALUES(1,6,'Preferred Author F27','Standard Regional Standup Note','{"checksum": "aa50040d", "received_via": "webhook", "record_type": "standup_note", "reference": "SN-000001", "source": "web_portal", "version": 2}');
+INSERT INTO "standup_notes" VALUES(2,2,'Primary Author E79','Consolidated Domestic Standup Note','{"checksum": "47f60a12", "received_via": "csv_import", "record_type": "standup_note", "reference": "SN-000002", "source": "edi_gateway", "version": 2}');
+INSERT INTO "standup_notes" VALUES(3,11,'Secondary Author F98','Consolidated Regional Standup Note','{"checksum": "7dd30ec6", "received_via": "webhook", "record_type": "standup_note", "reference": "SN-000003", "source": "edi_gateway", "version": 4}');
+INSERT INTO "standup_notes" VALUES(4,9,'General Author F68','Consolidated Internal Standup Note','{"checksum": "03dcbbc4", "received_via": "api", "record_type": "standup_note", "reference": "SN-000004", "source": "web_portal", "version": 2}');
+INSERT INTO "standup_notes" VALUES(5,3,'Preferred Author F14','Expedited Domestic Standup Note','{"checksum": "0a0a4c78", "received_via": "webhook", "record_type": "standup_note", "reference": "SN-000005", "source": "momo", "version": 4}');
+INSERT INTO "standup_notes" VALUES(6,12,'Standard Author D97','Standard Retail Standup Note','{"checksum": "e811b77a", "received_via": "api", "record_type": "standup_note", "reference": "SN-000006", "source": "erp_sync", "version": 1}');
+INSERT INTO "standup_notes" VALUES(7,4,'Primary Author B68','Consolidated Partner Standup Note','{"checksum": "263e5e1e", "received_via": "csv_import", "record_type": "standup_note", "reference": "SN-000007", "source": "cyberbiz", "version": 1}');
+INSERT INTO "standup_notes" VALUES(8,5,'Backup Author B39','Recurring Retail Standup Note','{"checksum": "7528e295", "received_via": "csv_import", "record_type": "standup_note", "reference": "SN-000008", "source": "erp_sync", "version": 2}');
+INSERT INTO "standup_notes" VALUES(9,4,'Preferred Author E63','Scheduled Partner Standup Note','{"checksum": "b7d41575", "received_via": "api", "record_type": "standup_note", "reference": "SN-000009", "source": "web_portal", "version": 1}');
+INSERT INTO "standup_notes" VALUES(10,8,'Regional Author B85','Priority Domestic Standup Note','{"checksum": "502cc3d7", "received_via": "manual_entry", "record_type": "standup_note", "reference": "SN-000010", "source": "momo", "version": 2}');
+INSERT INTO "standup_notes" VALUES(11,9,'Regional Author C55','Quarterly Internal Standup Note','{"checksum": "3a4fd466", "received_via": "csv_import", "record_type": "standup_note", "reference": "SN-000011", "source": "momo", "version": 1}');
+INSERT INTO "standup_notes" VALUES(12,9,'Regional Author E87','Consolidated Domestic Standup Note','{"checksum": "a0ed5418", "received_via": "csv_import", "record_type": "standup_note", "reference": "SN-000012", "source": "edi_gateway", "version": 3}');
+INSERT INTO "standup_notes" VALUES(13,1,'Consolidated Author B17','Recurring Domestic Standup Note','{"checksum": "75a564be", "received_via": "csv_import", "record_type": "standup_note", "reference": "SN-001001", "source": "shopee", "version": 2}');
+INSERT INTO "standup_notes" VALUES(14,17,'Preferred Author B77','Consolidated Retail Standup Note','{"checksum": "3230f3f2", "received_via": "webhook", "record_type": "standup_note", "reference": "SN-001002", "source": "shopee", "version": 4}');
