Upload folder using huggingface_hub

LICENSE-CODE

@@ -0,0 +1,2 @@
+Apache License 2.0
+https://www.apache.org/licenses/LICENSE-2.0

LICENSE-DATA

@@ -0,0 +1,2 @@
+Creative Commons Attribution 4.0 International
+https://creativecommons.org/licenses/by/4.0/

README.md

@@ -0,0 +1,101 @@
+---
+license: cc-by-4.0
+task_categories:
+- text-generation
+language:
+- en
+tags:
+- devops
+- sre
+- software-engineering
+- benchmark
+- agents
+- mcp
+- deterministic-evaluation
+pretty_name: DevOpsBench-100
+size_categories:
+- n<1K
+---
+
+# DevOpsBench-100
+
+DevOpsBench-100 is a synthetic long-horizon software-engineering / SRE agent
+benchmark: 100 tasks over one executable world ("NovaCart", a mid-size
+e-commerce SaaS) with 72 SQLite tables,
+1451 seeded rows, a 38-file monorepo with 417 commits,
+and 97 MCP tools spanning a first-party engineering stack
+(tickets, PRs, CI, deployments, canaries, migrations, feature flags, metrics,
+alerts, incidents, chat, knowledge base) plus deliberately disagreeing
+vendor-shaped surfaces (Jira, Linear, GitHub Issues, Prometheus, Sentry,
+PagerDuty, Confluence, spreadsheets) and Kubernetes.
+
+Tasks are outcome-only tickets (symptom + definition of done; company policy
+lives in the world's knowledge base, not the prompt). Reference trajectories
+run 4-34 tool calls (median 13).
+Acceptance is fully deterministic: each task ships an executable vcode
+verifier that checks the final world state and the append-only audit log,
+with anti-forgery table pins - no LLM judge, no network, no clock in the
+reward path.
+
+## What is included
+
+- `data/tasks.jsonl`: task records (`task_id`, `task_name`, `world_id`,
+  `prompt`, `context_files`, `rubric`, `gold_output`, `metadata`).
+- `tasks/`: one readable JSON record per task (includes the guided
+  instruction variant).
+- `world/`: the offline world source - stdlib MCP server, tool implementations,
+  seeded SQLite database, schema and seed SQL.
+- `verifiers/`: 100 standalone verifier scripts
+  (`python3 verify_<task>.py world.db` prints the full verdict).
+- `trajectories/`: one executed reference trajectory per task (JSONL).
+- `reports/`: measured build and qualification evidence.
+
+## Task families (19)
+
+| Family | Tasks |
+|---|---:|
+| Cross-system source of truth | 8 |
+| Error-rate SLO recovery | 8 |
+| Latency optimization | 8 |
+| API migration | 7 |
+| Change attribution | 7 |
+| Feature-flag operation | 7 |
+| Multi-service rollout | 7 |
+| Security incident response | 7 |
+| AIOps root-cause analysis | 6 |
+| Flaky-test remediation | 6 |
+| AIOps detection | 5 |
+| AIOps localization | 5 |
+| Cross-source reconciliation | 5 |
+| Code implementation | 4 |
+| Operational judgement / restraint | 4 |
+| Incident handover | 2 |
+| Long-horizon delivery | 2 |
+| Human-approval gated change | 1 |
+| Workspace scripting | 1 |
+
+## Objective release gates
+
+| Gate | Required | Measured |
+|---|---:|---:|
+| Tasks | 100 | 100 |
+| Oracle replays at reward 1.0 | 100/100 | see `reports/qualification.json` |
+| Deterministic verifier replays | 100/100 | see `reports/qualification.json` |
+| Negative-control false accepts | 0 | see `reports/qualification.json` |
+| LLM / network calls in verifier | 0 | 0 |
+
+## Provenance and contamination
+
+Every service, metric, document, commit, and incident is synthetic and was
+generated for the NovaCart world. 30 cross_system/handover tasks were ported
+from TheAgentCompany task *shapes* re-grounded onto NovaCart's own state; the
+AIOps families reproduce the microsoft/AIOpsLab task *structure*
+(detect / localize / analyze) against NovaCart. No third-party benchmark
+text, evidence, or gold answers are included. Gold outputs are public, so
+this release is appropriate for transparent evaluation and RL experiments
+rather than secret-test claims.
+
+## Licenses
+
+Synthetic task data and world content are CC-BY-4.0. Benchmark code,
+server, and verifiers are Apache-2.0.

reports/build.json

@@ -0,0 +1,53 @@
+{
+  "benchmark": "DevOpsBench-100",
+  "categories": {
+    "aiops_analysis": 6,
+    "aiops_detection": 5,
+    "aiops_localization": 5,
+    "api_migration": 7,
+    "attribution": 7,
+    "code_implementation": 4,
+    "cross_system": 8,
+    "error_rate_reduction": 8,
+    "feature_flag": 7,
+    "flaky_test": 6,
+    "handover": 2,
+    "horizon": 2,
+    "human_gated": 1,
+    "judgement": 4,
+    "latency_optimization": 8,
+    "multi_service_rollout": 7,
+    "reconciliation": 5,
+    "security_incident": 7,
+    "workspace": 1
+  },
+  "category_count": 19,
+  "difficulties": {
+    "easy": 2,
+    "expert": 38,
+    "hard": 31,
+    "medium": 29
+  },
+  "exact_duplicate_prompts": 0,
+  "reference_tool_calls": {
+    "max": 34,
+    "median": 13,
+    "min": 4,
+    "total": 1318
+  },
+  "schema_version": "1.0",
+  "task_count": 100,
+  "verifier": {
+    "deterministic": true,
+    "model_calls": 0,
+    "network_calls": 0,
+    "random_calls": 0,
+    "wall_clock_reads": 0
+  },
+  "version": "1.0.0",
+  "world_commits": 417,
+  "world_documents": 30,
+  "world_seeded_rows": 1451,
+  "world_tables": 72,
+  "world_tools": 97
+}

tasks/dob100-001-rca-analytics-egress-blocked.json

@@ -0,0 +1,112 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "service_localized",
+        "fault_type_correct",
+        "offending_key_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "evidence": "the path from analytics-worker to pg-replica is REFUSED at the transport layer rather than timing out, so this is a network policy applied on day 418 and not a capacity problem; a timeout would look like load and a refusal cannot",
+          "fault_detected": true,
+          "fault_type": "misconfig",
+          "offending_key": "pg-replica",
+          "scope": "analytics-connection-refused",
+          "service": "analytics-worker"
+        },
+        "tool": "submit_diagnosis"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-163 — Root cause: analytics connections are refused, not slow\n\nGet to the bottom of analytics-connection-refused: not just which service, but the mechanism and the exact setting responsible. The service's own code is not necessarily where this lives — confirm it against the state of the system rather than inferring it from the symptom alone. Submit your finding under scope 'analytics-connection-refused' with the responsible service, the fault type, the offending key, and your evidence.\n\nProcedure: read the production logs and the error tracker, inspect the deployed configuration, and then look one layer down - at the cluster the service runs on, its nodes, its pods and their events - because a fault there leaves no trace in the service's own code or metrics. Confirm the mechanism against that state before submitting.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "aiops_analysis",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "aiops_analysis",
+      "expert",
+      "9-call-oracle",
+      "curated",
+      "hardest-of-category-quota",
+      "controls-rejected:pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "expert",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-001-rca-analytics-egress-blocked",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 9,
+    "required_tools": [
+      "check_network_path",
+      "get_service",
+      "get_ticket",
+      "list_error_events",
+      "query_metrics",
+      "search_logs",
+      "submit_diagnosis",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_rca_analytics_egress_blocked",
+    "version": "1.0.0"
+  },
+  "prompt": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-163 — Root cause: analytics connections are refused, not slow\n\nGet to the bottom of analytics-connection-refused: not just which service, but the mechanism and the exact setting responsible. The service's own code is not necessarily where this lives — confirm it against the state of the system rather than inferring it from the symptom alone. Submit your finding under scope 'analytics-connection-refused' with the responsible service, the fault type, the offending key, and your evidence.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "service_localized",
+        "fault_type_correct",
+        "offending_key_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-001-rca-analytics-egress-blocked.py"
+  },
+  "task_id": "dob100-001-rca-analytics-egress-blocked",
+  "task_name": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-002-rca-analytics-revoked-grant.json

@@ -0,0 +1,113 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "service_localized",
+        "fault_type_correct",
+        "offending_key_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "evidence": "the analytics_ro role on pg-replica was dropped during the day-418 credential rotation, but the running analytics-worker pods were never restarted and still present it; the database is healthy and so is the service, and the grant between them is not",
+          "fault_detected": true,
+          "fault_type": "misconfig",
+          "offending_key": "analytics_ro",
+          "scope": "analytics-cannot-read-replica",
+          "service": "analytics-worker"
+        },
+        "tool": "submit_diagnosis"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-160 — Root cause: analytics cannot read the replica since the rotation\n\nGet to the bottom of analytics-cannot-read-replica: not just which service, but the mechanism and the exact setting responsible. The service's own code is not necessarily where this lives — confirm it against the state of the system rather than inferring it from the symptom alone. Submit your finding under scope 'analytics-cannot-read-replica' with the responsible service, the fault type, the offending key, and your evidence.\n\nProcedure: read the production logs and the error tracker, inspect the deployed configuration, and then look one layer down - at the cluster the service runs on, its nodes, its pods and their events - because a fault there leaves no trace in the service's own code or metrics. Confirm the mechanism against that state before submitting.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "aiops_analysis",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "aiops_analysis",
+      "expert",
+      "9-call-oracle",
+      "curated",
+      "hardest-of-category-quota",
+      "controls-rejected:pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "expert",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-002-rca-analytics-revoked-grant",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 9,
+    "required_tools": [
+      "get_service",
+      "get_ticket",
+      "list_db_grants",
+      "list_error_events",
+      "list_infra",
+      "query_metrics",
+      "search_logs",
+      "submit_diagnosis",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_rca_analytics_revoked_grant",
+    "version": "1.0.0"
+  },
+  "prompt": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-160 — Root cause: analytics cannot read the replica since the rotation\n\nGet to the bottom of analytics-cannot-read-replica: not just which service, but the mechanism and the exact setting responsible. The service's own code is not necessarily where this lives — confirm it against the state of the system rather than inferring it from the symptom alone. Submit your finding under scope 'analytics-cannot-read-replica' with the responsible service, the fault type, the offending key, and your evidence.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "service_localized",
+        "fault_type_correct",
+        "offending_key_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-002-rca-analytics-revoked-grant.py"
+  },
+  "task_id": "dob100-002-rca-analytics-revoked-grant",
+  "task_name": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-003-rca-checkout-unschedulable-replicas.json

@@ -0,0 +1,114 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "service_localized",
+        "fault_type_correct",
+        "offending_key_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "evidence": "the checkout deployment declares 64 desired replicas and has 6 ready; the 58 that cannot be admitted are Pending with '4 Insufficient cpu' and no node in the cluster has the capacity, so the shortfall is in the requested replica count rather than in the cluster or the service",
+          "fault_detected": true,
+          "fault_type": "misconfig",
+          "offending_key": "desired_replicas",
+          "scope": "checkout-capacity-shortfall",
+          "service": "checkout"
+        },
+        "tool": "submit_diagnosis"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-129 — Root cause: checkout is running at a fraction of capacity\n\nGet to the bottom of checkout-capacity-shortfall: not just which service, but the mechanism and the exact setting responsible. The service's own code is not necessarily where this lives — confirm it against the state of the system rather than inferring it from the symptom alone. Submit your finding under scope 'checkout-capacity-shortfall' with the responsible service, the fault type, the offending key, and your evidence.\n\nProcedure: read the production logs and the error tracker, inspect the deployed configuration, and then look one layer down - at the cluster the service runs on, its nodes, its pods and their events - because a fault there leaves no trace in the service's own code or metrics. Confirm the mechanism against that state before submitting.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "aiops_analysis",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "aiops_analysis",
+      "expert",
+      "10-call-oracle",
+      "curated",
+      "hardest-of-category-quota",
+      "controls-rejected:pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "expert",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-003-rca-checkout-unschedulable-replicas",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 10,
+    "required_tools": [
+      "get_service",
+      "get_ticket",
+      "k8s_deployments_list",
+      "k8s_nodes_list",
+      "k8s_pods_list",
+      "list_error_events",
+      "query_metrics",
+      "search_logs",
+      "submit_diagnosis",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_rca_checkout_unschedulable_replicas",
+    "version": "1.0.0"
+  },
+  "prompt": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-129 — Root cause: checkout is running at a fraction of capacity\n\nGet to the bottom of checkout-capacity-shortfall: not just which service, but the mechanism and the exact setting responsible. The service's own code is not necessarily where this lives — confirm it against the state of the system rather than inferring it from the symptom alone. Submit your finding under scope 'checkout-capacity-shortfall' with the responsible service, the fault type, the offending key, and your evidence.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "service_localized",
+        "fault_type_correct",
+        "offending_key_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-003-rca-checkout-unschedulable-replicas.py"
+  },
+  "task_id": "dob100-003-rca-checkout-unschedulable-replicas",
+  "task_name": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-004-rca-media-disk-pressure.json

@@ -0,0 +1,114 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "service_localized",
+        "fault_type_correct",
+        "offending_key_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "evidence": "both media-service replicas are scheduled on node-b3, whose kubelet reports DiskPressure at 97% of 200Gi ephemeral storage, and one replica has already been Evicted for ephemeral-storage three times; the pods themselves are Running with normal memory and the service's own code is unchanged",
+          "fault_detected": true,
+          "fault_type": "node_unhealthy",
+          "offending_key": "node-b3",
+          "scope": "media-upload-stalls",
+          "service": "media-service"
+        },
+        "tool": "submit_diagnosis"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-125 — Root cause: media uploads stalling\n\nGet to the bottom of media-upload-stalls: not just which service, but the mechanism and the exact setting responsible. The service's own code is not necessarily where this lives — confirm it against the state of the system rather than inferring it from the symptom alone. Submit your finding under scope 'media-upload-stalls' with the responsible service, the fault type, the offending key, and your evidence.\n\nProcedure: read the production logs and the error tracker, inspect the deployed configuration, and then look one layer down - at the cluster the service runs on, its nodes, its pods and their events - because a fault there leaves no trace in the service's own code or metrics. Confirm the mechanism against that state before submitting.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "aiops_analysis",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "aiops_analysis",
+      "expert",
+      "10-call-oracle",
+      "curated",
+      "hardest-of-category-quota",
+      "controls-rejected:pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "expert",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-004-rca-media-disk-pressure",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 10,
+    "required_tools": [
+      "get_service",
+      "get_ticket",
+      "k8s_events_list",
+      "k8s_nodes_list",
+      "k8s_pods_list",
+      "list_error_events",
+      "query_metrics",
+      "search_logs",
+      "submit_diagnosis",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_rca_media_disk_pressure",
+    "version": "1.0.0"
+  },
+  "prompt": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-125 — Root cause: media uploads stalling\n\nGet to the bottom of media-upload-stalls: not just which service, but the mechanism and the exact setting responsible. The service's own code is not necessarily where this lives — confirm it against the state of the system rather than inferring it from the symptom alone. Submit your finding under scope 'media-upload-stalls' with the responsible service, the fault type, the offending key, and your evidence.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "service_localized",
+        "fault_type_correct",
+        "offending_key_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-004-rca-media-disk-pressure.py"
+  },
+  "task_id": "dob100-004-rca-media-disk-pressure",
+  "task_name": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-005-rca-notifications-node-deadlock.json

