Thorsu/MIHA_HUNTING_Evidence
Question Answering • Updated
id stringlengths 14 14 | schema_version stringclasses 1
value | track stringlengths 11 23 | difficulty int64 1 4 | question stringlengths 31 75 | contract dict | sources listlengths 1 3 | neighbor_policy dict | output_requirements dict |
|---|---|---|---|---|---|---|---|---|
miha-bench-001 | 1.1 | runtime_gate | 1 | May release Q4 be marked green? | {
"as_of": "2026-10-05T13:00:00",
"required_search_spaces": [
"ci",
"runtime"
],
"allowed_source_ids": [
"a",
"b"
],
"forbidden_source_ids": []
} | [
{
"id": "a",
"space": "ci",
"text": "Q4 tests: 90 passed, 0 failed."
},
{
"id": "b",
"space": "runtime",
"text": "Q4 runtime probe timed out. Green requires both tests and runtime probe to pass."
}
] | {
"branching_factor": 4,
"restart_on_verified_find": true,
"max_depth": 10,
"max_admitted_paths_per_depth": 5
} | {
"answer_type": "string",
"must_return_used_sources": true,
"must_return_claims": true,
"must_return_dead_end": true,
"must_return_visited_paths": true
} |
miha-bench-002 | 1.1 | revision_supersession | 1 | Which inventory count is current? | {
"as_of": "2026-10-05T13:05:00",
"required_search_spaces": [
"inventory"
],
"allowed_source_ids": [
"a",
"b"
],
"forbidden_source_ids": []
} | [
{
"id": "a",
"space": "inventory",
"text": "Inventory v3 reports 18 nodes."
},
{
"id": "b",
"space": "inventory",
"text": "Inventory v4 revises v3 and reports 17 nodes."
}
] | {
"branching_factor": 4,
"restart_on_verified_find": true,
"max_depth": 10,
"max_admitted_paths_per_depth": 5
} | {
"answer_type": "string",
"must_return_used_sources": true,
"must_return_claims": true,
"must_return_dead_end": true,
"must_return_visited_paths": true
} |
miha-bench-003 | 1.1 | missing_primary | 2 | Can the claimed deployment date be verified? | {
"as_of": "2026-10-05T13:10:00",
"required_search_spaces": [
"deployment-registry"
],
"allowed_source_ids": [
"a"
],
"forbidden_source_ids": []
} | [
{
"id": "a",
"space": "forum",
"text": "A forum post claims deployment happened Friday; no registry record or primary citation is available."
}
] | {
"branching_factor": 4,
"restart_on_verified_find": true,
"max_depth": 10,
"max_admitted_paths_per_depth": 5
} | {
"answer_type": "string",
"must_return_used_sources": true,
"must_return_claims": true,
"must_return_dead_end": true,
"must_return_visited_paths": true
} |
miha-bench-004 | 1.1 | temporal_cutoff | 2 | As of 14:00Z, was service S known to be recovered? | {
"as_of": "2026-10-05T14:00:00",
"required_search_spaces": [
"status"
],
"allowed_source_ids": [
"a"
],
"forbidden_source_ids": [
"b"
]
} | [
{
"id": "a",
"space": "status",
"timestamp": "2026-10-05T13:50:00Z",
"text": "13:50Z: service S remains degraded."
},
{
"id": "b",
"space": "status",
"timestamp": "2026-10-05T14:07:00Z",
"text": "14:07Z: service S recovered."
}
] | {
"branching_factor": 4,
"restart_on_verified_find": true,
"max_depth": 10,
"max_admitted_paths_per_depth": 5
} | {
"answer_type": "string",
"must_return_used_sources": true,
"must_return_claims": true,
"must_return_dead_end": true,
"must_return_visited_paths": true
} |
miha-bench-005 | 1.1 | access_policy | 2 | May Noor read production traces? | {
"as_of": "2026-10-05T14:10:00",
"required_search_spaces": [
"identity",
"policy"
],
"allowed_source_ids": [
"a",
"b"
],
"forbidden_source_ids": []
} | [
{
"id": "a",
"space": "identity",
"text": "Noor role: reviewer."