+INSERT INTO "standup_notes" VALUES(15,6,'Preferred Author C10','Priority Partner Standup Note','{"checksum": "6f965453", "received_via": "csv_import", "record_type": "standup_note", "reference": "SN-001003", "source": "erp_sync", "version": 3}');
+INSERT INTO "standup_notes" VALUES(16,15,'Consolidated Author D95','Scheduled Internal Standup Note','{"checksum": "666cd64b", "received_via": "csv_import", "record_type": "standup_note", "reference": "SN-001004", "source": "edi_gateway", "version": 3}');
+INSERT INTO "standup_notes" VALUES(17,14,'Regional Author B83','Standard Domestic Standup Note','{"checksum": "1c9a4e1e", "received_via": "manual_entry", "record_type": "standup_note", "reference": "SN-001005", "source": "web_portal", "version": 4}');
+INSERT INTO "standup_notes" VALUES(9501,9001,'Sage Ivanov','monday','Picking up the backpressure PR.');
+INSERT INTO "standup_notes" VALUES(9502,9001,'Sage Ivanov','monday','Picking up the backpressure PR.');
+INSERT INTO "sqlite_sequence" VALUES('services',9002);
+INSERT INTO "sqlite_sequence" VALUES('repositories',9103);
+INSERT INTO "sqlite_sequence" VALUES('pull_requests',9203);
+INSERT INTO "sqlite_sequence" VALUES('adr_records',9301);
+INSERT INTO "sqlite_sequence" VALUES('incidents',9402);
+INSERT INTO "sqlite_sequence" VALUES('standup_notes',9502);

worlds/repobench-500/world.json

@@ -0,0 +1,88 @@
+{
+  "world_id": "env_blobfish_benchmarks_repobench_software_e_0af840ee",
+  "blobfish_version": "0.1.0",
+  "vertical": "engineering",
+  "tenant": "blobfish-benchmarks",
+  "brief": "",
+  "topic": "engineering standups code review architecture decisions incident response",
+  "domain": "engineering",
+  "engine": "builtin",
+  "seed": 20260901,
+  "created_at": 1787733366.716198,
+  "world_digest": "482b4cabbd31f60b18d71d5d709ae725e85c1a57f5736fdee8a40b4677546a4e",
+  "entities": [
+    "services",
+    "repositories",
+    "pull_requests",
+    "adr_records",
+    "incidents",
+    "standup_notes"
+  ],
+  "tables_without_tools": [],
+  "counts": {
+    "tables": 6,
+    "rows": 115,
+    "tools": 44,
+    "tasks": 4,
+    "tasks_rejected": 0
+  },
+  "difficulty": {
+    "hard": 1,
+    "medium": 2,
+    "easy": 1
+  },
+  "splits": {
+    "train": [
+      "eng_resolve_ingest_incident",
+      "eng_reopen_pr_with_red_ci",
+      "eng_duplicate_repository_name"
+    ],
+    "heldout": [
+      "eng_dedupe_standup_note"
+    ]
+  },
+  "task_creation": {
+    "schema_version": 1,
+    "status": "PASS",
+    "candidate_tasks": 4,
+    "accepted_tasks": 4,
+    "rejected_tasks": 0,
+    "rejections_by_gate": {},
+    "solvability": {
+      "oracle_replay_passed": 4,
+      "oracle_replay_failed": 0,
+      "all_admitted_tasks_oracle_solvable": true,
+      "impossible_candidates_removed_before_packaging": 0
+    },
+    "verifier_attribution": {
+      "all_admitted_tasks_attribution_safe": true,
+      "null_rollout_and_reference_replay_required": true,
+      "full_logical_state_protected_for_mutations": true
+    },
+    "workflow_coherence": {
+      "graph_candidates": 0,
+      "dependency_backed_candidates": 0,
+      "admitted_graph_tasks": 0,
+      "all_admitted_graph_tasks_dependency_backed": true,
+      "independent_connectivity_edges_allowed": false
+    },
+    "instruction_grounding": {
+      "graph_candidates": 0,
+      "fully_grounded_candidates": 0,