@@ -0,0 +1,114 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "service_localized",
+        "fault_type_correct",
+        "offending_key_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "evidence": "node-c1 is Ready=Unknown with a kernel soft lockup on CPU#3 and stopped posting node status 14 minutes ago; notifications-1b7d-hh88 still reports Running because the kubelet stopped reporting, not the pod, so pod phase alone says the service is healthy while it is unreachable",
+          "fault_detected": true,
+          "fault_type": "node_unhealthy",
+          "offending_key": "node-c1",
+          "scope": "notifications-unreachable",
+          "service": "notifications"
+        },
+        "tool": "submit_diagnosis"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-127 — Root cause: notifications are not being delivered\n\nGet to the bottom of notifications-unreachable: not just which service, but the mechanism and the exact setting responsible. The service's own code is not necessarily where this lives — confirm it against the state of the system rather than inferring it from the symptom alone. Submit your finding under scope 'notifications-unreachable' with the responsible service, the fault type, the offending key, and your evidence.\n\nProcedure: read the production logs and the error tracker, inspect the deployed configuration, and then look one layer down - at the cluster the service runs on, its nodes, its pods and their events - because a fault there leaves no trace in the service's own code or metrics. Confirm the mechanism against that state before submitting.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "aiops_analysis",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "aiops_analysis",
+      "expert",
+      "10-call-oracle",
+      "curated",
+      "hardest-of-category-quota",
+      "controls-rejected:pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "expert",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-005-rca-notifications-node-deadlock",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 10,
+    "required_tools": [
+      "get_service",
+      "get_ticket",
+      "k8s_events_list",
+      "k8s_nodes_list",
+      "k8s_pods_list",
+      "list_error_events",
+      "query_metrics",
+      "search_logs",
+      "submit_diagnosis",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_rca_notifications_node_deadlock",
+    "version": "1.0.0"
+  },
+  "prompt": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-127 — Root cause: notifications are not being delivered\n\nGet to the bottom of notifications-unreachable: not just which service, but the mechanism and the exact setting responsible. The service's own code is not necessarily where this lives — confirm it against the state of the system rather than inferring it from the symptom alone. Submit your finding under scope 'notifications-unreachable' with the responsible service, the fault type, the offending key, and your evidence.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "service_localized",
+        "fault_type_correct",
+        "offending_key_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-005-rca-notifications-node-deadlock.py"
+  },
+  "task_id": "dob100-005-rca-notifications-node-deadlock",
+  "task_name": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-006-rca-search-unscheduled-reindex.json

@@ -0,0 +1,114 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "service_localized",
+        "fault_type_correct",
+        "offending_key_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "evidence": "search-reindex-8c2a-gg77 has been Pending since day 419 with 214 FailedScheduling events reading 0/4 nodes available: no node matches the pod's accelerator=gpu-a100 selector, and no node in the cluster carries that label; the reindex has therefore never run once and the index is stale",
+          "fault_detected": true,
+          "fault_type": "misconfig",
+          "offending_key": "accelerator=gpu-a100",
+          "scope": "search-stale-index",
+          "service": "search"
+        },
+        "tool": "submit_diagnosis"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-126 — Root cause: search results are stale\n\nGet to the bottom of search-stale-index: not just which service, but the mechanism and the exact setting responsible. The service's own code is not necessarily where this lives — confirm it against the state of the system rather than inferring it from the symptom alone. Submit your finding under scope 'search-stale-index' with the responsible service, the fault type, the offending key, and your evidence.\n\nProcedure: read the production logs and the error tracker, inspect the deployed configuration, and then look one layer down - at the cluster the service runs on, its nodes, its pods and their events - because a fault there leaves no trace in the service's own code or metrics. Confirm the mechanism against that state before submitting.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "aiops_analysis",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "aiops_analysis",
+      "expert",
+      "10-call-oracle",
+      "curated",
+      "hardest-of-category-quota",
+      "controls-rejected:pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "expert",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-006-rca-search-unscheduled-reindex",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 10,
+    "required_tools": [
+      "get_service",
+      "get_ticket",
+      "k8s_events_list",
+      "k8s_nodes_list",
+      "k8s_pods_list",
+      "list_error_events",
+      "query_metrics",
+      "search_logs",
+      "submit_diagnosis",
+      "update_ticket"
+    ],
+    "source_split": "heldout",
+    "source_task_id": "tsk_rca_search_unscheduled_reindex",
+    "version": "1.0.0"
+  },
+  "prompt": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-126 — Root cause: search results are stale\n\nGet to the bottom of search-stale-index: not just which service, but the mechanism and the exact setting responsible. The service's own code is not necessarily where this lives — confirm it against the state of the system rather than inferring it from the symptom alone. Submit your finding under scope 'search-stale-index' with the responsible service, the fault type, the offending key, and your evidence.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "service_localized",
+        "fault_type_correct",
+        "offending_key_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-006-rca-search-unscheduled-reindex.py"
+  },
+  "task_id": "dob100-006-rca-search-unscheduled-reindex",
+  "task_name": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-007-detect-checkout-errors.json

@@ -0,0 +1,106 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "detection_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "evidence": "checkout error_rate_pct is 5.5 against a 1.0 SLO while latency_p99_ms 180 is within its 400 SLO",
+          "fault_detected": true,
+          "fault_type": "feature_flag_regression",
+          "offending_key": "instant_refunds",
+          "scope": "checkout",
+          "service": "checkout"
+        },
+        "tool": "submit_diagnosis"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-104 — Detection: is checkout violating its SLOs?\n\nDetermine whether checkout is currently violating any of its service level objectives, and submit the result under scope 'checkout' with the evidence you relied on. Report honestly either way: a clean bill of health is a valid finding.\n\nProcedure: compare live production metrics against the service's SLOs and check the alarm state, then submit. If healthy, submit fault_detected=false with fault_type='none'.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "aiops_detection",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "aiops_detection",
+      "medium",
+      "6-call-oracle",
+      "curated",
+      "hardest-of-category-quota",
+      "controls-rejected:pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "medium",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-007-detect-checkout-errors",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 6,
+    "required_tools": [
+      "get_slo_status",
+      "get_ticket",
+      "list_alerts",
+      "query_metrics",
+      "submit_diagnosis",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_detect_checkout_errors",
+    "version": "1.0.0"
+  },
+  "prompt": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-104 — Detection: is checkout violating its SLOs?\n\nDetermine whether checkout is currently violating any of its service level objectives, and submit the result under scope 'checkout' with the evidence you relied on. Report honestly either way: a clean bill of health is a valid finding.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "detection_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-007-detect-checkout-errors.py"
+  },
+  "task_id": "dob100-007-detect-checkout-errors",
+  "task_name": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-008-detect-inventory.json

@@ -0,0 +1,106 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "detection_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "evidence": "inventory error_rate_pct is 4.7 against a 1.0 SLO; alarm 9606 is firing",
+          "fault_detected": true,
+          "fault_type": "resource_exhaustion",
+          "offending_key": "db_pool_size",
+          "scope": "inventory",
+          "service": "inventory"
+        },
+        "tool": "submit_diagnosis"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-102 — Detection: is inventory violating its SLOs?\n\nDetermine whether inventory is currently violating any of its service level objectives, and submit the result under scope 'inventory' with the evidence you relied on. Report honestly either way: a clean bill of health is a valid finding.\n\nProcedure: compare live production metrics against the service's SLOs and check the alarm state, then submit. If healthy, submit fault_detected=false with fault_type='none'.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "aiops_detection",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "aiops_detection",
+      "easy",
+      "6-call-oracle",
+      "curated",
+      "hardest-of-category-quota",
+      "controls-rejected:pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "easy",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-008-detect-inventory",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 6,
+    "required_tools": [
+      "get_slo_status",
+      "get_ticket",
+      "list_alerts",
+      "query_metrics",
+      "submit_diagnosis",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_detect_inventory",
+    "version": "1.0.0"
+  },
+  "prompt": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-102 — Detection: is inventory violating its SLOs?\n\nDetermine whether inventory is currently violating any of its service level objectives, and submit the result under scope 'inventory' with the evidence you relied on. Report honestly either way: a clean bill of health is a valid finding.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "detection_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-008-detect-inventory.py"
+  },
+  "task_id": "dob100-008-detect-inventory",
+  "task_name": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-009-detect-payments.json

@@ -0,0 +1,106 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "detection_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "evidence": "payments error_rate_pct is 4.2 against a 1.0 SLO; alarm 9601 is firing",
+          "fault_detected": true,
+          "fault_type": "missing_retry",
+          "offending_key": "notifications_retry_max_attempts",
+          "scope": "payments",
+          "service": "payments"
+        },
+        "tool": "submit_diagnosis"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-101 — Detection: is payments violating its SLOs?\n\nDetermine whether payments is currently violating any of its service level objectives, and submit the result under scope 'payments' with the evidence you relied on. Report honestly either way: a clean bill of health is a valid finding.\n\nProcedure: compare live production metrics against the service's SLOs and check the alarm state, then submit. If healthy, submit fault_detected=false with fault_type='none'.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "aiops_detection",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "aiops_detection",
+      "easy",
+      "6-call-oracle",
+      "curated",
+      "hardest-of-category-quota",
+      "controls-rejected:pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "easy",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-009-detect-payments",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 6,
+    "required_tools": [
+      "get_slo_status",
+      "get_ticket",
+      "list_alerts",
+      "query_metrics",
+      "submit_diagnosis",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_detect_payments",
+    "version": "1.0.0"
+  },
+  "prompt": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-101 — Detection: is payments violating its SLOs?\n\nDetermine whether payments is currently violating any of its service level objectives, and submit the result under scope 'payments' with the evidence you relied on. Report honestly either way: a clean bill of health is a valid finding.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "detection_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-009-detect-payments.py"
+  },
+  "task_id": "dob100-009-detect-payments",
+  "task_name": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-010-detect-status-page-recurrence.json

@@ -0,0 +1,109 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "detection_correct",
+        "service_identified",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "evidence": "the status page post 7001 is marked resolved, but checkout is breaching both of its SLOs right now - error_rate_pct 5.5 against a 1.0 threshold and latency_p99_ms 530 against 400 - and alarms 9603 and 9610 are both firing; the status page lags internal state and is not evidence of health",
+          "fault_detected": true,
+          "fault_type": "unclassified",
+          "offending_key": "",
+          "scope": "status-7001-recurrence",
+          "service": "checkout"
+        },
+        "tool": "submit_diagnosis"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-131 — Customers still cannot check out, and the status page says we fixed it\n\nSupport has had eleven tickets this morning from customers who cannot complete a purchase - they get an error at the final step and give up. Our status page has nothing open: the only related post, 'Degraded checkout performance', was marked resolved on day 412. Support wants to know whether to keep telling customers this is their browser.\n\nWork out which service this is about, determine whether it is currently violating any of its service level objectives, and submit the result under scope 'status-7001-recurrence' — naming the service in the `service` field — with the evidence you relied on. Report honestly either way: a clean bill of health is a valid finding, and so is contradicting a system that says everything is fine.\n\nProcedure: compare live production metrics against the service's SLOs and check the alarm state, then submit. If healthy, submit fault_detected=false with fault_type='none'.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "aiops_detection",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "aiops_detection",
+      "expert",
+      "8-call-oracle",
+      "curated",
+      "hardest-of-category-quota",
+      "controls-rejected:pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "expert",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-010-detect-status-page-recurrence",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 8,
+    "required_tools": [
+      "get_slo_status",
+      "get_ticket",
+      "list_alerts",
+      "list_status_page_posts",
+      "query_metrics",
+      "submit_diagnosis",
+      "update_ticket"
+    ],
+    "source_split": "heldout",
+    "source_task_id": "tsk_detect_status_page_recurrence",
+    "version": "1.0.0"
+  },
+  "prompt": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-131 — Customers still cannot check out, and the status page says we fixed it\n\nSupport has had eleven tickets this morning from customers who cannot complete a purchase - they get an error at the final step and give up. Our status page has nothing open: the only related post, 'Degraded checkout performance', was marked resolved on day 412. Support wants to know whether to keep telling customers this is their browser.\n\nWork out which service this is about, determine whether it is currently violating any of its service level objectives, and submit the result under scope 'status-7001-recurrence' — naming the service in the `service` field — with the evidence you relied on. Report honestly either way: a clean bill of health is a valid finding, and so is contradicting a system that says everything is fine.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "detection_correct",
+        "service_identified",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-010-detect-status-page-recurrence.py"
+  },
+  "task_id": "dob100-010-detect-status-page-recurrence",
+  "task_name": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-011-detect-storefront-healthy.json

@@ -0,0 +1,106 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "detection_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "evidence": "storefront-web latency_p99_ms is 220 against a 500 SLO and no alarm is firing",
+          "fault_detected": false,
+          "fault_type": "none",
+          "offending_key": "",
+          "scope": "storefront-web",
+          "service": ""
+        },
+        "tool": "submit_diagnosis"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-103 — Detection: is storefront-web violating its SLOs?\n\nDetermine whether storefront-web is currently violating any of its service level objectives, and submit the result under scope 'storefront-web' with the evidence you relied on. Report honestly either way: a clean bill of health is a valid finding.\n\nProcedure: compare live production metrics against the service's SLOs and check the alarm state, then submit. If healthy, submit fault_detected=false with fault_type='none'.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "aiops_detection",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "aiops_detection",
+      "medium",
+      "6-call-oracle",
+      "curated",
+      "hardest-of-category-quota",
+      "controls-rejected:pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "medium",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-011-detect-storefront-healthy",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 6,
+    "required_tools": [
+      "get_slo_status",
+      "get_ticket",
+      "list_alerts",
+      "query_metrics",
+      "submit_diagnosis",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_detect_storefront_healthy",
+    "version": "1.0.0"
+  },
+  "prompt": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-103 — Detection: is storefront-web violating its SLOs?\n\nDetermine whether storefront-web is currently violating any of its service level objectives, and submit the result under scope 'storefront-web' with the evidence you relied on. Report honestly either way: a clean bill of health is a valid finding.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "detection_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-011-detect-storefront-healthy.py"
+  },
+  "task_id": "dob100-011-detect-storefront-healthy",
+  "task_name": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-012-localize-analytics-crashloop.json

