Blobfish Litigation Bench v1.0 — 100 tasks, median 108 verified steps, deterministic verifiers

README.md

@@ -0,0 +1,125 @@
+---
+license: cc-by-4.0
+task_categories:
+- text-generation
+language:
+- en
+tags:
+- benchmarking
+- agent-evaluation
+- deterministic-verification
+- reinforcement-learning
+- legal
+- litigation
+- long-horizon
+- mcp
+size_categories:
+- n<1K
+configs:
+- config_name: default
+  data_files:
+  - split: dev
+    path: data/dev.jsonl
+---
+
+# Blobfish Litigation Bench
+
+Executable, long-horizon litigation environments for training and evaluating
+tool-using agents. **100 tasks, 104-113 verified tool-calling steps each (median 108),
+graded by deterministic state-diff — no LLM judge anywhere on the reward path.**
+
+Every task runs inside a stateful law-firm simulation exposed over MCP: nine
+litigation systems of record (matters, parties, claims, pleadings, motions, discovery_requests, depositions, exhibits, settlements) behind 47 product-shaped tools. The
+agent works a real matter through those systems while reading a mounted file of
+6-7 synthetic documents.
+
+> **Simulation only.** Every matter, party, document, attorney and figure is synthetic.
+
+## At a glance
+
+| | |
+|---|---|
+| Tasks | 100 |
+| Verified steps per task | 104-113 (median 108) |
+| Distinct scenarios | 10 |
+| Unique prompt skeletons | 100/100 |
+| Documents per task | 6-7 (docx/eml/pptx/xlsx) |
+| Grading | Deterministic state-diff + trace assertions |
+| LLM judge | **None** |
+| Oracle pass rate | 100/100 (1.00) |
+
+## Why deterministic grading
+
+Rubric-graded agent benchmarks ask a model whether another model's prose
+satisfied a criterion. That conflates the grader's judgement with the agent's
+capability, and it cannot tell a rubric miss from a model miss.
+
+Here every task ships an executable reference walk, and grading is a diff of
+world state before and after plus assertions over the recorded tool trace. A
+zero means the agent failed, not that a judge disagreed.
+
+## First results
+
+| Agent | Reward | Conditions failed | Tools missed | Systems never touched |
+|---|---|---|---|---|
+| Oracle (reference walk) | **1.00** | 0 | 0 | — |
+| Claude Sonnet 4.5 · motion to compel | 0.00 | 40 | 32 | calendar, contacts, documents, efiling, matters, time |
+| Claude Sonnet 4.5 · case chronology | 0.00 | 29 | 25 | calendar, contacts, matters, time |
+
+The oracle finishes every task. A frontier coding agent, given the same
+environment and the same mounted matter file, returns 0.0 on both — and the
+deterministic verifier says exactly which tools it never called and which work
+products it never created. The failure is not refusal or malformed output: the
+agent worked the litigation system and never completed the cross-application
+chain.
+
+**Two tasks is not a leaderboard.** These are the first end-to-end runs, not a
+scored evaluation, and no confidence interval is claimed. They are reported
+because they establish the thing that matters before any leaderboard: the
+benchmark discriminates.
+
+## Data schema
+
+| Field | Type | Description |
+|---|---|---|
+| `task_id` | string | Stable identifier |
+| `task_name` | string | Human-readable name |
+| `world_id` | string | Environment pointer |
+| `prompt` | string | Full agent instruction |
+| `context_files` | list | Mounted document manifest (name + role) |
+| `rubric` | list | Deterministic assertion names checked by the verifier |
+| `gold_output` | string | Reference walk summary (tool sequence) |
+| `metadata` | dict | scenario, work_type, posture, complication, phases, steps, docket, court |
+
+## Scenario coverage
+
+- **Motion to Compel Deficient Discovery Responses** — draft (10 variants, 9 phases)
+- **Privilege Review and Log for a Rolling Production** — review (10 variants, 8 phases)
+- **Deposition Preparation and Exhibit Assembly** — draft (10 variants, 9 phases)
+- **Litigation Hold Scoping and Custodian Identification** — advise (10 variants, 8 phases)
+- **Opposition to Motion for Summary Judgment** — draft (10 variants, 8 phases)
+- **Settlement Valuation and Authority Memorandum** — analyze (10 variants, 7 phases)
+- **Production Completeness Audit Against Requests** — compare (10 variants, 8 phases)
+- **Motion for Sanctions for Discovery Abuse** — draft (10 variants, 8 phases)
+- **Expert Disclosure Review and Exclusion Motion** — analyze (10 variants, 8 phases)
+- **Case Chronology Reconstruction from the Record** — analyze (10 variants, 8 phases)
+
+## Splits and release policy
+
+`dev` carries all 100 tasks with their full deterministic assertion sets. The
+environment image and reference walks are published alongside so results are
+reproducible rather than self-reported.
+
+## Running the tasks
+
+The executable environments are published to the Harbor registry:
+
+```bash
+harbor run -d blobfishai/litigation -a <your-agent>
+```
+
+## Intended use and licensing
+
+Released CC BY 4.0 for benchmarking and agent training. The corpus is entirely
+synthetic and contains no client data. Nothing here is legal advice; every task
+carries an attorney-validation limitation that a correct answer must preserve.