+      "admitted_graph_tasks": 0,
+      "all_admitted_graph_tasks_fully_determined_by_prompt": true,
+      "oracle_only_hidden_plans_allowed": false
+    },
+    "difficulty_calibration": {
+      "status": "NOT_MEASURED",
+      "claim_allowed": false,
+      "reason": "oracle success and verifier discrimination prove solvability, not model difficulty; a provider-identified stronger/less-capable model pair is required for a capability-separation claim"
+    }
+  },
+  "personas": [
+    "acc12e461ac38edf",
+    "487662642a40a65b",
+    "7a339fdf0ee2881e",
+    "00957e782ef2bd6d"
+  ]
+}

worlds/servicebench-375/schema.sql

@@ -0,0 +1,7 @@
+CREATE TABLE escalations (escalation_id INTEGER PRIMARY KEY AUTOINCREMENT, ticket_id INTEGER REFERENCES tickets(ticket_id), reason TEXT, tier TEXT DEFAULT 'tier2', status TEXT DEFAULT 'open', CHECK ("status" IS NULL OR "status" IN ('open', 'closed', 'on_hold')));
+CREATE TABLE kb_articles (article_id INTEGER PRIMARY KEY AUTOINCREMENT, title TEXT NOT NULL, body TEXT, status TEXT DEFAULT 'draft', related_ticket_id INTEGER REFERENCES tickets(ticket_id), CHECK ("status" IS NULL OR "status" IN ('approved', 'draft', 'final', 'published', 'submitted')));
+CREATE TABLE sla_policies (policy_id INTEGER PRIMARY KEY AUTOINCREMENT, name TEXT NOT NULL, plan TEXT, first_response_minutes INTEGER DEFAULT 60, resolve_minutes INTEGER DEFAULT 480);
+CREATE TABLE support_customers (customer_id INTEGER PRIMARY KEY AUTOINCREMENT, name TEXT NOT NULL, plan TEXT DEFAULT 'standard', region TEXT);
+CREATE TABLE ticket_messages (message_id INTEGER PRIMARY KEY AUTOINCREMENT, ticket_id INTEGER REFERENCES tickets(ticket_id), author TEXT, body TEXT, sent_minute INTEGER DEFAULT 0, channel TEXT DEFAULT 'email');
+CREATE TABLE tickets (ticket_id INTEGER PRIMARY KEY AUTOINCREMENT, customer_id INTEGER REFERENCES support_customers(customer_id), subject TEXT NOT NULL, status TEXT DEFAULT 'open', priority TEXT DEFAULT 'normal', opened_minute INTEGER DEFAULT 0, first_response_minute INTEGER, resolved_minute INTEGER, CHECK ("status" IS NULL OR "status" IN ('closed', 'escalated', 'on_hold', 'open', 'resolved')));
+DELETE FROM "sqlite_sequence";

worlds/servicebench-375/seed.sql

@@ -0,0 +1,121 @@
+INSERT INTO "escalations" VALUES(1,4,'Verified and approved','tier2','on_hold');
+INSERT INTO "escalations" VALUES(2,7,'Routine escalation entry','tier2','closed');
+INSERT INTO "escalations" VALUES(3,6,'Follow-up required','tier2','open');
+INSERT INTO "escalations" VALUES(4,12,'Follow-up required','tier2','closed');
+INSERT INTO "escalations" VALUES(5,11,'Escalated for attention','tier2','open');
+INSERT INTO "escalations" VALUES(6,2,'Follow-up required','tier2','closed');
+INSERT INTO "escalations" VALUES(7,10,'Follow-up required','tier2','on_hold');
+INSERT INTO "escalations" VALUES(8,10,'Updated escalation per request','tier2','open');
+INSERT INTO "escalations" VALUES(9,5,'Pending review','tier2','closed');
+INSERT INTO "escalations" VALUES(10,8,'Routine escalation entry','tier2','closed');
+INSERT INTO "escalations" VALUES(11,10,'Updated escalation per request','tier2','closed');
+INSERT INTO "escalations" VALUES(12,4,'Updated escalation per request','tier2','closed');
+INSERT INTO "escalations" VALUES(13,14,'Follow-up required','tier2','closed');
+INSERT INTO "escalations" VALUES(14,13,'Verified and approved','tier2','closed');