@@ -0,0 +1,109 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "service_localized",
+        "fault_confirmed",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "evidence": "the kubelet records 47 OOMKilled events and a CrashLoopBackOff on analytics-worker-7d9f-x2k1 at a 512Mi limit against 511Mi usage; the error tracker shows nothing for this service because the process is killed before its SDK can flush",
+          "fault_detected": true,
+          "fault_type": "resource_exhaustion",
+          "offending_key": "memory_limit_mb",
+          "scope": "9609",
+          "service": "analytics-worker"
+        },
+        "tool": "submit_diagnosis"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-116 — Localize alarm 9609 — the error tracker shows nothing\n\nAlarm 9609 is firing. Work out which service is actually responsible — the service the alarm names is not automatically the culprit — and submit your finding under scope '9609' with the evidence you relied on.\n\nProcedure: trace the symptom through the metrics, the service dependency graph and the logs, then submit the responsible service.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "aiops_localization",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "aiops_localization",
+      "expert",
+      "7-call-oracle",
+      "curated",
+      "hardest-of-category-quota",
+      "controls-rejected:pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "expert",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-012-localize-analytics-crashloop",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 7,
+    "required_tools": [
+      "get_service",
+      "get_slo_status",
+      "get_ticket",
+      "list_alerts",
+      "search_logs",
+      "submit_diagnosis",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_localize_analytics_crashloop",
+    "version": "1.0.0"
+  },
+  "prompt": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-116 — Localize alarm 9609 — the error tracker shows nothing\n\nAlarm 9609 is firing. Work out which service is actually responsible — the service the alarm names is not automatically the culprit — and submit your finding under scope '9609' with the evidence you relied on.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "service_localized",
+        "fault_confirmed",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-012-localize-analytics-crashloop.py"
+  },
+  "task_id": "dob100-012-localize-analytics-crashloop",
+  "task_name": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-013-localize-analytics-errors.json

@@ -0,0 +1,109 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "service_localized",
+        "fault_confirmed",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "evidence": "analytics-worker error_rate_pct 6.0; the consumer OOMs with unlimited queue prefetch",
+          "fault_detected": true,
+          "fault_type": "unbounded_prefetch",
+          "offending_key": "prefetch_count",
+          "scope": "9609",
+          "service": "analytics-worker"
+        },
+        "tool": "submit_diagnosis"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-113 — Localize alarm 9609 (analytics-worker errors)\n\nAlarm 9609 is firing. Work out which service is actually responsible — the service the alarm names is not automatically the culprit — and submit your finding under scope '9609' with the evidence you relied on.\n\nProcedure: trace the symptom through the metrics, the service dependency graph and the logs, then submit the responsible service.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "aiops_localization",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "aiops_localization",
+      "medium",
+      "7-call-oracle",
+      "curated",
+      "hardest-of-category-quota",
+      "controls-rejected:pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "medium",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-013-localize-analytics-errors",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 7,
+    "required_tools": [
+      "get_service",
+      "get_slo_status",
+      "get_ticket",
+      "list_alerts",
+      "search_logs",
+      "submit_diagnosis",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_localize_analytics_errors",
+    "version": "1.0.0"
+  },
+  "prompt": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-113 — Localize alarm 9609 (analytics-worker errors)\n\nAlarm 9609 is firing. Work out which service is actually responsible — the service the alarm names is not automatically the culprit — and submit your finding under scope '9609' with the evidence you relied on.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "service_localized",
+        "fault_confirmed",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-013-localize-analytics-errors.py"
+  },
+  "task_id": "dob100-013-localize-analytics-errors",
+  "task_name": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-014-localize-checkout-latency.json

@@ -0,0 +1,109 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "service_localized",
+        "fault_confirmed",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "evidence": "checkout p99 530ms is spent waiting on payments, which blocks for up to 30s on its notifications call; checkout's own configuration is within policy",
+          "fault_detected": true,
+          "fault_type": "misconfig",
+          "offending_key": "notifications_timeout_ms",
+          "scope": "9610",
+          "service": "payments"
+        },
+        "tool": "submit_diagnosis"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-115 — Localize alarm 9610 (checkout latency)\n\nAlarm 9610 is firing. Work out which service is actually responsible — the service the alarm names is not automatically the culprit — and submit your finding under scope '9610' with the evidence you relied on.\n\nProcedure: trace the symptom through the metrics, the service dependency graph and the logs, then submit the responsible service.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "aiops_localization",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "aiops_localization",
+      "hard",
+      "7-call-oracle",
+      "curated",
+      "hardest-of-category-quota",
+      "controls-rejected:shortcut,pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "hard",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-014-localize-checkout-latency",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 7,
+    "required_tools": [
+      "get_service",
+      "get_slo_status",
+      "get_ticket",
+      "list_alerts",
+      "search_logs",
+      "submit_diagnosis",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_localize_checkout_latency",
+    "version": "1.0.0"
+  },
+  "prompt": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-115 — Localize alarm 9610 (checkout latency)\n\nAlarm 9610 is firing. Work out which service is actually responsible — the service the alarm names is not automatically the culprit — and submit your finding under scope '9610' with the evidence you relied on.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "service_localized",
+        "fault_confirmed",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-014-localize-checkout-latency.py"
+  },
+  "task_id": "dob100-014-localize-checkout-latency",
+  "task_name": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-015-localize-gateway-latency.json

@@ -0,0 +1,109 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "service_localized",
+        "fault_confirmed",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "evidence": "p99 1030ms began immediately after v5.1.0 was promoted; the pool opens a connection per request",
+          "fault_detected": true,
+          "fault_type": "bad_release",
+          "offending_key": "v5.1.0",
+          "scope": "9604",
+          "service": "api-gateway"
+        },
+        "tool": "submit_diagnosis"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-111 — Localize alarm 9604 (api-gateway latency)\n\nAlarm 9604 is firing. Work out which service is actually responsible — the service the alarm names is not automatically the culprit — and submit your finding under scope '9604' with the evidence you relied on.\n\nProcedure: trace the symptom through the metrics, the service dependency graph and the logs, then submit the responsible service.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "aiops_localization",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "aiops_localization",
+      "medium",
+      "7-call-oracle",
+      "curated",
+      "hardest-of-category-quota",
+      "controls-rejected:pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "medium",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-015-localize-gateway-latency",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 7,
+    "required_tools": [
+      "get_service",
+      "get_slo_status",
+      "get_ticket",
+      "list_alerts",
+      "search_logs",
+      "submit_diagnosis",
+      "update_ticket"
+    ],
+    "source_split": "heldout",
+    "source_task_id": "tsk_localize_gateway_latency",
+    "version": "1.0.0"
+  },
+  "prompt": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-111 — Localize alarm 9604 (api-gateway latency)\n\nAlarm 9604 is firing. Work out which service is actually responsible — the service the alarm names is not automatically the culprit — and submit your finding under scope '9604' with the evidence you relied on.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "service_localized",
+        "fault_confirmed",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-015-localize-gateway-latency.py"
+  },
+  "task_id": "dob100-015-localize-gateway-latency",
+  "task_name": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-016-localize-media-latency.json

@@ -0,0 +1,109 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "service_localized",
+        "fault_confirmed",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "evidence": "media-service p99 800ms with every asset request served from the origin object store",
+          "fault_detected": true,
+          "fault_type": "cdn_bypass",
+          "offending_key": "cdn_enabled",
+          "scope": "9607",
+          "service": "media-service"
+        },
+        "tool": "submit_diagnosis"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-114 — Localize alarm 9607 (media-service latency)\n\nAlarm 9607 is firing. Work out which service is actually responsible — the service the alarm names is not automatically the culprit — and submit your finding under scope '9607' with the evidence you relied on.\n\nProcedure: trace the symptom through the metrics, the service dependency graph and the logs, then submit the responsible service.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "aiops_localization",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "aiops_localization",
+      "medium",
+      "7-call-oracle",
+      "curated",
+      "hardest-of-category-quota",
+      "controls-rejected:pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "medium",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-016-localize-media-latency",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 7,
+    "required_tools": [
+      "get_service",
+      "get_slo_status",
+      "get_ticket",
+      "list_alerts",
+      "search_logs",
+      "submit_diagnosis",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_localize_media_latency",
+    "version": "1.0.0"
+  },
+  "prompt": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-114 — Localize alarm 9607 (media-service latency)\n\nAlarm 9607 is firing. Work out which service is actually responsible — the service the alarm names is not automatically the culprit — and submit your finding under scope '9607' with the evidence you relied on.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "diagnosis_submitted",
+        "service_localized",
+        "fault_confirmed",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "efficient_investigation",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-016-localize-media-latency.py"
+  },
+  "task_id": "dob100-016-localize-media-latency",
+  "task_name": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-017-auth-v1-to-v2.json

@@ -0,0 +1,124 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "legacy_retired",
+        "replacement_active",
+        "legacy_traffic_drained",
+        "replacement_serving_all_traffic",
+        "ci_all_stages_green",
+        "consumer_migrated",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "deprecate_before_shift",
+        "staged_traffic_shifts",
+        "drain_before_retire",
+        "consumer_before_retire",
+        "staging_first",
+        "canary_then_promote",
+        "staging_first",
+        "canary_then_promote",
+        "no_alarming_deploys"
+      ],
+      "quality": [
+        "pr_linked_to_ticket",
+        "ticket_closed",
+        "closed_after_the_work",
+        "pr_has_description",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "ENG-2402 — Migrate /v1/auth traffic to /v2/auth and retire v1\n\nThe /v1/auth API on api-gateway is superseded by /v2/auth. Retire the legacy path and move its production traffic across, without dropping requests. NovaCart's engineering policies are documented in the knowledge base and are not optional; CI enforces some of them and will tell you what it is unhappy about.\n\nDone when: /v2/auth serves 100% of production traffic, /v1/auth is retired in production, and ENG-2402 is closed.\n\nProcedure: migrate any consumer service to the new version first and deploy it; deprecate the legacy endpoint and deploy that; shift traffic in steps of at most 50 percentage points until the replacement serves 100%; only then retire the legacy endpoint in a second pull request and deploy it. Every deploy is staging-first, tier-1 canary then promote.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "api_migration",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "api_migration",
+      "expert",
+      "34-call-oracle",
+      "curated",
+      "category-shipped-whole",
+      "controls-rejected:naive,pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "expert",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-017-auth-v1-to-v2",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 34,
+    "required_tools": [
+      "assess_canary",
+      "deploy_service",
+      "get_document",
+      "get_ticket",
+      "list_api_endpoints",
+      "merge_pull_request",
+      "open_pull_request",
+      "promote_canary",
+      "run_ci",
+      "search_docs",
+      "shift_endpoint_traffic",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_auth_v1_to_v2",
+    "version": "1.0.0"
+  },
+  "prompt": "ENG-2402 — Migrate /v1/auth traffic to /v2/auth and retire v1\n\nThe /v1/auth API on api-gateway is superseded by /v2/auth. Retire the legacy path and move its production traffic across, without dropping requests. NovaCart's engineering policies are documented in the knowledge base and are not optional; CI enforces some of them and will tell you what it is unhappy about.\n\nDone when: /v2/auth serves 100% of production traffic, /v1/auth is retired in production, and ENG-2402 is closed.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "legacy_retired",
+        "replacement_active",
+        "legacy_traffic_drained",
+        "replacement_serving_all_traffic",
+        "ci_all_stages_green",
+        "consumer_migrated",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "deprecate_before_shift",
+        "staged_traffic_shifts",
+        "drain_before_retire",
+        "consumer_before_retire",
+        "staging_first",
+        "canary_then_promote",
+        "staging_first",
+        "canary_then_promote",
+        "no_alarming_deploys"
+      ],
+      "quality": [
+        "pr_linked_to_ticket",
+        "ticket_closed",
+        "closed_after_the_work",
+        "pr_has_description",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-017-auth-v1-to-v2.py"
+  },
+  "task_id": "dob100-017-auth-v1-to-v2",
+  "task_name": "ENG-2402 — Migrate /v1/auth traffic to /v2/auth and retire v1",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-018-checkout-v1-to-v2.json

@@ -0,0 +1,124 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "legacy_retired",
+        "replacement_active",
+        "legacy_traffic_drained",
+        "replacement_serving_all_traffic",
+        "ci_all_stages_green",
+        "consumer_migrated",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "deprecate_before_shift",
+        "staged_traffic_shifts",
+        "drain_before_retire",
+        "consumer_before_retire",
+        "staging_first",
+        "canary_then_promote",
+        "staging_first",
+        "canary_then_promote",
+        "no_alarming_deploys"
+      ],
+      "quality": [
+        "pr_linked_to_ticket",
+        "ticket_closed",
+        "closed_after_the_work",
+        "pr_has_description",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "ENG-2403 — Migrate /v1/checkout traffic to /v2/checkout and retire v1\n\nThe /v1/checkout API on api-gateway is superseded by /v2/checkout. Retire the legacy path and move its production traffic across, without dropping requests. NovaCart's engineering policies are documented in the knowledge base and are not optional; CI enforces some of them and will tell you what it is unhappy about.\n\nDone when: /v2/checkout serves 100% of production traffic, /v1/checkout is retired in production, and ENG-2403 is closed.\n\nProcedure: migrate any consumer service to the new version first and deploy it; deprecate the legacy endpoint and deploy that; shift traffic in steps of at most 50 percentage points until the replacement serves 100%; only then retire the legacy endpoint in a second pull request and deploy it. Every deploy is staging-first, tier-1 canary then promote.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "api_migration",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "api_migration",
+      "expert",
+      "34-call-oracle",
+      "curated",
+      "category-shipped-whole",
+      "controls-rejected:naive,pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "expert",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-018-checkout-v1-to-v2",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 34,
+    "required_tools": [
+      "assess_canary",
+      "deploy_service",
+      "get_document",
+      "get_ticket",
+      "list_api_endpoints",
+      "merge_pull_request",
+      "open_pull_request",
+      "promote_canary",
+      "run_ci",
+      "search_docs",
+      "shift_endpoint_traffic",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_checkout_v1_to_v2",
+    "version": "1.0.0"
+  },
+  "prompt": "ENG-2403 — Migrate /v1/checkout traffic to /v2/checkout and retire v1\n\nThe /v1/checkout API on api-gateway is superseded by /v2/checkout. Retire the legacy path and move its production traffic across, without dropping requests. NovaCart's engineering policies are documented in the knowledge base and are not optional; CI enforces some of them and will tell you what it is unhappy about.\n\nDone when: /v2/checkout serves 100% of production traffic, /v1/checkout is retired in production, and ENG-2403 is closed.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "legacy_retired",
+        "replacement_active",
+        "legacy_traffic_drained",
+        "replacement_serving_all_traffic",
+        "ci_all_stages_green",
+        "consumer_migrated",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "deprecate_before_shift",
+        "staged_traffic_shifts",
+        "drain_before_retire",
+        "consumer_before_retire",
+        "staging_first",
+        "canary_then_promote",
+        "staging_first",
+        "canary_then_promote",
+        "no_alarming_deploys"
+      ],
+      "quality": [
+        "pr_linked_to_ticket",
+        "ticket_closed",
+        "closed_after_the_work",
+        "pr_has_description",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-018-checkout-v1-to-v2.py"
+  },
+  "task_id": "dob100-018-checkout-v1-to-v2",
+  "task_name": "ENG-2403 — Migrate /v1/checkout traffic to /v2/checkout and retire v1",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-019-inventory-v1-to-v2.json