},
{
"id": "b",
"space": "policy",
"text": "Reviewers may read production traces but may not write them."
}
] | {
"branching_factor": 4,
"restart_on_verified_find": true,
"max_depth": 10,
"max_admitted_paths_per_depth": 5
} | {
"answer_type": "string",
"must_return_used_sources": true,
"must_return_claims": true,
"must_return_dead_end": true,
"must_return_visited_paths": true
} |
miha-bench-006 | 1.1 | verified_absence | 2 | Is a rollback report absent from the declared search space? | {
"as_of": "2026-10-05T14:20:00",
"required_search_spaces": [
"reports",
"attachments"
],
"allowed_source_ids": [
"a",
"b"
],
"forbidden_source_ids": []
} | [
{
"id": "a",
"space": "reports",
"text": "Complete reports search for rollback R2 returned zero records."
},
{
"id": "b",
"space": "attachments",
"text": "Complete attachments search for rollback R2 returned zero report files."
}
] | {
"branching_factor": 4,
"restart_on_verified_find": true,
"max_depth": 10,
"max_admitted_paths_per_depth": 5
} | {
"answer_type": "string",
"must_return_used_sources": true,
"must_return_claims": true,
"must_return_dead_end": true,
"must_return_visited_paths": true
} |
miha-bench-007 | 1.1 | contradiction | 3 | Is build B17 checksum verified? | {
"as_of": "2026-10-05T15:00:00",
"required_search_spaces": [
"release-registry",
"deployment-log"
],
"allowed_source_ids": [
"a",
"b"
],
"forbidden_source_ids": []
} | [
{
"id": "a",
"space": "release-registry",
"text": "Build B17 checksum is sha256:aaa111."
},
{
"id": "b",
"space": "deployment-log",
"text": "Deployment log for Build B17 records checksum sha256:bbb222."
}
] | {
"branching_factor": 4,
"restart_on_verified_find": true,
"max_depth": 10,
"max_admitted_paths_per_depth": 5
} | {
"answer_type": "string",
"must_return_used_sources": true,
"must_return_claims": true,
"must_return_dead_end": true,
"must_return_visited_paths": true
} |
miha-bench-008 | 1.1 | identity_disambiguation | 3 | Can Nora#17 rotate production keys? | {
"as_of": "2026-10-05T15:05:00",
"required_search_spaces": [
"identity",
"policy"
],
"allowed_source_ids": [
"a",
"c"
],
"forbidden_source_ids": [
"b"
]
} | [
{
"id": "a",
"space": "identity",
"text": "Nora#17 role: reviewer."
},
{
"id": "b",
"space": "identity",
"text": "Nora#71 role: admin."
},
{
"id": "c",
"space": "policy",
"text": "Only admins may rotate production keys."
}
] | {
"branching_factor": 4,
"restart_on_verified_find": true,
"max_depth": 10,
"max_admitted_paths_per_depth": 5
} | {
"answer_type": "string",
"must_return_used_sources": true,
"must_return_claims": true,
"must_return_dead_end": true,
"must_return_visited_paths": true
} |
miha-bench-009 | 1.1 | privacy_export_gate | 3 | Can this case be auto-exported to a public evidence dataset? | {
"as_of": "2026-10-05T15:10:00",
"required_search_spaces": [
"source-policy"
],
"allowed_source_ids": [
"a",
"b"
],
"forbidden_source_ids": []
} | [
{
"id": "a",
"space": "source-policy",
"visibility": "public",
"text": "Synthetic control cases may be exported when no private content is included."