+INSERT INTO "escalations" VALUES(15,13,'Updated escalation per request','tier2','closed');
+INSERT INTO "escalations" VALUES(16,11,'Follow-up required','tier2','on_hold');
+INSERT INTO "escalations" VALUES(17,16,'Updated escalation per request','tier2','open');
+INSERT INTO "escalations" VALUES(9401,9203,'customer_request','tier2','closed');
+INSERT INTO "kb_articles" VALUES(1,'Quarterly Partner Kb Article','{"checksum": "8a09158b", "received_via": "webhook", "record_type": "kb_article", "reference": "KA-000001", "source": "web_portal", "version": 1}','final',10);
+INSERT INTO "kb_articles" VALUES(2,'Quarterly Domestic Kb Article','{"checksum": "2fd4ab60", "received_via": "manual_entry", "record_type": "kb_article", "reference": "KA-000002", "source": "erp_sync", "version": 1}','final',4);
+INSERT INTO "kb_articles" VALUES(3,'Consolidated Internal Kb Article','{"checksum": "3a0947b3", "received_via": "csv_import", "record_type": "kb_article", "reference": "KA-000003", "source": "cyberbiz", "version": 1}','submitted',10);
+INSERT INTO "kb_articles" VALUES(4,'Priority Partner Kb Article','{"checksum": "f4266090", "received_via": "api", "record_type": "kb_article", "reference": "KA-000004", "source": "shopee", "version": 3}','final',5);
+INSERT INTO "kb_articles" VALUES(5,'Quarterly Domestic Kb Article','{"checksum": "01d3ade5", "received_via": "api", "record_type": "kb_article", "reference": "KA-000005", "source": "edi_gateway", "version": 3}','approved',4);
+INSERT INTO "kb_articles" VALUES(6,'Quarterly Retail Kb Article','{"checksum": "fd263100", "received_via": "manual_entry", "record_type": "kb_article", "reference": "KA-000006", "source": "momo", "version": 1}','submitted',7);
+INSERT INTO "kb_articles" VALUES(7,'Quarterly Retail Kb Article','{"checksum": "e023e637", "received_via": "csv_import", "record_type": "kb_article", "reference": "KA-000007", "source": "cyberbiz", "version": 4}','submitted',7);
+INSERT INTO "kb_articles" VALUES(8,'Priority Internal Kb Article','{"checksum": "eab270f0", "received_via": "webhook", "record_type": "kb_article", "reference": "KA-000008", "source": "edi_gateway", "version": 4}','approved',12);
+INSERT INTO "kb_articles" VALUES(9,'Recurring Regional Kb Article','{"checksum": "798657f7", "received_via": "manual_entry", "record_type": "kb_article", "reference": "KA-000009", "source": "momo", "version": 2}','draft',8);
+INSERT INTO "kb_articles" VALUES(10,'Standard Retail Kb Article','{"checksum": "af008030", "received_via": "api", "record_type": "kb_article", "reference": "KA-000010", "source": "cyberbiz", "version": 1}','final',9);
+INSERT INTO "kb_articles" VALUES(11,'Consolidated Retail Kb Article','{"checksum": "9334fb0a", "received_via": "api", "record_type": "kb_article", "reference": "KA-000011", "source": "shopee", "version": 1}','published',11);
+INSERT INTO "kb_articles" VALUES(12,'Priority Domestic Kb Article','{"checksum": "f7db0943", "received_via": "manual_entry", "record_type": "kb_article", "reference": "KA-000012", "source": "edi_gateway", "version": 2}','approved',10);
+INSERT INTO "kb_articles" VALUES(13,'Quarterly Export Kb Article','{"checksum": "9ec5fbdf", "received_via": "api", "record_type": "kb_article", "reference": "KA-001001", "source": "web_portal", "version": 1}','approved',17);
+INSERT INTO "kb_articles" VALUES(14,'Priority Export Kb Article','{"checksum": "30a52261", "received_via": "manual_entry", "record_type": "kb_article", "reference": "KA-001002", "source": "momo", "version": 4}','draft',13);