@@ -0,0 +1,116 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "legacy_retired",
+        "replacement_active",
+        "legacy_traffic_drained",
+        "replacement_serving_all_traffic",
+        "ci_all_stages_green",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "deprecate_before_shift",
+        "staged_traffic_shifts",
+        "drain_before_retire",
+        "staging_first",
+        "canary_then_promote",
+        "no_alarming_deploys"
+      ],
+      "quality": [
+        "pr_linked_to_ticket",
+        "ticket_closed",
+        "closed_after_the_work",
+        "pr_has_description",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "ENG-2406 — Migrate /v1/inventory traffic to /v2/inventory and retire v1\n\nThe /v1/inventory API on api-gateway is superseded by /v2/inventory. Retire the legacy path and move its production traffic across, without dropping requests. NovaCart's engineering policies are documented in the knowledge base and are not optional; CI enforces some of them and will tell you what it is unhappy about.\n\nDone when: /v2/inventory serves 100% of production traffic, /v1/inventory is retired in production, and ENG-2406 is closed.\n\nProcedure: migrate any consumer service to the new version first and deploy it; deprecate the legacy endpoint and deploy that; shift traffic in steps of at most 50 percentage points until the replacement serves 100%; only then retire the legacy endpoint in a second pull request and deploy it. Every deploy is staging-first, tier-1 canary then promote.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "api_migration",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "api_migration",
+      "hard",
+      "27-call-oracle",
+      "curated",
+      "category-shipped-whole",
+      "controls-rejected:naive,pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "hard",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-019-inventory-v1-to-v2",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 27,
+    "required_tools": [
+      "assess_canary",
+      "deploy_service",
+      "get_document",
+      "get_ticket",
+      "list_api_endpoints",
+      "merge_pull_request",
+      "open_pull_request",
+      "promote_canary",
+      "run_ci",
+      "search_docs",
+      "shift_endpoint_traffic",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_inventory_v1_to_v2",
+    "version": "1.0.0"
+  },
+  "prompt": "ENG-2406 — Migrate /v1/inventory traffic to /v2/inventory and retire v1\n\nThe /v1/inventory API on api-gateway is superseded by /v2/inventory. Retire the legacy path and move its production traffic across, without dropping requests. NovaCart's engineering policies are documented in the knowledge base and are not optional; CI enforces some of them and will tell you what it is unhappy about.\n\nDone when: /v2/inventory serves 100% of production traffic, /v1/inventory is retired in production, and ENG-2406 is closed.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "legacy_retired",
+        "replacement_active",
+        "legacy_traffic_drained",
+        "replacement_serving_all_traffic",
+        "ci_all_stages_green",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "deprecate_before_shift",
+        "staged_traffic_shifts",
+        "drain_before_retire",
+        "staging_first",
+        "canary_then_promote",
+        "no_alarming_deploys"
+      ],
+      "quality": [
+        "pr_linked_to_ticket",
+        "ticket_closed",
+        "closed_after_the_work",
+        "pr_has_description",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-019-inventory-v1-to-v2.py"
+  },
+  "task_id": "dob100-019-inventory-v1-to-v2",
+  "task_name": "ENG-2406 — Migrate /v1/inventory traffic to /v2/inventory and retire v1",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-020-media-v1-to-v2.json

@@ -0,0 +1,116 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "legacy_retired",
+        "replacement_active",
+        "legacy_traffic_drained",
+        "replacement_serving_all_traffic",
+        "ci_all_stages_green",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "deprecate_before_shift",
+        "staged_traffic_shifts",
+        "drain_before_retire",
+        "staging_first",
+        "canary_then_promote",
+        "no_alarming_deploys"
+      ],
+      "quality": [
+        "pr_linked_to_ticket",
+        "ticket_closed",
+        "closed_after_the_work",
+        "pr_has_description",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "ENG-2405 — Migrate /v1/media traffic to /v2/media and retire v1\n\nThe /v1/media API on api-gateway is superseded by /v2/media. Retire the legacy path and move its production traffic across, without dropping requests. NovaCart's engineering policies are documented in the knowledge base and are not optional; CI enforces some of them and will tell you what it is unhappy about.\n\nDone when: /v2/media serves 100% of production traffic, /v1/media is retired in production, and ENG-2405 is closed.\n\nProcedure: migrate any consumer service to the new version first and deploy it; deprecate the legacy endpoint and deploy that; shift traffic in steps of at most 50 percentage points until the replacement serves 100%; only then retire the legacy endpoint in a second pull request and deploy it. Every deploy is staging-first, tier-1 canary then promote.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "api_migration",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "api_migration",
+      "hard",
+      "27-call-oracle",
+      "curated",
+      "category-shipped-whole",
+      "controls-rejected:naive,pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "hard",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-020-media-v1-to-v2",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 27,
+    "required_tools": [
+      "assess_canary",
+      "deploy_service",
+      "get_document",
+      "get_ticket",
+      "list_api_endpoints",
+      "merge_pull_request",
+      "open_pull_request",
+      "promote_canary",
+      "run_ci",
+      "search_docs",
+      "shift_endpoint_traffic",
+      "update_ticket"
+    ],
+    "source_split": "heldout",
+    "source_task_id": "tsk_media_v1_to_v2",
+    "version": "1.0.0"
+  },
+  "prompt": "ENG-2405 — Migrate /v1/media traffic to /v2/media and retire v1\n\nThe /v1/media API on api-gateway is superseded by /v2/media. Retire the legacy path and move its production traffic across, without dropping requests. NovaCart's engineering policies are documented in the knowledge base and are not optional; CI enforces some of them and will tell you what it is unhappy about.\n\nDone when: /v2/media serves 100% of production traffic, /v1/media is retired in production, and ENG-2405 is closed.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "legacy_retired",
+        "replacement_active",
+        "legacy_traffic_drained",
+        "replacement_serving_all_traffic",
+        "ci_all_stages_green",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "deprecate_before_shift",
+        "staged_traffic_shifts",
+        "drain_before_retire",
+        "staging_first",
+        "canary_then_promote",
+        "no_alarming_deploys"
+      ],
+      "quality": [
+        "pr_linked_to_ticket",
+        "ticket_closed",
+        "closed_after_the_work",
+        "pr_has_description",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-020-media-v1-to-v2.py"
+  },
+  "task_id": "dob100-020-media-v1-to-v2",
+  "task_name": "ENG-2405 — Migrate /v1/media traffic to /v2/media and retire v1",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-021-notify-v1-to-v2.json

@@ -0,0 +1,116 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "legacy_retired",
+        "replacement_active",
+        "legacy_traffic_drained",
+        "replacement_serving_all_traffic",
+        "ci_all_stages_green",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "deprecate_before_shift",
+        "staged_traffic_shifts",
+        "drain_before_retire",
+        "staging_first",
+        "canary_then_promote",
+        "no_alarming_deploys"
+      ],
+      "quality": [
+        "pr_linked_to_ticket",
+        "ticket_closed",
+        "closed_after_the_work",
+        "pr_has_description",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "ENG-2407 — Migrate /v1/notify traffic to /v2/notify and retire v1\n\nThe /v1/notify API on api-gateway is superseded by /v2/notify. Retire the legacy path and move its production traffic across, without dropping requests. NovaCart's engineering policies are documented in the knowledge base and are not optional; CI enforces some of them and will tell you what it is unhappy about.\n\nDone when: /v2/notify serves 100% of production traffic, /v1/notify is retired in production, and ENG-2407 is closed.\n\nProcedure: migrate any consumer service to the new version first and deploy it; deprecate the legacy endpoint and deploy that; shift traffic in steps of at most 50 percentage points until the replacement serves 100%; only then retire the legacy endpoint in a second pull request and deploy it. Every deploy is staging-first, tier-1 canary then promote.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "api_migration",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "api_migration",
+      "hard",
+      "27-call-oracle",
+      "curated",
+      "category-shipped-whole",
+      "controls-rejected:naive,pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "hard",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-021-notify-v1-to-v2",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 27,
+    "required_tools": [
+      "assess_canary",
+      "deploy_service",
+      "get_document",
+      "get_ticket",
+      "list_api_endpoints",
+      "merge_pull_request",
+      "open_pull_request",
+      "promote_canary",
+      "run_ci",
+      "search_docs",
+      "shift_endpoint_traffic",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_notify_v1_to_v2",
+    "version": "1.0.0"
+  },
+  "prompt": "ENG-2407 — Migrate /v1/notify traffic to /v2/notify and retire v1\n\nThe /v1/notify API on api-gateway is superseded by /v2/notify. Retire the legacy path and move its production traffic across, without dropping requests. NovaCart's engineering policies are documented in the knowledge base and are not optional; CI enforces some of them and will tell you what it is unhappy about.\n\nDone when: /v2/notify serves 100% of production traffic, /v1/notify is retired in production, and ENG-2407 is closed.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "legacy_retired",
+        "replacement_active",
+        "legacy_traffic_drained",
+        "replacement_serving_all_traffic",
+        "ci_all_stages_green",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "deprecate_before_shift",
+        "staged_traffic_shifts",
+        "drain_before_retire",
+        "staging_first",
+        "canary_then_promote",
+        "no_alarming_deploys"
+      ],
+      "quality": [
+        "pr_linked_to_ticket",
+        "ticket_closed",
+        "closed_after_the_work",
+        "pr_has_description",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-021-notify-v1-to-v2.py"
+  },
+  "task_id": "dob100-021-notify-v1-to-v2",
+  "task_name": "ENG-2407 — Migrate /v1/notify traffic to /v2/notify and retire v1",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-022-orders-v1-to-v2.json

@@ -0,0 +1,124 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "legacy_retired",
+        "replacement_active",
+        "legacy_traffic_drained",
+        "replacement_serving_all_traffic",
+        "ci_all_stages_green",
+        "consumer_migrated",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "deprecate_before_shift",
+        "staged_traffic_shifts",
+        "drain_before_retire",
+        "consumer_before_retire",
+        "staging_first",
+        "canary_then_promote",
+        "staging_first",
+        "canary_then_promote",
+        "no_alarming_deploys"
+      ],
+      "quality": [
+        "pr_linked_to_ticket",
+        "ticket_closed",
+        "closed_after_the_work",
+        "pr_has_description",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "ENG-2401 — Migrate /v1/orders traffic to /v2/orders and retire v1\n\nThe /v1/orders API on api-gateway is superseded by /v2/orders. Retire the legacy path and move its production traffic across, without dropping requests. NovaCart's engineering policies are documented in the knowledge base and are not optional; CI enforces some of them and will tell you what it is unhappy about.\n\nDone when: /v2/orders serves 100% of production traffic, /v1/orders is retired in production, and ENG-2401 is closed.\n\nProcedure: migrate any consumer service to the new version first and deploy it; deprecate the legacy endpoint and deploy that; shift traffic in steps of at most 50 percentage points until the replacement serves 100%; only then retire the legacy endpoint in a second pull request and deploy it. Every deploy is staging-first, tier-1 canary then promote.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "api_migration",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "api_migration",
+      "expert",
+      "34-call-oracle",
+      "curated",
+      "category-shipped-whole",
+      "controls-rejected:naive,pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "expert",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-022-orders-v1-to-v2",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 34,
+    "required_tools": [
+      "assess_canary",
+      "deploy_service",
+      "get_document",
+      "get_ticket",
+      "list_api_endpoints",
+      "merge_pull_request",
+      "open_pull_request",
+      "promote_canary",
+      "run_ci",
+      "search_docs",
+      "shift_endpoint_traffic",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_orders_v1_to_v2",
+    "version": "1.0.0"
+  },
+  "prompt": "ENG-2401 — Migrate /v1/orders traffic to /v2/orders and retire v1\n\nThe /v1/orders API on api-gateway is superseded by /v2/orders. Retire the legacy path and move its production traffic across, without dropping requests. NovaCart's engineering policies are documented in the knowledge base and are not optional; CI enforces some of them and will tell you what it is unhappy about.\n\nDone when: /v2/orders serves 100% of production traffic, /v1/orders is retired in production, and ENG-2401 is closed.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "legacy_retired",
+        "replacement_active",
+        "legacy_traffic_drained",
+        "replacement_serving_all_traffic",
+        "ci_all_stages_green",
+        "consumer_migrated",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "deprecate_before_shift",
+        "staged_traffic_shifts",
+        "drain_before_retire",
+        "consumer_before_retire",
+        "staging_first",
+        "canary_then_promote",
+        "staging_first",
+        "canary_then_promote",
+        "no_alarming_deploys"
+      ],
+      "quality": [
+        "pr_linked_to_ticket",
+        "ticket_closed",
+        "closed_after_the_work",
+        "pr_has_description",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-022-orders-v1-to-v2.py"
+  },
+  "task_id": "dob100-022-orders-v1-to-v2",
+  "task_name": "ENG-2401 — Migrate /v1/orders traffic to /v2/orders and retire v1",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-023-search-v1-to-v2.json

@@ -0,0 +1,116 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "legacy_retired",
+        "replacement_active",
+        "legacy_traffic_drained",
+        "replacement_serving_all_traffic",
+        "ci_all_stages_green",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "deprecate_before_shift",
+        "staged_traffic_shifts",
+        "drain_before_retire",
+        "staging_first",
+        "canary_then_promote",
+        "no_alarming_deploys"
+      ],
+      "quality": [
+        "pr_linked_to_ticket",
+        "ticket_closed",
+        "closed_after_the_work",
+        "pr_has_description",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "ENG-2404 — Migrate /v1/search traffic to /v2/search and retire v1\n\nThe /v1/search API on api-gateway is superseded by /v2/search. Retire the legacy path and move its production traffic across, without dropping requests. NovaCart's engineering policies are documented in the knowledge base and are not optional; CI enforces some of them and will tell you what it is unhappy about.\n\nDone when: /v2/search serves 100% of production traffic, /v1/search is retired in production, and ENG-2404 is closed.\n\nProcedure: migrate any consumer service to the new version first and deploy it; deprecate the legacy endpoint and deploy that; shift traffic in steps of at most 50 percentage points until the replacement serves 100%; only then retire the legacy endpoint in a second pull request and deploy it. Every deploy is staging-first, tier-1 canary then promote.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "api_migration",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "api_migration",
+      "hard",
+      "27-call-oracle",
+      "curated",
+      "category-shipped-whole",
+      "controls-rejected:naive,pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "hard",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-023-search-v1-to-v2",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 27,
+    "required_tools": [
+      "assess_canary",
+      "deploy_service",
+      "get_document",
+      "get_ticket",
+      "list_api_endpoints",
+      "merge_pull_request",
+      "open_pull_request",
+      "promote_canary",
+      "run_ci",
+      "search_docs",
+      "shift_endpoint_traffic",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_search_v1_to_v2",
+    "version": "1.0.0"
+  },
+  "prompt": "ENG-2404 — Migrate /v1/search traffic to /v2/search and retire v1\n\nThe /v1/search API on api-gateway is superseded by /v2/search. Retire the legacy path and move its production traffic across, without dropping requests. NovaCart's engineering policies are documented in the knowledge base and are not optional; CI enforces some of them and will tell you what it is unhappy about.\n\nDone when: /v2/search serves 100% of production traffic, /v1/search is retired in production, and ENG-2404 is closed.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "legacy_retired",
+        "replacement_active",
+        "legacy_traffic_drained",
+        "replacement_serving_all_traffic",
+        "ci_all_stages_green",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "deprecate_before_shift",
+        "staged_traffic_shifts",
+        "drain_before_retire",
+        "staging_first",
+        "canary_then_promote",
+        "no_alarming_deploys"
+      ],
+      "quality": [
+        "pr_linked_to_ticket",
+        "ticket_closed",
+        "closed_after_the_work",
+        "pr_has_description",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-023-search-v1-to-v2.py"
+  },
+  "task_id": "dob100-023-search-v1-to-v2",
+  "task_name": "ENG-2404 — Migrate /v1/search traffic to /v2/search and retire v1",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-024-attr-three-at-once.json