},
{
"id": "b",
"space": "source-policy",
"visibility": "private",
"text": "The case includes an internal email body. Private conn... | {
"branching_factor": 4,
"restart_on_verified_find": true,
"max_depth": 10,
"max_admitted_paths_per_depth": 5
} | {
"answer_type": "string",
"must_return_used_sources": true,
"must_return_claims": true,
"must_return_dead_end": true,
"must_return_visited_paths": true
} |
miha-bench-010 | 1.1 | multi_parent_deduction | 3 | Is artifact A12 eligible for release? | {
"as_of": "2026-10-05T15:20:00",
"required_search_spaces": [
"license",
"ci",
"runtime"
],
"allowed_source_ids": [
"a",
"b",
"c"
],
"forbidden_source_ids": []
} | [
{
"id": "a",
"space": "license",
"text": "Artifact A12 license scan: Apache-2.0 compatible."
},
{
"id": "b",
"space": "ci",
"text": "Artifact A12 regression suite: passed."
},
{
"id": "c",
"space": "runtime",
"text": "Artifact A12 runtime smoke probe: passed."
}
] | {
"branching_factor": 4,
"restart_on_verified_find": true,
"max_depth": 10,
"max_admitted_paths_per_depth": 5
} | {
"answer_type": "string",
"must_return_used_sources": true,
"must_return_claims": true,
"must_return_dead_end": true,
"must_return_visited_paths": true
} |
miha-bench-011 | 1.1 | causality_guard | 3 | May event E1 be claimed as the cause of outage O9? | {
"as_of": "2026-10-05T15:30:00",
"required_search_spaces": [
"events",
"joins"
],
"allowed_source_ids": [
"a",
"b"
],
"forbidden_source_ids": []
} | [
{
"id": "a",
"space": "events",
"text": "Event E1 occurred at 10:00Z. Outage O9 started at 10:03Z."
},
{
"id": "b",
"space": "joins",
"text": "No shared incident_id, service_id, or ticket_id links E1 to O9."
}
] | {
"branching_factor": 4,
"restart_on_verified_find": true,
"max_depth": 10,
"max_admitted_paths_per_depth": 5
} | {
"answer_type": "string",
"must_return_used_sources": true,
"must_return_claims": true,
"must_return_dead_end": true,
"must_return_visited_paths": true
} |
miha-bench-012 | 1.1 | locator_dead_end | 3 | Can the audit claim be admitted as verified? | {
"as_of": "2026-10-05T15:40:00",
"required_search_spaces": [
"audit-log"
],
"allowed_source_ids": [
"a"
],
"forbidden_source_ids": []
} | [
{
"id": "a",
"space": "secondary-summary",
"text": "A summary says an audit log exists but gives no document ID, timestamp, page, or path."
}
] | {
"branching_factor": 4,
"restart_on_verified_find": true,
"max_depth": 10,
"max_admitted_paths_per_depth": 5
} | {
"answer_type": "string",
"must_return_used_sources": true,
"must_return_claims": true,
"must_return_dead_end": true,
"must_return_visited_paths": true
} |
miha-bench-013 | 1.1 | policy_revision | 3 | As of 16:00Z, may reviewers read production traces? | {
"as_of": "2026-10-05T16:00:00",
"required_search_spaces": [
"policy"
],
"allowed_source_ids": [
"a",
"b"
],
"forbidden_source_ids": []
} | [
{
"id": "a",
"space": "policy",
"timestamp": "2026-10-05T14:00:00Z",
"text": "Policy v2: reviewers may read production traces."
},
{
"id": "b",
"space": "policy",
"timestamp": "2026-10-05T15:30:00Z",
"text": "Policy v3 revises v2: reviewers may not read production traces."