+INSERT INTO "kb_articles" VALUES(15,'Annual Regional Kb Article','{"checksum": "604927d4", "received_via": "api", "record_type": "kb_article", "reference": "KA-001003", "source": "web_portal", "version": 3}','submitted',12);
+INSERT INTO "kb_articles" VALUES(16,'Annual Regional Kb Article','{"checksum": "9e963238", "received_via": "manual_entry", "record_type": "kb_article", "reference": "KA-001004", "source": "erp_sync", "version": 4}','final',5);
+INSERT INTO "kb_articles" VALUES(17,'Quarterly Export Kb Article','{"checksum": "4008f63a", "received_via": "manual_entry", "record_type": "kb_article", "reference": "KA-001005", "source": "web_portal", "version": 3}','published',3);
+INSERT INTO "kb_articles" VALUES(9501,'Recovering an offline fleet controller','Power-cycle then re-pair.','published',9201);
+INSERT INTO "kb_articles" VALUES(9502,'Recovering an offline fleet controller','Power-cycle then re-pair.','draft',9201);
+INSERT INTO "sla_policies" VALUES(1,'Consolidated Partner Sla Policy','Standard Plan C97',60,480);
+INSERT INTO "sla_policies" VALUES(2,'Annual Export Sla Policy','Standard Plan E36',60,480);
+INSERT INTO "sla_policies" VALUES(3,'Annual Domestic Sla Policy','Regional Plan B69',60,480);
+INSERT INTO "sla_policies" VALUES(4,'Priority Regional Sla Policy','Preferred Plan E64',60,480);
+INSERT INTO "sla_policies" VALUES(5,'Recurring Partner Sla Policy','Consolidated Plan C87',60,480);
+INSERT INTO "sla_policies" VALUES(6,'Consolidated Partner Sla Policy','Primary Plan F40',60,480);
+INSERT INTO "sla_policies" VALUES(7,'Expedited Regional Sla Policy','Preferred Plan B76',60,480);
+INSERT INTO "sla_policies" VALUES(8,'Priority Partner Sla Policy','Backup Plan E88',60,480);
+INSERT INTO "sla_policies" VALUES(9,'Scheduled Export Sla Policy','Standard Plan E75',60,480);
+INSERT INTO "sla_policies" VALUES(10,'Standard Regional Sla Policy','Preferred Plan F78',60,480);
+INSERT INTO "sla_policies" VALUES(11,'Annual Export Sla Policy','Secondary Plan D97',60,480);
+INSERT INTO "sla_policies" VALUES(12,'Annual Domestic Sla Policy','Consolidated Plan D34',60,480);
+INSERT INTO "sla_policies" VALUES(13,'Consolidated Export Sla Policy','Regional Plan F43',60,480);
+INSERT INTO "sla_policies" VALUES(14,'Expedited Regional Sla Policy','Consolidated Plan C46',60,480);
+INSERT INTO "sla_policies" VALUES(15,'Expedited Retail Sla Policy','Secondary Plan C59',60,480);
+INSERT INTO "sla_policies" VALUES(16,'Consolidated Export Sla Policy','Regional Plan C55',60,480);
+INSERT INTO "sla_policies" VALUES(17,'Priority Export Sla Policy','Standard Plan B98',60,480);
+INSERT INTO "sla_policies" VALUES(9101,'Enterprise SLA','enterprise',30,240);
+INSERT INTO "sla_policies" VALUES(9102,'Standard SLA','standard',120,960);
+INSERT INTO "support_customers" VALUES(1,'Wilson Northern Holdings','standard','Regional Region A73');
+INSERT INTO "support_customers" VALUES(2,'White Solutions','standard','Primary Region E37');
+INSERT INTO "support_customers" VALUES(3,'Ramirez Pacific Partners','standard','Consolidated Region C94');
+INSERT INTO "support_customers" VALUES(4,'Miller Corp','standard','Preferred Region C88');
+INSERT INTO "support_customers" VALUES(5,'Lopez Pacific Corp','standard','Standard Region F24');
+INSERT INTO "support_customers" VALUES(6,'Thomas Meridian Solutions','standard','Preferred Region C27');