@@ -0,0 +1,141 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "attributed_media_upload_stalls",
+        "key_for_media_upload_stalls",
+        "attributed_gateway_latency_surge",
+        "key_for_gateway_latency_surge",
+        "attributed_checkout_error_spike",
+        "key_for_checkout_error_spike",
+        "causes_are_distinct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "evidence": "both media-service replicas are scheduled on node-b3, which reports DiskPressure at 97% of its ephemeral storage, and one replica has already been Evicted for ephemeral-storage",
+          "fault_detected": true,
+          "fault_type": "node_unhealthy",
+          "offending_key": "node-b3",
+          "scope": "media-upload-stalls",
+          "service": "media-service"
+        },
+        "tool": "submit_diagnosis"
+      },
+      {
+        "arguments": {
+          "evidence": "p99 moved when v5.1.0 was promoted; that release opens an upstream connection per request without releasing it, so the leak is in every version at or above v5.1.0 and only a rollback recovers it",
+          "fault_detected": true,
+          "fault_type": "bad_release",
+          "offending_key": "v5.1.0",
+          "scope": "gateway-latency-surge",
+          "service": "api-gateway"
+        },
+        "tool": "submit_diagnosis"
+      },
+      {
+        "arguments": {
+          "evidence": "checkout errors track the instant_refunds rollout rather than any deploy, and disabling the flag returns the error rate to baseline",
+          "fault_detected": true,
+          "fault_type": "feature_flag_regression",
+          "offending_key": "instant_refunds",
+          "scope": "checkout-error-spike",
+          "service": "checkout"
+        },
+        "tool": "submit_diagnosis"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-150 — Three alarms in twenty minutes - is this one incident or three?\n\n3 things are wrong at once this morning:\n\n  - media uploads are stalling for customers\n  - api-gateway p99 is 1030ms against a 250ms SLO\n  - checkout error rate is 5.5% against a 1.0% SLO\n\nOps would like to treat this as a single incident with a single cause, because that is one rollback instead of three investigations. Work out whether that is true. Submit one diagnosis per scope listed above — the scope strings are 'media-upload-stalls', 'gateway-latency-surge', 'checkout-error-spike' — each naming the service, the fault type and the specific setting or version responsible.\n\nThey may share a cause and they may not. Deciding that is the task; do not assume either way, and do not attribute a symptom to a fault you happened to find while looking at something else.\n\nProcedure: take each symptom on its own. For each, follow the evidence from the affected service down through its pods, its node and its events before looking anywhere else, and only then decide whether any two share a cause.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "attribution",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "attribution",
+      "expert",
+      "13-call-oracle",
+      "curated",
+      "hardest-of-category-quota",
+      "controls-rejected:pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "expert",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-024-attr-three-at-once",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 13,
+    "required_tools": [
+      "get_ticket",
+      "k8s_events_list",
+      "k8s_nodes_list",
+      "k8s_pods_list",
+      "list_alerts",
+      "list_deployments",
+      "list_feature_flags",
+      "query_metrics",
+      "search_logs",
+      "submit_diagnosis",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_attr_three_at_once",
+    "version": "1.0.0"
+  },
+  "prompt": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nOPS-150 — Three alarms in twenty minutes - is this one incident or three?\n\n3 things are wrong at once this morning:\n\n  - media uploads are stalling for customers\n  - api-gateway p99 is 1030ms against a 250ms SLO\n  - checkout error rate is 5.5% against a 1.0% SLO\n\nOps would like to treat this as a single incident with a single cause, because that is one rollback instead of three investigations. Work out whether that is true. Submit one diagnosis per scope listed above — the scope strings are 'media-upload-stalls', 'gateway-latency-surge', 'checkout-error-spike' — each naming the service, the fault type and the specific setting or version responsible.\n\nThey may share a cause and they may not. Deciding that is the task; do not assume either way, and do not attribute a symptom to a fault you happened to find while looking at something else.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "attributed_media_upload_stalls",
+        "key_for_media_upload_stalls",
+        "attributed_gateway_latency_surge",
+        "key_for_gateway_latency_surge",
+        "attributed_checkout_error_spike",
+        "key_for_checkout_error_spike",
+        "causes_are_distinct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-024-attr-three-at-once.py"
+  },
+  "task_id": "dob100-024-attr-three-at-once",
+  "task_name": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-025-w5-attr-media-analytics-catalog.json

@@ -0,0 +1,138 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "attributed_media_upload_stalls",
+        "key_for_media_upload_stalls",
+        "attributed_analytics_connection_refused",
+        "key_for_analytics_connection_refused",
+        "attributed_catalog_latency_sawtooth",
+        "key_for_catalog_latency_sawtooth",
+        "causes_are_distinct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "evidence": "both replicas sit on node-b3, which reports DiskPressure at 97%",
+          "fault_detected": true,
+          "fault_type": "node_unhealthy",
+          "offending_key": "node-b3",
+          "scope": "media-upload-stalls",
+          "service": "media-service"
+        },
+        "tool": "submit_diagnosis"
+      },
+      {
+        "arguments": {
+          "evidence": "the path is REFUSED at the transport layer, so a network policy rather than capacity",
+          "fault_detected": true,
+          "fault_type": "misconfig",
+          "offending_key": "pg-replica",
+          "scope": "analytics-connection-refused",
+          "service": "analytics-worker"
+        },
+        "tool": "submit_diagnosis"
+      },
+      {
+        "arguments": {
+          "evidence": "94% heap, 780ms p99 GC pause, 41 collections a minute",
+          "fault_detected": true,
+          "fault_type": "resource_exhaustion",
+          "offending_key": "heap_used_pct",
+          "scope": "catalog-latency-sawtooth",
+          "service": "catalog"
+        },
+        "tool": "submit_diagnosis"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nMULTI-7607 — Three alarms at once - one incident or three?\n\n3 things are wrong at once this morning:\n\n  - media uploads are stalling\n  - analytics cannot reach the replica\n  - catalog latency arrives in waves\n\nOps would like to treat this as a single incident with a single cause, because that is one rollback instead of three investigations. Work out whether that is true. Submit one diagnosis per scope listed above — the scope strings are 'media-upload-stalls', 'analytics-connection-refused', 'catalog-latency-sawtooth' — each naming the service, the fault type and the specific setting or version responsible.\n\nThey may share a cause and they may not. Deciding that is the task; do not assume either way, and do not attribute a symptom to a fault you happened to find while looking at something else.\n\nProcedure: take each symptom on its own. For each, follow the evidence from the affected service down through its pods, its node and its events before looking anywhere else, and only then decide whether any two share a cause.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "attribution",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "attribution",
+      "expert",
+      "10-call-oracle",
+      "wave-generated",
+      "hardest-of-category-quota",
+      "controls-rejected:pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "expert",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-025-w5-attr-media-analytics-catalog",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 10,
+    "required_tools": [
+      "check_network_path",
+      "get_runtime_stats",
+      "get_ticket",
+      "k8s_nodes_list",
+      "k8s_pods_list",
+      "list_alerts",
+      "submit_diagnosis",
+      "update_ticket"
+    ],
+    "source_split": "heldout",
+    "source_task_id": "tsk_w5_attr_media_analytics_catalog",
+    "version": "1.0.0"
+  },
+  "prompt": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nMULTI-7607 — Three alarms at once - one incident or three?\n\n3 things are wrong at once this morning:\n\n  - media uploads are stalling\n  - analytics cannot reach the replica\n  - catalog latency arrives in waves\n\nOps would like to treat this as a single incident with a single cause, because that is one rollback instead of three investigations. Work out whether that is true. Submit one diagnosis per scope listed above �� the scope strings are 'media-upload-stalls', 'analytics-connection-refused', 'catalog-latency-sawtooth' — each naming the service, the fault type and the specific setting or version responsible.\n\nThey may share a cause and they may not. Deciding that is the task; do not assume either way, and do not attribute a symptom to a fault you happened to find while looking at something else.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "attributed_media_upload_stalls",
+        "key_for_media_upload_stalls",
+        "attributed_analytics_connection_refused",
+        "key_for_analytics_connection_refused",
+        "attributed_catalog_latency_sawtooth",
+        "key_for_catalog_latency_sawtooth",
+        "causes_are_distinct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-025-w5-attr-media-analytics-catalog.py"
+  },
+  "task_id": "dob100-025-w5-attr-media-analytics-catalog",
+  "task_name": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-026-w5-attr-media-analytics-payments.json

@@ -0,0 +1,138 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "attributed_media_upload_stalls",
+        "key_for_media_upload_stalls",
+        "attributed_analytics_connection_refused",
+        "key_for_analytics_connection_refused",
+        "attributed_payments_error_rate",
+        "key_for_payments_error_rate",
+        "causes_are_distinct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "evidence": "both replicas sit on node-b3, which reports DiskPressure at 97%",
+          "fault_detected": true,
+          "fault_type": "node_unhealthy",
+          "offending_key": "node-b3",
+          "scope": "media-upload-stalls",
+          "service": "media-service"
+        },
+        "tool": "submit_diagnosis"
+      },
+      {
+        "arguments": {
+          "evidence": "the path is REFUSED at the transport layer, so a network policy rather than capacity",
+          "fault_detected": true,
+          "fault_type": "misconfig",
+          "offending_key": "pg-replica",
+          "scope": "analytics-connection-refused",
+          "service": "analytics-worker"
+        },
+        "tool": "submit_diagnosis"
+      },
+      {
+        "arguments": {
+          "evidence": "the notifications call times out with no retry configured",
+          "fault_detected": true,
+          "fault_type": "missing_retry",
+          "offending_key": "notifications_retry_max_attempts",
+          "scope": "payments-error-rate",
+          "service": "payments"
+        },
+        "tool": "submit_diagnosis"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nMULTI-7608 — Three alarms at once - one incident or three?\n\n3 things are wrong at once this morning:\n\n  - media uploads are stalling\n  - analytics cannot reach the replica\n  - payments errors are 4.2% against 1.0%\n\nOps would like to treat this as a single incident with a single cause, because that is one rollback instead of three investigations. Work out whether that is true. Submit one diagnosis per scope listed above — the scope strings are 'media-upload-stalls', 'analytics-connection-refused', 'payments-error-rate' — each naming the service, the fault type and the specific setting or version responsible.\n\nThey may share a cause and they may not. Deciding that is the task; do not assume either way, and do not attribute a symptom to a fault you happened to find while looking at something else.\n\nProcedure: take each symptom on its own. For each, follow the evidence from the affected service down through its pods, its node and its events before looking anywhere else, and only then decide whether any two share a cause.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "attribution",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "attribution",
+      "expert",
+      "10-call-oracle",
+      "wave-generated",
+      "hardest-of-category-quota",
+      "controls-rejected:pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "expert",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-026-w5-attr-media-analytics-payments",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 10,
+    "required_tools": [
+      "check_network_path",
+      "get_ticket",
+      "k8s_nodes_list",
+      "k8s_pods_list",
+      "list_alerts",
+      "search_logs",
+      "submit_diagnosis",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_w5_attr_media_analytics_payments",
+    "version": "1.0.0"
+  },
+  "prompt": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nMULTI-7608 — Three alarms at once - one incident or three?\n\n3 things are wrong at once this morning:\n\n  - media uploads are stalling\n  - analytics cannot reach the replica\n  - payments errors are 4.2% against 1.0%\n\nOps would like to treat this as a single incident with a single cause, because that is one rollback instead of three investigations. Work out whether that is true. Submit one diagnosis per scope listed above — the scope strings are 'media-upload-stalls', 'analytics-connection-refused', 'payments-error-rate' — each naming the service, the fault type and the specific setting or version responsible.\n\nThey may share a cause and they may not. Deciding that is the task; do not assume either way, and do not attribute a symptom to a fault you happened to find while looking at something else.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "attributed_media_upload_stalls",
+        "key_for_media_upload_stalls",
+        "attributed_analytics_connection_refused",
+        "key_for_analytics_connection_refused",
+        "attributed_payments_error_rate",
+        "key_for_payments_error_rate",
+        "causes_are_distinct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-026-w5-attr-media-analytics-payments.py"
+  },
+  "task_id": "dob100-026-w5-attr-media-analytics-payments",
+  "task_name": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-027-w5-attr-media-api-analytics.json