}
] | {
"branching_factor": 4,
"restart_on_verified_find": true,
"max_depth": 10,
"max_admitted_paths_per_depth": 5
} | {
"answer_type": "string",
"must_return_used_sources": true,
"must_return_claims": true,
"must_return_dead_end": true,
"must_return_visited_paths": true
} |
miha-bench-014 | 1.1 | policy_cutoff | 3 | As of 15:00Z, may reviewers read production traces? | {
"as_of": "2026-10-05T15:00:00",
"required_search_spaces": [
"policy"
],
"allowed_source_ids": [
"a"
],
"forbidden_source_ids": [
"b"
]
} | [
{
"id": "a",
"space": "policy",
"timestamp": "2026-10-05T14:00:00Z",
"text": "Policy v2: reviewers may read production traces."
},
{
"id": "b",
"space": "policy",
"timestamp": "2026-10-05T15:30:00Z",
"text": "Policy v3 revises v2: reviewers may not read production traces."
}
] | {
"branching_factor": 4,
"restart_on_verified_find": true,
"max_depth": 10,
"max_admitted_paths_per_depth": 5
} | {
"answer_type": "string",
"must_return_used_sources": true,
"must_return_claims": true,
"must_return_dead_end": true,
"must_return_visited_paths": true
} |
miha-bench-015 | 1.1 | required_path_coverage | 3 | Is incident I5 root-cause analysis available and complete? | {
"as_of": "2026-10-05T16:10:00",
"required_search_spaces": [
"incidents",
"attachments"
],
"allowed_source_ids": [
"a",
"b"
],
"forbidden_source_ids": []
} | [
{
"id": "a",
"space": "incidents",
"text": "Incident I5 record links RCA attachment rca-i5.md."
},
{
"id": "b",
"space": "attachments",
"text": "Attachment rca-i5.md contains root_cause, impact, remediation, and owner fields."
}
] | {
"branching_factor": 4,
"restart_on_verified_find": true,
"max_depth": 10,
"max_admitted_paths_per_depth": 5
} | {
"answer_type": "string",
"must_return_used_sources": true,
"must_return_claims": true,
"must_return_dead_end": true,
"must_return_visited_paths": true
} |
miha-bench-016 | 1.1 | canonical_vs_mirror | 4 | Which MIHA auto-row count is authoritative? | {
"as_of": "2026-10-05T16:20:00",
"required_search_spaces": [
"canonical-manifest",
"mirror-status"
],
"allowed_source_ids": [
"a",
"b"
],
"forbidden_source_ids": []
} | [
{
"id": "a",
"space": "mirror-status",
"text": "Cached viewer mirror shows 8 rows and is marked stale."
},
{
"id": "b",
"space": "canonical-manifest",
"text": "Canonical manifest version 12 lists 12 auto_rows and supersedes stale mirrors."
}
] | {
"branching_factor": 4,
"restart_on_verified_find": true,
"max_depth": 10,
"max_admitted_paths_per_depth": 5
} | {
"answer_type": "string",
"must_return_used_sources": true,
"must_return_claims": true,
"must_return_dead_end": true,
"must_return_visited_paths": true
} |
miha-bench-017 | 1.1 | graph_integrity | 4 | Is the evidence graph structurally valid? | {
"as_of": "2026-10-05T16:30:00",
"required_search_spaces": [
"graph-nodes",
"graph-edges"
],
"allowed_source_ids": [
"a",
"b"
],
"forbidden_source_ids": []
} | [
{
"id": "a",
"space": "graph-nodes",
"text": "Node set contains n1 and n2 only."
},
{
"id": "b",
"space": "graph-edges",
"text": "Edge e1 references source n1 and target n3."
}
] | {
"branching_factor": 4,
"restart_on_verified_find": true,
"max_depth": 10,
"max_admitted_paths_per_depth": 5
} | {
"answer_type": "string",
"must_return_used_sources": true,
"must_return_claims": true,
"must_return_dead_end": true,
"must_return_visited_paths": true
} |
miha-bench-018 | 1.1 | holdout_boundary | 4 | May this live hunt checkpoint be copied into miha-bench automatically? | {
"as_of": "2026-10-05T16:40:00",
"required_search_spaces": [
"miha-policy",
"checkpoint"
],
"allowed_source_ids": [
"a",
"b"
],
"forbidden_source_ids": []
} | [
{
"id": "a",
"space": "checkpoint",
"text": "Checkpoint C9 was generated from the current live hunt."