+INSERT INTO "support_customers" VALUES(7,'Robinson Meridian Trading Co','standard','Primary Region D39');
+INSERT INTO "support_customers" VALUES(8,'Brown Allied Logistics','standard','Secondary Region C43');
+INSERT INTO "support_customers" VALUES(9,'Martin Corp','standard','Backup Region B18');
+INSERT INTO "support_customers" VALUES(10,'Jackson Group','standard','Secondary Region A27');
+INSERT INTO "support_customers" VALUES(11,'Hernandez LLC','standard','Regional Region F32');
+INSERT INTO "support_customers" VALUES(12,'Clark Meridian Holdings','standard','Standard Region D64');
+INSERT INTO "support_customers" VALUES(13,'Garcia Supply Co','standard','Standard Region D86');
+INSERT INTO "support_customers" VALUES(14,'Davis Coastal Group','standard','Preferred Region B34');
+INSERT INTO "support_customers" VALUES(15,'King Allied Supply Co','standard','Regional Region C87');
+INSERT INTO "support_customers" VALUES(16,'Jones Trading Co','standard','General Region C79');
+INSERT INTO "support_customers" VALUES(17,'Thomas Meridian Supply Co','standard','Primary Region E24');
+INSERT INTO "support_customers" VALUES(9001,'Bluepeak Robotics','enterprise','EMEA');
+INSERT INTO "support_customers" VALUES(9002,'Bluepeak Robotics','standard','APAC');
+INSERT INTO "support_customers" VALUES(9003,'Corvus Analytics','standard','NA');
+INSERT INTO "ticket_messages" VALUES(1,2,'Regional Author C26','{"checksum": "9f0eefc0", "received_via": "api", "record_type": "ticket_message", "reference": "TM-000001", "source": "erp_sync", "version": 3}',0,'email');
+INSERT INTO "ticket_messages" VALUES(2,8,'Standard Author F18','{"checksum": "a9890112", "received_via": "api", "record_type": "ticket_message", "reference": "TM-000002", "source": "shopee", "version": 4}',0,'email');
+INSERT INTO "ticket_messages" VALUES(3,2,'Consolidated Author E81','{"checksum": "d996e4da", "received_via": "webhook", "record_type": "ticket_message", "reference": "TM-000003", "source": "cyberbiz", "version": 2}',0,'email');
+INSERT INTO "ticket_messages" VALUES(4,1,'Secondary Author B95','{"checksum": "a12182cf", "received_via": "webhook", "record_type": "ticket_message", "reference": "TM-000004", "source": "cyberbiz", "version": 4}',0,'email');
+INSERT INTO "ticket_messages" VALUES(5,10,'Standard Author B91','{"checksum": "5a66b808", "received_via": "csv_import", "record_type": "ticket_message", "reference": "TM-000005", "source": "shopee", "version": 3}',0,'email');
+INSERT INTO "ticket_messages" VALUES(6,11,'Secondary Author E88','{"checksum": "bcb92175", "received_via": "webhook", "record_type": "ticket_message", "reference": "TM-000006", "source": "web_portal", "version": 2}',0,'email');
+INSERT INTO "ticket_messages" VALUES(7,4,'Preferred Author D83','{"checksum": "8447c9b5", "received_via": "webhook", "record_type": "ticket_message", "reference": "TM-000007", "source": "cyberbiz", "version": 1}',0,'email');
+INSERT INTO "ticket_messages" VALUES(8,12,'Preferred Author D73','{"checksum": "8d234806", "received_via": "manual_entry", "record_type": "ticket_message", "reference": "TM-000008", "source": "cyberbiz", "version": 2}',0,'email');
+INSERT INTO "ticket_messages" VALUES(9,12,'Regional Author F15','{"checksum": "9ce00fd1", "received_via": "manual_entry", "record_type": "ticket_message", "reference": "TM-000009", "source": "erp_sync", "version": 2}',0,'email');