@@ -0,0 +1,138 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "attributed_media_upload_stalls",
+        "key_for_media_upload_stalls",
+        "attributed_gateway_latency_surge",
+        "key_for_gateway_latency_surge",
+        "attributed_analytics_connection_refused",
+        "key_for_analytics_connection_refused",
+        "causes_are_distinct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "evidence": "both replicas sit on node-b3, which reports DiskPressure at 97%",
+          "fault_detected": true,
+          "fault_type": "node_unhealthy",
+          "offending_key": "node-b3",
+          "scope": "media-upload-stalls",
+          "service": "media-service"
+        },
+        "tool": "submit_diagnosis"
+      },
+      {
+        "arguments": {
+          "evidence": "p99 moved when v5.1.0 was promoted; it leaks an upstream connection per request, so only a rollback recovers it",
+          "fault_detected": true,
+          "fault_type": "bad_release",
+          "offending_key": "v5.1.0",
+          "scope": "gateway-latency-surge",
+          "service": "api-gateway"
+        },
+        "tool": "submit_diagnosis"
+      },
+      {
+        "arguments": {
+          "evidence": "the path is REFUSED at the transport layer, so a network policy rather than capacity",
+          "fault_detected": true,
+          "fault_type": "misconfig",
+          "offending_key": "pg-replica",
+          "scope": "analytics-connection-refused",
+          "service": "analytics-worker"
+        },
+        "tool": "submit_diagnosis"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nMULTI-7601 — Three alarms at once - one incident or three?\n\n3 things are wrong at once this morning:\n\n  - media uploads are stalling\n  - api-gateway p99 is 1030ms against 250ms\n  - analytics cannot reach the replica\n\nOps would like to treat this as a single incident with a single cause, because that is one rollback instead of three investigations. Work out whether that is true. Submit one diagnosis per scope listed above — the scope strings are 'media-upload-stalls', 'gateway-latency-surge', 'analytics-connection-refused' — each naming the service, the fault type and the specific setting or version responsible.\n\nThey may share a cause and they may not. Deciding that is the task; do not assume either way, and do not attribute a symptom to a fault you happened to find while looking at something else.\n\nProcedure: take each symptom on its own. For each, follow the evidence from the affected service down through its pods, its node and its events before looking anywhere else, and only then decide whether any two share a cause.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "attribution",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "attribution",
+      "expert",
+      "10-call-oracle",
+      "wave-generated",
+      "hardest-of-category-quota",
+      "controls-rejected:pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "expert",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-027-w5-attr-media-api-analytics",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 10,
+    "required_tools": [
+      "check_network_path",
+      "get_ticket",
+      "k8s_nodes_list",
+      "k8s_pods_list",
+      "list_alerts",
+      "list_deployments",
+      "submit_diagnosis",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_w5_attr_media_api_analytics",
+    "version": "1.0.0"
+  },
+  "prompt": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nMULTI-7601 — Three alarms at once - one incident or three?\n\n3 things are wrong at once this morning:\n\n  - media uploads are stalling\n  - api-gateway p99 is 1030ms against 250ms\n  - analytics cannot reach the replica\n\nOps would like to treat this as a single incident with a single cause, because that is one rollback instead of three investigations. Work out whether that is true. Submit one diagnosis per scope listed above — the scope strings are 'media-upload-stalls', 'gateway-latency-surge', 'analytics-connection-refused' — each naming the service, the fault type and the specific setting or version responsible.\n\nThey may share a cause and they may not. Deciding that is the task; do not assume either way, and do not attribute a symptom to a fault you happened to find while looking at something else.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "attributed_media_upload_stalls",
+        "key_for_media_upload_stalls",
+        "attributed_gateway_latency_surge",
+        "key_for_gateway_latency_surge",
+        "attributed_analytics_connection_refused",
+        "key_for_analytics_connection_refused",
+        "causes_are_distinct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-027-w5-attr-media-api-analytics.py"
+  },
+  "task_id": "dob100-027-w5-attr-media-api-analytics",
+  "task_name": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-028-w5-attr-media-api-catalog.json

@@ -0,0 +1,138 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "attributed_media_upload_stalls",
+        "key_for_media_upload_stalls",
+        "attributed_gateway_latency_surge",
+        "key_for_gateway_latency_surge",
+        "attributed_catalog_latency_sawtooth",
+        "key_for_catalog_latency_sawtooth",
+        "causes_are_distinct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "evidence": "both replicas sit on node-b3, which reports DiskPressure at 97%",
+          "fault_detected": true,
+          "fault_type": "node_unhealthy",
+          "offending_key": "node-b3",
+          "scope": "media-upload-stalls",
+          "service": "media-service"
+        },
+        "tool": "submit_diagnosis"
+      },
+      {
+        "arguments": {
+          "evidence": "p99 moved when v5.1.0 was promoted; it leaks an upstream connection per request, so only a rollback recovers it",
+          "fault_detected": true,
+          "fault_type": "bad_release",
+          "offending_key": "v5.1.0",
+          "scope": "gateway-latency-surge",
+          "service": "api-gateway"
+        },
+        "tool": "submit_diagnosis"
+      },
+      {
+        "arguments": {
+          "evidence": "94% heap, 780ms p99 GC pause, 41 collections a minute",
+          "fault_detected": true,
+          "fault_type": "resource_exhaustion",
+          "offending_key": "heap_used_pct",
+          "scope": "catalog-latency-sawtooth",
+          "service": "catalog"
+        },
+        "tool": "submit_diagnosis"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nMULTI-7602 — Three alarms at once - one incident or three?\n\n3 things are wrong at once this morning:\n\n  - media uploads are stalling\n  - api-gateway p99 is 1030ms against 250ms\n  - catalog latency arrives in waves\n\nOps would like to treat this as a single incident with a single cause, because that is one rollback instead of three investigations. Work out whether that is true. Submit one diagnosis per scope listed above — the scope strings are 'media-upload-stalls', 'gateway-latency-surge', 'catalog-latency-sawtooth' — each naming the service, the fault type and the specific setting or version responsible.\n\nThey may share a cause and they may not. Deciding that is the task; do not assume either way, and do not attribute a symptom to a fault you happened to find while looking at something else.\n\nProcedure: take each symptom on its own. For each, follow the evidence from the affected service down through its pods, its node and its events before looking anywhere else, and only then decide whether any two share a cause.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "attribution",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "attribution",
+      "expert",
+      "10-call-oracle",
+      "wave-generated",
+      "hardest-of-category-quota",
+      "controls-rejected:pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "expert",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-028-w5-attr-media-api-catalog",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 10,
+    "required_tools": [
+      "get_runtime_stats",
+      "get_ticket",
+      "k8s_nodes_list",
+      "k8s_pods_list",
+      "list_alerts",
+      "list_deployments",
+      "submit_diagnosis",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_w5_attr_media_api_catalog",
+    "version": "1.0.0"
+  },
+  "prompt": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nMULTI-7602 — Three alarms at once - one incident or three?\n\n3 things are wrong at once this morning:\n\n  - media uploads are stalling\n  - api-gateway p99 is 1030ms against 250ms\n  - catalog latency arrives in waves\n\nOps would like to treat this as a single incident with a single cause, because that is one rollback instead of three investigations. Work out whether that is true. Submit one diagnosis per scope listed above — the scope strings are 'media-upload-stalls', 'gateway-latency-surge', 'catalog-latency-sawtooth' — each naming the service, the fault type and the specific setting or version responsible.\n\nThey may share a cause and they may not. Deciding that is the task; do not assume either way, and do not attribute a symptom to a fault you happened to find while looking at something else.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "attributed_media_upload_stalls",
+        "key_for_media_upload_stalls",
+        "attributed_gateway_latency_surge",
+        "key_for_gateway_latency_surge",
+        "attributed_catalog_latency_sawtooth",
+        "key_for_catalog_latency_sawtooth",
+        "causes_are_distinct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-028-w5-attr-media-api-catalog.py"
+  },
+  "task_id": "dob100-028-w5-attr-media-api-catalog",
+  "task_name": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-029-w5-attr-media-api-checkout.json

@@ -0,0 +1,138 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "attributed_media_upload_stalls",
+        "key_for_media_upload_stalls",
+        "attributed_gateway_latency_surge",
+        "key_for_gateway_latency_surge",
+        "attributed_checkout_error_spike",
+        "key_for_checkout_error_spike",
+        "causes_are_distinct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "evidence": "both replicas sit on node-b3, which reports DiskPressure at 97%",
+          "fault_detected": true,
+          "fault_type": "node_unhealthy",
+          "offending_key": "node-b3",
+          "scope": "media-upload-stalls",
+          "service": "media-service"
+        },
+        "tool": "submit_diagnosis"
+      },
+      {
+        "arguments": {
+          "evidence": "p99 moved when v5.1.0 was promoted; it leaks an upstream connection per request, so only a rollback recovers it",
+          "fault_detected": true,
+          "fault_type": "bad_release",
+          "offending_key": "v5.1.0",
+          "scope": "gateway-latency-surge",
+          "service": "api-gateway"
+        },
+        "tool": "submit_diagnosis"
+      },
+      {
+        "arguments": {
+          "evidence": "checkout errors track the instant_refunds rollout rather than any deploy",
+          "fault_detected": true,
+          "fault_type": "feature_flag_regression",
+          "offending_key": "instant_refunds",
+          "scope": "checkout-error-spike",
+          "service": "checkout"
+        },
+        "tool": "submit_diagnosis"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nMULTI-7600 — Three alarms at once - one incident or three?\n\n3 things are wrong at once this morning:\n\n  - media uploads are stalling\n  - api-gateway p99 is 1030ms against 250ms\n  - checkout errors are 5.5% against 1.0%\n\nOps would like to treat this as a single incident with a single cause, because that is one rollback instead of three investigations. Work out whether that is true. Submit one diagnosis per scope listed above — the scope strings are 'media-upload-stalls', 'gateway-latency-surge', 'checkout-error-spike' — each naming the service, the fault type and the specific setting or version responsible.\n\nThey may share a cause and they may not. Deciding that is the task; do not assume either way, and do not attribute a symptom to a fault you happened to find while looking at something else.\n\nProcedure: take each symptom on its own. For each, follow the evidence from the affected service down through its pods, its node and its events before looking anywhere else, and only then decide whether any two share a cause.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "attribution",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "attribution",
+      "expert",
+      "10-call-oracle",
+      "wave-generated",
+      "hardest-of-category-quota",
+      "controls-rejected:pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "expert",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-029-w5-attr-media-api-checkout",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 10,
+    "required_tools": [
+      "get_ticket",
+      "k8s_nodes_list",
+      "k8s_pods_list",
+      "list_alerts",
+      "list_deployments",
+      "list_feature_flags",
+      "submit_diagnosis",
+      "update_ticket"
+    ],
+    "source_split": "heldout",
+    "source_task_id": "tsk_w5_attr_media_api_checkout",
+    "version": "1.0.0"
+  },
+  "prompt": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nMULTI-7600 — Three alarms at once - one incident or three?\n\n3 things are wrong at once this morning:\n\n  - media uploads are stalling\n  - api-gateway p99 is 1030ms against 250ms\n  - checkout errors are 5.5% against 1.0%\n\nOps would like to treat this as a single incident with a single cause, because that is one rollback instead of three investigations. Work out whether that is true. Submit one diagnosis per scope listed above — the scope strings are 'media-upload-stalls', 'gateway-latency-surge', 'checkout-error-spike' — each naming the service, the fault type and the specific setting or version responsible.\n\nThey may share a cause and they may not. Deciding that is the task; do not assume either way, and do not attribute a symptom to a fault you happened to find while looking at something else.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "attributed_media_upload_stalls",
+        "key_for_media_upload_stalls",
+        "attributed_gateway_latency_surge",
+        "key_for_gateway_latency_surge",
+        "attributed_checkout_error_spike",
+        "key_for_checkout_error_spike",
+        "causes_are_distinct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-029-w5-attr-media-api-checkout.py"
+  },
+  "task_id": "dob100-029-w5-attr-media-api-checkout",
+  "task_name": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-030-w5-attr-media-api-payments.json

@@ -0,0 +1,138 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "attributed_media_upload_stalls",
+        "key_for_media_upload_stalls",
+        "attributed_gateway_latency_surge",
+        "key_for_gateway_latency_surge",
+        "attributed_payments_error_rate",
+        "key_for_payments_error_rate",
+        "causes_are_distinct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "evidence": "both replicas sit on node-b3, which reports DiskPressure at 97%",
+          "fault_detected": true,
+          "fault_type": "node_unhealthy",
+          "offending_key": "node-b3",
+          "scope": "media-upload-stalls",
+          "service": "media-service"
+        },
+        "tool": "submit_diagnosis"
+      },
+      {
+        "arguments": {
+          "evidence": "p99 moved when v5.1.0 was promoted; it leaks an upstream connection per request, so only a rollback recovers it",
+          "fault_detected": true,
+          "fault_type": "bad_release",
+          "offending_key": "v5.1.0",
+          "scope": "gateway-latency-surge",
+          "service": "api-gateway"
+        },
+        "tool": "submit_diagnosis"
+      },
+      {
+        "arguments": {
+          "evidence": "the notifications call times out with no retry configured",
+          "fault_detected": true,
+          "fault_type": "missing_retry",
+          "offending_key": "notifications_retry_max_attempts",
+          "scope": "payments-error-rate",
+          "service": "payments"
+        },
+        "tool": "submit_diagnosis"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nMULTI-7603 — Three alarms at once - one incident or three?\n\n3 things are wrong at once this morning:\n\n  - media uploads are stalling\n  - api-gateway p99 is 1030ms against 250ms\n  - payments errors are 4.2% against 1.0%\n\nOps would like to treat this as a single incident with a single cause, because that is one rollback instead of three investigations. Work out whether that is true. Submit one diagnosis per scope listed above — the scope strings are 'media-upload-stalls', 'gateway-latency-surge', 'payments-error-rate' — each naming the service, the fault type and the specific setting or version responsible.\n\nThey may share a cause and they may not. Deciding that is the task; do not assume either way, and do not attribute a symptom to a fault you happened to find while looking at something else.\n\nProcedure: take each symptom on its own. For each, follow the evidence from the affected service down through its pods, its node and its events before looking anywhere else, and only then decide whether any two share a cause.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "attribution",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "attribution",
+      "expert",
+      "10-call-oracle",
+      "wave-generated",
+      "hardest-of-category-quota",
+      "controls-rejected:pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "expert",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-030-w5-attr-media-api-payments",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 10,
+    "required_tools": [
+      "get_ticket",
+      "k8s_nodes_list",
+      "k8s_pods_list",
+      "list_alerts",
+      "list_deployments",
+      "search_logs",
+      "submit_diagnosis",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_w5_attr_media_api_payments",
+    "version": "1.0.0"
+  },
+  "prompt": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).\n\nMULTI-7603 — Three alarms at once - one incident or three?\n\n3 things are wrong at once this morning:\n\n  - media uploads are stalling\n  - api-gateway p99 is 1030ms against 250ms\n  - payments errors are 4.2% against 1.0%\n\nOps would like to treat this as a single incident with a single cause, because that is one rollback instead of three investigations. Work out whether that is true. Submit one diagnosis per scope listed above — the scope strings are 'media-upload-stalls', 'gateway-latency-surge', 'payments-error-rate' — each naming the service, the fault type and the specific setting or version responsible.\n\nThey may share a cause and they may not. Deciding that is the task; do not assume either way, and do not attribute a symptom to a fault you happened to find while looking at something else.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "attributed_media_upload_stalls",
+        "key_for_media_upload_stalls",
+        "attributed_gateway_latency_surge",
+        "key_for_gateway_latency_surge",
+        "attributed_payments_error_rate",
+        "key_for_payments_error_rate",
+        "causes_are_distinct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "investigation_was_read_only"
+      ],
+      "quality": [
+        "evidence_recorded",
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-030-w5-attr-media-api-payments.py"
+  },
+  "task_id": "dob100-030-w5-attr-media-api-payments",
+  "task_name": "This is an investigation, not a change. Do NOT modify production state: no pull requests, deploys, flag toggles, migrations or resolutions. Investigate with the read tools and report what you find with submit_diagnosis(...).",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-031-impl-backoff.json