+INSERT INTO "ticket_messages" VALUES(10,2,'Regional Author C90','{"checksum": "9d2daac3", "received_via": "csv_import", "record_type": "ticket_message", "reference": "TM-000010", "source": "cyberbiz", "version": 3}',0,'email');
+INSERT INTO "ticket_messages" VALUES(11,8,'Regional Author F11','{"checksum": "3deb466a", "received_via": "csv_import", "record_type": "ticket_message", "reference": "TM-000011", "source": "momo", "version": 3}',0,'email');
+INSERT INTO "ticket_messages" VALUES(12,12,'Standard Author F90','{"checksum": "e67357e2", "received_via": "csv_import", "record_type": "ticket_message", "reference": "TM-000012", "source": "erp_sync", "version": 2}',0,'email');
+INSERT INTO "ticket_messages" VALUES(13,12,'General Author A85','{"checksum": "38cb8bda", "received_via": "manual_entry", "record_type": "ticket_message", "reference": "TM-001001", "source": "web_portal", "version": 3}',0,'email');
+INSERT INTO "ticket_messages" VALUES(14,4,'General Author F81','{"checksum": "76b1c76f", "received_via": "csv_import", "record_type": "ticket_message", "reference": "TM-001002", "source": "cyberbiz", "version": 2}',0,'email');
+INSERT INTO "ticket_messages" VALUES(15,14,'Consolidated Author D79','{"checksum": "d1bc72dc", "received_via": "csv_import", "record_type": "ticket_message", "reference": "TM-001003", "source": "cyberbiz", "version": 2}',0,'email');
+INSERT INTO "ticket_messages" VALUES(16,8,'General Author C56','{"checksum": "fefd3e75", "received_via": "csv_import", "record_type": "ticket_message", "reference": "TM-001004", "source": "web_portal", "version": 3}',0,'email');
+INSERT INTO "ticket_messages" VALUES(17,5,'Backup Author A62','{"checksum": "4964dfe8", "received_via": "manual_entry", "record_type": "ticket_message", "reference": "TM-001005", "source": "momo", "version": 1}',0,'email');
+INSERT INTO "ticket_messages" VALUES(9301,9201,'customer','Our whole fleet dropped offline.',100,'email');
+INSERT INTO "ticket_messages" VALUES(9302,9202,'agent','Patch deployed, please confirm.',190,'email');
+INSERT INTO "tickets" VALUES(1,5,'Preferred Subject F21','on_hold','normal',0,47,13);
+INSERT INTO "tickets" VALUES(2,5,'Preferred Subject B96','open','normal',0,84,96);
+INSERT INTO "tickets" VALUES(3,1,'Backup Subject C88','escalated','normal',0,64,3);
+INSERT INTO "tickets" VALUES(4,8,'Regional Subject C70','resolved','normal',0,66,57);
+INSERT INTO "tickets" VALUES(5,6,'Regional Subject E74','on_hold','normal',0,3,66);
+INSERT INTO "tickets" VALUES(6,2,'Primary Subject F46','on_hold','normal',0,41,61);
+INSERT INTO "tickets" VALUES(7,5,'General Subject B85','escalated','normal',0,1,95);
+INSERT INTO "tickets" VALUES(8,10,'Primary Subject A16','resolved','normal',0,56,42);
+INSERT INTO "tickets" VALUES(9,1,'General Subject C11','on_hold','normal',0,23,27);
+INSERT INTO "tickets" VALUES(10,7,'Secondary Subject E27','escalated','normal',0,48,62);
+INSERT INTO "tickets" VALUES(11,11,'Secondary Subject C43','on_hold','normal',0,72,71);
+INSERT INTO "tickets" VALUES(12,11,'Preferred Subject E68','escalated','normal',0,4,13);
+INSERT INTO "tickets" VALUES(13,14,'General Subject E44','open','normal',0,45,61);
+INSERT INTO "tickets" VALUES(14,8,'Consolidated Subject B42','open','normal',0,17,6);
+INSERT INTO "tickets" VALUES(15,1,'Backup Subject C62','escalated','normal',0,59,90);