@@ -0,0 +1,91 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "implementation_written",
+        "visible_tests_pass",
+        "hidden_tests_pass",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "executed_before_finishing"
+      ],
+      "quality": [
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "OPS-140 — Implement the payments retry backoff schedule\n\nsrc/payments/backoff.py is not implemented. Read the specification with read_exercise('src/payments/backoff.py'), write your implementation with write_implementation, and run it with run_exercise_tests until the visible tests pass.\n\nThe visible tests are not the whole specification. Other tests run at grading time and are not shown to you, so satisfying what you can see is necessary and not sufficient — read the specification and implement what it actually says.\n\nProcedure: read the specification, write the implementation, run the visible tests, and re-read the specification for the cases the visible tests do not cover before you finish.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "code_implementation",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "code_implementation",
+      "hard",
+      "5-call-oracle",
+      "curated",
+      "category-shipped-whole",
+      "controls-rejected:naive,pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "hard",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-031-impl-backoff",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 5,
+    "required_tools": [
+      "get_ticket",
+      "read_exercise",
+      "run_exercise_tests",
+      "update_ticket",
+      "write_implementation"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_impl_backoff",
+    "version": "1.0.0"
+  },
+  "prompt": "OPS-140 — Implement the payments retry backoff schedule\n\nsrc/payments/backoff.py is not implemented. Read the specification with read_exercise('src/payments/backoff.py'), write your implementation with write_implementation, and run it with run_exercise_tests until the visible tests pass.\n\nThe visible tests are not the whole specification. Other tests run at grading time and are not shown to you, so satisfying what you can see is necessary and not sufficient — read the specification and implement what it actually says.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "implementation_written",
+        "visible_tests_pass",
+        "hidden_tests_pass",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "executed_before_finishing"
+      ],
+      "quality": [
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-031-impl-backoff.py"
+  },
+  "task_id": "dob100-031-impl-backoff",
+  "task_name": "OPS-140 — Implement the payments retry backoff schedule",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-032-impl-cachekey.json

@@ -0,0 +1,91 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "implementation_written",
+        "visible_tests_pass",
+        "hidden_tests_pass",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "executed_before_finishing"
+      ],
+      "quality": [
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "OPS-142 — Implement the search cache key\n\nsrc/search/cache_key.py is not implemented. Read the specification with read_exercise('src/search/cache_key.py'), write your implementation with write_implementation, and run it with run_exercise_tests until the visible tests pass.\n\nThe visible tests are not the whole specification. Other tests run at grading time and are not shown to you, so satisfying what you can see is necessary and not sufficient — read the specification and implement what it actually says.\n\nProcedure: read the specification, write the implementation, run the visible tests, and re-read the specification for the cases the visible tests do not cover before you finish.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "code_implementation",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "code_implementation",
+      "hard",
+      "5-call-oracle",
+      "curated",
+      "category-shipped-whole",
+      "controls-rejected:naive,pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "hard",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-032-impl-cachekey",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 5,
+    "required_tools": [
+      "get_ticket",
+      "read_exercise",
+      "run_exercise_tests",
+      "update_ticket",
+      "write_implementation"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_impl_cachekey",
+    "version": "1.0.0"
+  },
+  "prompt": "OPS-142 — Implement the search cache key\n\nsrc/search/cache_key.py is not implemented. Read the specification with read_exercise('src/search/cache_key.py'), write your implementation with write_implementation, and run it with run_exercise_tests until the visible tests pass.\n\nThe visible tests are not the whole specification. Other tests run at grading time and are not shown to you, so satisfying what you can see is necessary and not sufficient — read the specification and implement what it actually says.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "implementation_written",
+        "visible_tests_pass",
+        "hidden_tests_pass",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "executed_before_finishing"
+      ],
+      "quality": [
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-032-impl-cachekey.py"
+  },
+  "task_id": "dob100-032-impl-cachekey",
+  "task_name": "OPS-142 — Implement the search cache key",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-033-impl-chunk.json

@@ -0,0 +1,91 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "implementation_written",
+        "visible_tests_pass",
+        "hidden_tests_pass",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "executed_before_finishing"
+      ],
+      "quality": [
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "OPS-141 — Implement settlement batch chunking\n\nsrc/payments/chunking.py is not implemented. Read the specification with read_exercise('src/payments/chunking.py'), write your implementation with write_implementation, and run it with run_exercise_tests until the visible tests pass.\n\nThe visible tests are not the whole specification. Other tests run at grading time and are not shown to you, so satisfying what you can see is necessary and not sufficient — read the specification and implement what it actually says.\n\nProcedure: read the specification, write the implementation, run the visible tests, and re-read the specification for the cases the visible tests do not cover before you finish.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "code_implementation",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "code_implementation",
+      "medium",
+      "5-call-oracle",
+      "curated",
+      "category-shipped-whole",
+      "controls-rejected:naive,pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "medium",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-033-impl-chunk",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 5,
+    "required_tools": [
+      "get_ticket",
+      "read_exercise",
+      "run_exercise_tests",
+      "update_ticket",
+      "write_implementation"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_impl_chunk",
+    "version": "1.0.0"
+  },
+  "prompt": "OPS-141 — Implement settlement batch chunking\n\nsrc/payments/chunking.py is not implemented. Read the specification with read_exercise('src/payments/chunking.py'), write your implementation with write_implementation, and run it with run_exercise_tests until the visible tests pass.\n\nThe visible tests are not the whole specification. Other tests run at grading time and are not shown to you, so satisfying what you can see is necessary and not sufficient — read the specification and implement what it actually says.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "implementation_written",
+        "visible_tests_pass",
+        "hidden_tests_pass",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "executed_before_finishing"
+      ],
+      "quality": [
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-033-impl-chunk.py"
+  },
+  "task_id": "dob100-033-impl-chunk",
+  "task_name": "OPS-141 — Implement settlement batch chunking",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-034-impl-ratelimit.json

@@ -0,0 +1,91 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "implementation_written",
+        "visible_tests_pass",
+        "hidden_tests_pass",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "executed_before_finishing"
+      ],
+      "quality": [
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "OPS-143 — Implement per-client rate limiting at the edge\n\nsrc/gateway/token_bucket.py is not implemented. Read the specification with read_exercise('src/gateway/token_bucket.py'), write your implementation with write_implementation, and run it with run_exercise_tests until the visible tests pass.\n\nThe visible tests are not the whole specification. Other tests run at grading time and are not shown to you, so satisfying what you can see is necessary and not sufficient — read the specification and implement what it actually says.\n\nProcedure: read the specification, write the implementation, run the visible tests, and re-read the specification for the cases the visible tests do not cover before you finish.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "code_implementation",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "code_implementation",
+      "expert",
+      "5-call-oracle",
+      "curated",
+      "category-shipped-whole",
+      "controls-rejected:naive,pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "expert",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-034-impl-ratelimit",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 5,
+    "required_tools": [
+      "get_ticket",
+      "read_exercise",
+      "run_exercise_tests",
+      "update_ticket",
+      "write_implementation"
+    ],
+    "source_split": "heldout",
+    "source_task_id": "tsk_impl_ratelimit",
+    "version": "1.0.0"
+  },
+  "prompt": "OPS-143 — Implement per-client rate limiting at the edge\n\nsrc/gateway/token_bucket.py is not implemented. Read the specification with read_exercise('src/gateway/token_bucket.py'), write your implementation with write_implementation, and run it with run_exercise_tests until the visible tests pass.\n\nThe visible tests are not the whole specification. Other tests run at grading time and are not shown to you, so satisfying what you can see is necessary and not sufficient — read the specification and implement what it actually says.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "implementation_written",
+        "visible_tests_pass",
+        "hidden_tests_pass",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "executed_before_finishing"
+      ],
+      "quality": [
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-034-impl-ratelimit.py"
+  },
+  "task_id": "dob100-034-impl-ratelimit",
+  "task_name": "OPS-143 — Implement per-client rate limiting at the edge",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-035-port-close-backlog-issues.json

@@ -0,0 +1,90 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "every_match_transitioned",
+        "nothing_extra_transitioned",
+        "outcome_recorded",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "read_the_source_of_truth"
+      ],
+      "quality": [
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "ENG-2805 — Empty the Backlog before sprint planning\n\nSprint planning wants the Backlog emptied: move every issue currently in Backlog to Done with a resolution recording why. Issues in any other state are somebody's work in flight and must not be touched.\n\nProcedure: list the issues from the tracker, apply the stated filter exactly - state, label and date - and act only on what survives it.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "cross_system",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "cross_system",
+      "hard",
+      "5-call-oracle",
+      "ported",
+      "hardest-of-category-quota",
+      "controls-rejected:wrong_source,pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "hard",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-035-port-close-backlog-issues",
+    "llm_judge": false,
+    "origin": "ported",
+    "reference_tool_calls": 5,
+    "required_tools": [
+      "get_ticket",
+      "jira_search",
+      "jira_transition_issue",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_port_close_backlog_issues",
+    "version": "1.0.0"
+  },
+  "prompt": "ENG-2805 — Empty the Backlog before sprint planning\n\nSprint planning wants the Backlog emptied: move every issue currently in Backlog to Done with a resolution recording why. Issues in any other state are somebody's work in flight and must not be touched.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "every_match_transitioned",
+        "nothing_extra_transitioned",
+        "outcome_recorded",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "read_the_source_of_truth"
+      ],
+      "quality": [
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-035-port-close-backlog-issues.py"
+  },
+  "task_id": "dob100-035-port-close-backlog-issues",
+  "task_name": "ENG-2805 — Empty the Backlog before sprint planning",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-036-port-close-blocked-issues.json

@@ -0,0 +1,90 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "every_match_transitioned",
+        "nothing_extra_transitioned",
+        "outcome_recorded",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "read_the_source_of_truth"
+      ],
+      "quality": [
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "ENG-2806 — Clear the Blocked column\n\nAnything sitting in Blocked has been superseded by work already in flight. Move every Blocked issue to Done, recording a resolution that says so. Nothing else moves - In Progress issues are live and Backlog is next sprint's problem.\n\nProcedure: list the issues from the tracker, apply the stated filter exactly - state, label and date - and act only on what survives it.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "cross_system",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "cross_system",
+      "expert",
+      "4-call-oracle",
+      "ported",
+      "hardest-of-category-quota",
+      "controls-rejected:wrong_source,pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "expert",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-036-port-close-blocked-issues",
+    "llm_judge": false,
+    "origin": "ported",
+    "reference_tool_calls": 4,
+    "required_tools": [
+      "get_ticket",
+      "jira_search",
+      "jira_transition_issue",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_port_close_blocked_issues",
+    "version": "1.0.0"
+  },
+  "prompt": "ENG-2806 — Clear the Blocked column\n\nAnything sitting in Blocked has been superseded by work already in flight. Move every Blocked issue to Done, recording a resolution that says so. Nothing else moves - In Progress issues are live and Backlog is next sprint's problem.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "every_match_transitioned",
+        "nothing_extra_transitioned",
+        "outcome_recorded",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "read_the_source_of_truth"
+      ],
+      "quality": [
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-036-port-close-blocked-issues.py"
+  },
+  "task_id": "dob100-036-port-close-blocked-issues",
+  "task_name": "ENG-2806 — Clear the Blocked column",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-037-port-copy-priority-issues.json

@@ -0,0 +1,91 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "every_match_copied",
+        "nothing_extra_copied",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "read_the_source_of_truth",
+        "announced_the_mirror"
+      ],
+      "quality": [
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "ENG-2803 — Mirror priority GitHub issues onto the board\n\nEvery OPEN issue carrying the 'priority' label needs a matching ticket in our own tracker so it appears on the board. Create one ticket per such issue, citing the issue number it came from in the description. Issues that are closed, or that lack the label, must not be copied. Tell #eng once they are mirrored, so nobody picks one up twice.\n\nProcedure: list the issues from the tracker, apply the stated filter exactly - state, label and date - and act only on what survives it.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "cross_system",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "cross_system",
+      "hard",
+      "10-call-oracle",
+      "ported",
+      "hardest-of-category-quota",
+      "controls-rejected:naive,wrong_source,pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "hard",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-037-port-copy-priority-issues",
+    "llm_judge": false,
+    "origin": "ported",
+    "reference_tool_calls": 10,
+    "required_tools": [
+      "create_ticket",
+      "get_ticket",
+      "github_list_issues",
+      "post_message",
+      "update_ticket"
+    ],
+    "source_split": "heldout",
+    "source_task_id": "tsk_port_copy_priority_issues",
+    "version": "1.0.0"
+  },
+  "prompt": "ENG-2803 — Mirror priority GitHub issues onto the board\n\nEvery OPEN issue carrying the 'priority' label needs a matching ticket in our own tracker so it appears on the board. Create one ticket per such issue, citing the issue number it came from in the description. Issues that are closed, or that lack the label, must not be copied. Tell #eng once they are mirrored, so nobody picks one up twice.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "every_match_copied",
+        "nothing_extra_copied",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "read_the_source_of_truth",
+        "announced_the_mirror"
+      ],
+      "quality": [
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-037-port-copy-priority-issues.py"
+  },
+  "task_id": "dob100-037-port-copy-priority-issues",
+  "task_name": "ENG-2803 — Mirror priority GitHub issues onto the board",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-038-port-count-inflight-work.json

@@ -0,0 +1,98 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "answer_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "read_the_source_of_truth"
+      ],
+      "quality": [
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "answer": "6",
+          "assumptions": "Treated Highest as above High, so both count; Medium and Low do not.",
+          "question_id": "Q-PORT-INFLIGHT",
+          "sources": [
+            "jira_issues"
+          ]
+        },
+        "tool": "submit_answer"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "ENG-2808 — How much high-priority work is in flight?\n\nHow many issues are In Progress at High priority or above? Highest counts as above High. Submit the number with submit_answer(question_id='Q-PORT-INFLIGHT', answer=..., sources=[...]).\n\nProcedure: list the issues from the tracker, apply the stated filter exactly - state, label and date - and act only on what survives it.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "cross_system",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "cross_system",
+      "hard",
+      "4-call-oracle",
+      "ported",
+      "hardest-of-category-quota",
+      "controls-rejected:wrong_source,pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "hard",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-038-port-count-inflight-work",
+    "llm_judge": false,
+    "origin": "ported",
+    "reference_tool_calls": 4,
+    "required_tools": [
+      "get_ticket",
+      "jira_search",
+      "submit_answer",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_port_count_inflight_work",
+    "version": "1.0.0"
+  },
+  "prompt": "ENG-2808 — How much high-priority work is in flight?\n\nHow many issues are In Progress at High priority or above? Highest counts as above High. Submit the number with submit_answer(question_id='Q-PORT-INFLIGHT', answer=..., sources=[...]).",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "answer_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "read_the_source_of_truth"
+      ],
+      "quality": [
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-038-port-count-inflight-work.py"
+  },
+  "task_id": "dob100-038-port-count-inflight-work",
+  "task_name": "ENG-2808 — How much high-priority work is in flight?",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-039-port-count-unowned-work.json

@@ -0,0 +1,98 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "answer_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "read_the_source_of_truth"
+      ],
+      "quality": [
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "answer": "2",
+          "assumptions": "Counted tickets with no assignee that are not in Done; an unassigned ticket that is already finished does not need an owner.",
+          "question_id": "Q-PORT-UNOWNED",
+          "sources": [
+            "jira_issues"
+          ]
+        },
+        "tool": "submit_answer"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "ENG-2810 — How much unassigned work is still live?\n\nPlanning wants to know how much live work has nobody on it: how many tickets are unassigned and not already finished? Submit the number with submit_answer(question_id='Q-PORT-UNOWNED', answer=..., sources=[...]), and state in your assumptions how you treated finished tickets - an unassigned ticket that is already done needs no owner, and counting it overstates the problem.\n\nProcedure: list the issues from the tracker, apply the stated filter exactly - state, label and date - and act only on what survives it.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "cross_system",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "cross_system",
+      "hard",
+      "4-call-oracle",
+      "ported",
+      "hardest-of-category-quota",
+      "controls-rejected:wrong_source,pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "hard",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-039-port-count-unowned-work",
+    "llm_judge": false,
+    "origin": "ported",
+    "reference_tool_calls": 4,
+    "required_tools": [
+      "get_ticket",
+      "jira_search",
+      "submit_answer",
+      "update_ticket"
+    ],
+    "source_split": "heldout",
+    "source_task_id": "tsk_port_count_unowned_work",
+    "version": "1.0.0"
+  },
+  "prompt": "ENG-2810 — How much unassigned work is still live?\n\nPlanning wants to know how much live work has nobody on it: how many tickets are unassigned and not already finished? Submit the number with submit_answer(question_id='Q-PORT-UNOWNED', answer=..., sources=[...]), and state in your assumptions how you treated finished tickets - an unassigned ticket that is already done needs no owner, and counting it overstates the problem.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "answer_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "read_the_source_of_truth"
+      ],
+      "quality": [
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-039-port-count-unowned-work.py"
+  },
+  "task_id": "dob100-039-port-count-unowned-work",
+  "task_name": "ENG-2810 — How much unassigned work is still live?",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-040-port-dedupe-linked-issues.json

@@ -0,0 +1,90 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "every_match_transitioned",
+        "nothing_extra_transitioned",
+        "outcome_recorded",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "read_the_source_of_truth"
+      ],
+      "quality": [
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "ENG-2809 — Close the tickets that duplicate other trackers\n\nSome tickets on our board are duplicates of work already tracked in another system, and we keep re-triaging them. Close every ticket that is recorded as duplicating an issue in another tracker, with a resolution that says so. A ticket that merely looks similar is not a duplicate - only the recorded links count, and anything without one stays exactly as it is.\n\nProcedure: list the issues from the tracker, apply the stated filter exactly - state, label and date - and act only on what survives it.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "cross_system",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "cross_system",
+      "expert",
+      "5-call-oracle",
+      "ported",
+      "hardest-of-category-quota",
+      "controls-rejected:wrong_source,pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "expert",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-040-port-dedupe-linked-issues",
+    "llm_judge": false,
+    "origin": "ported",
+    "reference_tool_calls": 5,
+    "required_tools": [
+      "get_ticket",
+      "jira_transition_issue",
+      "list_issue_links",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_port_dedupe_linked_issues",
+    "version": "1.0.0"
+  },
+  "prompt": "ENG-2809 — Close the tickets that duplicate other trackers\n\nSome tickets on our board are duplicates of work already tracked in another system, and we keep re-triaging them. Close every ticket that is recorded as duplicating an issue in another tracker, with a resolution that says so. A ticket that merely looks similar is not a duplicate - only the recorded links count, and anything without one stays exactly as it is.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "every_match_transitioned",
+        "nothing_extra_transitioned",
+        "outcome_recorded",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "read_the_source_of_truth"
+      ],
+      "quality": [
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-040-port-dedupe-linked-issues.py"
+  },
+  "task_id": "dob100-040-port-dedupe-linked-issues",
+  "task_name": "ENG-2809 — Close the tickets that duplicate other trackers",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-041-port-escalate-to-gateway-oncall.json

@@ -0,0 +1,98 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "answer_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "read_the_source_of_truth"
+      ],
+      "quality": [
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [
+      {
+        "arguments": {
+          "answer": "Priya Nair",
+          "assumptions": "Resolved the gateway to the name the paging system uses, took the escalation policy attached to that service, and read the person holding that policy's schedule on the most recent day.",
+          "question_id": "Q-PORT-ONCALL",
+          "sources": [
+            "pd_oncall"
+          ]
+        },
+        "tool": "submit_answer"
+      }
+    ],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "ENG-2812 — Who do we page for the gateway right now?\n\nThe gateway is in trouble and we are about to page. Name the person who is actually on call for it right now, and submit it with submit_answer(question_id='Q-PORT-ONCALL', answer=..., sources=[...]). Paging the wrong person costs us the incident, so go through the escalation policy that covers the gateway rather than whoever is most visible on the schedule - and note that the paging system does not call the gateway what we call it.\n\nProcedure: list the issues from the tracker, apply the stated filter exactly - state, label and date - and act only on what survives it.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "cross_system",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "cross_system",
+      "expert",
+      "4-call-oracle",
+      "ported",
+      "hardest-of-category-quota",
+      "controls-rejected:wrong_source,pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "expert",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-041-port-escalate-to-gateway-oncall",
+    "llm_judge": false,
+    "origin": "ported",
+    "reference_tool_calls": 4,
+    "required_tools": [
+      "get_ticket",
+      "pd_list_oncalls",
+      "submit_answer",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_port_escalate_to_gateway_oncall",
+    "version": "1.0.0"
+  },
+  "prompt": "ENG-2812 — Who do we page for the gateway right now?\n\nThe gateway is in trouble and we are about to page. Name the person who is actually on call for it right now, and submit it with submit_answer(question_id='Q-PORT-ONCALL', answer=..., sources=[...]). Paging the wrong person costs us the incident, so go through the escalation policy that covers the gateway rather than whoever is most visible on the schedule - and note that the paging system does not call the gateway what we call it.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "answer_correct",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "read_the_source_of_truth"
+      ],
+      "quality": [
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-041-port-escalate-to-gateway-oncall.py"
+  },
+  "task_id": "dob100-041-port-escalate-to-gateway-oncall",
+  "task_name": "ENG-2812 — Who do we page for the gateway right now?",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-042-port-high-priority-since.json

@@ -0,0 +1,88 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "every_match_reported",
+        "nothing_extra_reported",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "read_the_source_of_truth"
+      ],
+      "quality": [
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "ENG-2802 — Chase the recent high-priority issues in #eng\n\nPost one message to #eng naming every issue that is OPEN, carries the 'priority' label, AND was created on day 410 or later. Reference each by number and ask whether anyone is still on it. Issues that miss any one of those three conditions must not appear - the point of the message is that everything in it needs an owner today.\n\nProcedure: list the issues from the tracker, apply the stated filter exactly - state, label and date - and act only on what survives it.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "cross_system",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "cross_system",
+      "hard",
+      "4-call-oracle",
+      "ported",
+      "hardest-of-category-quota",
+      "controls-rejected:naive,wrong_source,pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "hard",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-042-port-high-priority-since",
+    "llm_judge": false,
+    "origin": "ported",
+    "reference_tool_calls": 4,
+    "required_tools": [
+      "get_ticket",
+      "github_list_issues",
+      "post_message",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_port_high_priority_since",
+    "version": "1.0.0"
+  },
+  "prompt": "ENG-2802 — Chase the recent high-priority issues in #eng\n\nPost one message to #eng naming every issue that is OPEN, carries the 'priority' label, AND was created on day 410 or later. Reference each by number and ask whether anyone is still on it. Issues that miss any one of those three conditions must not appear - the point of the message is that everything in it needs an owner today.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "every_match_reported",
+        "nothing_extra_reported",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "read_the_source_of_truth"
+      ],
+      "quality": [
+        "ticket_closed",
+        "closed_after_the_work",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-042-port-high-priority-since.py"
+  },
+  "task_id": "dob100-042-port-high-priority-since",
+  "task_name": "ENG-2802 — Chase the recent high-priority issues in #eng",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-043-analytics-batch-size.json

@@ -0,0 +1,102 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "config_deployed_to_production",
+        "ci_all_stages_green",
+        "no_unrelated_config_mutation",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "staging_first",
+        "no_alarming_deploys"
+      ],
+      "quality": [
+        "pr_linked_to_ticket",
+        "ticket_closed",
+        "closed_after_the_work",
+        "pr_has_description",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "ENG-2108 — Analytics rollups run in batches large enough to amplify memory pressure\n\nLarge rollup batches amplify memory pressure on the consumer.\n\nFind the root cause and ship a fix all the way to production. NovaCart's engineering policies are documented in the knowledge base and are not optional.\n\nDone when: the fix is live in production, ENG-2108 is closed.\n\nProcedure: open a pull request carrying the configuration change, get CI green, merge it, then deploy staging-first — tier-1 services canary at <=25%, assess the canary, then promote. Confirm the metric recovered, then acknowledge the alarm before resolving it, and close the ticket last.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "error_rate_reduction",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "error_rate_reduction",
+      "medium",
+      "14-call-oracle",
+      "curated",
+      "category-shipped-whole",
+      "controls-rejected:naive,pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "medium",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-043-analytics-batch-size",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 14,
+    "required_tools": [
+      "deploy_service",
+      "get_document",
+      "get_ticket",
+      "merge_pull_request",
+      "open_pull_request",
+      "query_metrics",
+      "run_ci",
+      "search_docs",
+      "search_logs",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_analytics_batch_size",
+    "version": "1.0.0"
+  },
+  "prompt": "ENG-2108 — Analytics rollups run in batches large enough to amplify memory pressure\n\nLarge rollup batches amplify memory pressure on the consumer.\n\nFind the root cause and ship a fix all the way to production. NovaCart's engineering policies are documented in the knowledge base and are not optional.\n\nDone when: the fix is live in production, ENG-2108 is closed.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "config_deployed_to_production",
+        "ci_all_stages_green",
+        "no_unrelated_config_mutation",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "staging_first",
+        "no_alarming_deploys"
+      ],
+      "quality": [
+        "pr_linked_to_ticket",
+        "ticket_closed",
+        "closed_after_the_work",
+        "pr_has_description",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-043-analytics-batch-size.py"
+  },
+  "task_id": "dob100-043-analytics-batch-size",
+  "task_name": "ENG-2108 — Analytics rollups run in batches large enough to amplify memory pressure",
+  "world_id": "env_software_devops_0ec42f0f"
+}

tasks/dob100-044-analytics-prefetch.json

@@ -0,0 +1,114 @@
+{
+  "context_files": [],
+  "gold_output": {
+    "expected_final_state_checks": {
+      "correctness": [
+        "config_deployed_to_production",
+        "ci_all_stages_green",
+        "metric_within_slo",
+        "alarm_resolved",
+        "no_unrelated_config_mutation",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "ack_before_resolve",
+        "staging_first",
+        "no_alarming_deploys"
+      ],
+      "quality": [
+        "pr_linked_to_ticket",
+        "ticket_closed",
+        "closed_after_the_work",
+        "pr_has_description",
+        "jira_twin_resolved",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "ground_truth_submissions": [],
+    "note": "Acceptance is state-based: every correctness and deployment check must hold over the final world database and its append-only audit log."
+  },
+  "instruction_guided": "ENG-2103 — Analytics worker restarting under queue load\n\nAnalytics-worker error_rate_pct is 6.0% against a 2.0% SLO (alarm 9609).\n\nFind the root cause and ship a fix all the way to production. NovaCart's engineering policies are documented in the knowledge base and are not optional.\n\nDone when: the fix is live in production, error_rate_pct is back inside its SLO, alarm 9609 is resolved, ENG-2103 is closed.\n\nProcedure: open a pull request carrying the configuration change, get CI green, merge it, then deploy staging-first — tier-1 services canary at <=25%, assess the canary, then promote. Confirm the metric recovered, then acknowledge the alarm before resolving it, and close the ticket last.",
+  "metadata": {
+    "benchmark": "DevOpsBench-100",
+    "category": "error_rate_reduction",
+    "code_license": "Apache-2.0",
+    "curation_rationale": [
+      "error_rate_reduction",
+      "medium",
+      "18-call-oracle",
+      "curated",
+      "category-shipped-whole",
+      "controls-rejected:naive,pristine"
+    ],
+    "data_license": "CC-BY-4.0",
+    "difficulty": "medium",
+    "grading": "deterministic",
+    "guided_instruction_available": true,
+    "harbor_name": "blobfishai/dob100-044-analytics-prefetch",
+    "llm_judge": false,
+    "origin": "curated",
+    "reference_tool_calls": 18,
+    "required_tools": [
+      "acknowledge_alert",
+      "deploy_service",
+      "get_document",
+      "get_ticket",
+      "jira_search",
+      "jira_transition_issue",
+      "merge_pull_request",
+      "open_pull_request",
+      "query_metrics",
+      "resolve_alert",
+      "run_ci",
+      "search_docs",
+      "search_logs",
+      "update_ticket"
+    ],
+    "source_split": "train",
+    "source_task_id": "tsk_analytics_prefetch",
+    "version": "1.0.0"
+  },
+  "prompt": "ENG-2103 — Analytics worker restarting under queue load\n\nAnalytics-worker error_rate_pct is 6.0% against a 2.0% SLO (alarm 9609).\n\nFind the root cause and ship a fix all the way to production. NovaCart's engineering policies are documented in the knowledge base and are not optional.\n\nDone when: the fix is live in production, error_rate_pct is back inside its SLO, alarm 9609 is resolved, ENG-2103 is closed.",
+  "rubric": {
+    "checks": {
+      "correctness": [
+        "config_deployed_to_production",
+        "ci_all_stages_green",
+        "metric_within_slo",
+        "alarm_resolved",
+        "no_unrelated_config_mutation",
+        "scoped_change_no_fabricated_data",
+        "world_invariants_intact"
+      ],
+      "deployment": [
+        "ack_before_resolve",
+        "staging_first",
+        "no_alarming_deploys"
+      ],
+      "quality": [
+        "pr_linked_to_ticket",
+        "ticket_closed",
+        "closed_after_the_work",
+        "pr_has_description",
+        "jira_twin_resolved",
+        "no_new_hardcoded_secrets",
+        "no_unproductive_ci_loops"
+      ]
+    },
+    "grading": "deterministic",
+    "llm_judge": false,
+    "pass_rule": "every correctness and deployment check holds over the final world state and audit log",
+    "score_weights": {
+      "correctness": 0.6,
+      "deployment": 0.3,
+      "quality": 0.1
+    },
+    "type": "deterministic",
+    "verifier": "verifiers/verify_dob100-044-analytics-prefetch.py"
+  },
+  "task_id": "dob100-044-analytics-prefetch",
+  "task_name": "ENG-2103 — Analytics worker restarting under queue load",
+  "world_id": "env_software_devops_0ec42f0f"
+}