+INSERT INTO "tickets" VALUES(16,3,'General Subject F35','closed','normal',0,94,29);
+INSERT INTO "tickets" VALUES(17,15,'Consolidated Subject A13','on_hold','normal',0,19,11);
+INSERT INTO "tickets" VALUES(9201,9001,'Fleet controller offline','open','normal',100,NULL,NULL);
+INSERT INTO "tickets" VALUES(9202,9003,'Export job fails','open','normal',60,75,200);
+INSERT INTO "tickets" VALUES(9203,9002,'Password reset loop','open','low',150,165,NULL);
+INSERT INTO "sqlite_sequence" VALUES('support_customers',9003);
+INSERT INTO "sqlite_sequence" VALUES('sla_policies',9102);
+INSERT INTO "sqlite_sequence" VALUES('tickets',9203);
+INSERT INTO "sqlite_sequence" VALUES('ticket_messages',9302);
+INSERT INTO "sqlite_sequence" VALUES('escalations',9401);
+INSERT INTO "sqlite_sequence" VALUES('kb_articles',9502);

worlds/servicebench-375/world.json

@@ -0,0 +1,88 @@
+{
+  "world_id": "env_blobfish_benchmarks_servicebench_custome_79e1b30f",
+  "blobfish_version": "0.1.0",
+  "vertical": "customer_support",
+  "tenant": "blobfish-benchmarks",
+  "brief": "",
+  "topic": "customer support tickets sla escalation knowledge base",
+  "domain": "support",
+  "engine": "builtin",
+  "seed": 20260828,
+  "created_at": 1787733344.299031,
+  "world_digest": "ab0d887316ac287e98532ff96cd1ec24f979922a054a8e02b58e4d2e75f37260",
+  "entities": [
+    "support_customers",
+    "sla_policies",
+    "tickets",
+    "ticket_messages",
+    "escalations",
+    "kb_articles"
+  ],
+  "tables_without_tools": [],
+  "counts": {
+    "tables": 6,
+    "rows": 115,
+    "tools": 43,
+    "tasks": 4,
+    "tasks_rejected": 0
+  },
+  "difficulty": {
+    "hard": 1,
+    "medium": 2,
+    "easy": 1
+  },
+  "splits": {
+    "train": [
+      "cs_escalate_sla_breach",
+      "cs_close_resolved_tickets",
+      "cs_duplicate_customer_name"
+    ],
+    "heldout": [
+      "cs_dedupe_kb_article"
+    ]
+  },
+  "task_creation": {
+    "schema_version": 1,
+    "status": "PASS",
+    "candidate_tasks": 4,
+    "accepted_tasks": 4,
+    "rejected_tasks": 0,
+    "rejections_by_gate": {},
+    "solvability": {
+      "oracle_replay_passed": 4,
+      "oracle_replay_failed": 0,
+      "all_admitted_tasks_oracle_solvable": true,
+      "impossible_candidates_removed_before_packaging": 0
+    },
+    "verifier_attribution": {
+      "all_admitted_tasks_attribution_safe": true,
+      "null_rollout_and_reference_replay_required": true,
+      "full_logical_state_protected_for_mutations": true
+    },
+    "workflow_coherence": {
+      "graph_candidates": 0,
+      "dependency_backed_candidates": 0,
+      "admitted_graph_tasks": 0,
+      "all_admitted_graph_tasks_dependency_backed": true,
+      "independent_connectivity_edges_allowed": false
+    },
+    "instruction_grounding": {
+      "graph_candidates": 0,
+      "fully_grounded_candidates": 0,
+      "admitted_graph_tasks": 0,
+      "all_admitted_graph_tasks_fully_determined_by_prompt": true,
+      "oracle_only_hidden_plans_allowed": false
+    },
+    "difficulty_calibration": {
+      "status": "NOT_MEASURED",
+      "claim_allowed": false,
+      "reason": "oracle success and verifier discrimination prove solvability, not model difficulty; a provider-identified stronger/less-capable model pair is required for a capability-separation claim"
+    }
+  },
+  "personas": [
+    "88d4da33d27b185d",
+    "693444da41966b33",
+    "edcc3a23753f0620",
+    "d3120004bdac6490"
+  ]
+}