},
{
"id": "b",
"space": "miha-policy",
"text": "miha-bench is a separate curated holdout; benchmark gold cases are not generated automatically from the same hunts."
}
] | {
"branching_factor": 4,
"restart_on_verified_find": true,
"max_depth": 10,
"max_admitted_paths_per_depth": 5
} | {
"answer_type": "string",
"must_return_used_sources": true,
"must_return_claims": true,
"must_return_dead_end": true,
"must_return_visited_paths": true
} |
miha-bench-019 | 1.1 | path_budget | 4 | May the sixth same-depth candidate be admitted in the current matrix level? | {
"as_of": "2026-10-05T16:50:00",
"required_search_spaces": [
"matrix-policy",
"candidate-list"
],
"allowed_source_ids": [
"a",
"b"
],
"forbidden_source_ids": []
} | [
{
"id": "a",
"space": "matrix-policy",
"text": "At each depth, at most five paths may be admitted overall; extra candidates go to backlog."
},
{
"id": "b",
"space": "candidate-list",
"text": "The current depth already has five admitted paths; candidate P6 is next by priority."
}
] | {
"branching_factor": 4,
"restart_on_verified_find": true,
"max_depth": 10,
"max_admitted_paths_per_depth": 5
} | {
"answer_type": "string",
"must_return_used_sources": true,
"must_return_claims": true,
"must_return_dead_end": true,
"must_return_visited_paths": true
} |
miha-bench-020 | 1.1 | dual_readback | 4 | Is the dataset export readback verified? | {
"as_of": "2026-10-05T17:00:00",
"required_search_spaces": [
"canonical-readback",
"viewer-readback"
],
"allowed_source_ids": [
"a",
"b"
],
"forbidden_source_ids": []
} | [
{
"id": "a",
"space": "canonical-readback",
"text": "Canonical record readback: record_id r7, content_sha256 h1."
},
{
"id": "b",
"space": "viewer-readback",
"text": "Viewer row readback: record_id r7, content_sha256 h2."
}
] | {
"branching_factor": 4,
"restart_on_verified_find": true,
"max_depth": 10,
"max_admitted_paths_per_depth": 5
} | {
"answer_type": "string",
"must_return_used_sources": true,
"must_return_claims": true,
"must_return_dead_end": true,
"must_return_visited_paths": true
} |
MIHA-Bench
MIHA-Bench is the deterministic holdout benchmark for MIHA-style evidence-path reasoning. It scores structured answers, cited source IDs, required search-space coverage, dead-end handling, temporal cutoffs, access boundaries, revision supersession, graph integrity, and dual-readback claims without an LLM judge.
The benchmark is deliberately separate from live MIHA hunts. Live hunts may export to Thorsu/miha-evidence; they must not automatically become MIHA-Bench gold cases.
Data
data/test.jsonl contains 20 curated synthetic control tasks. Each row declares:
question: the task to answer,contract: the as-of timestamp, required search spaces, allowed source IDs, and forbidden source IDs,sources: the task-local evidence package,neighbor_policy: MIHA's 4-neighbor restart rule with the global 10 x 5 bound,output_requirements: fields expected from a submission.
gold/answers.jsonl contains the deterministic gold labels for regression and harness development. For an external leaderboard, publish the task split and keep gold labels in a separate evaluator service or sealed release.
Submission format
One JSON object per test record:
{"id":"miha-bench-001","answer":"no","used_sources":["a","b"],"claims":["tests_passed"],"dead_end":false,"visited_paths":["ci","runtime"]}
Run the evaluator
python eval.py gold/answers.jsonl predictions.jsonl --tasks data/test.jsonl
For a smoke test:
python tools/validate_bench.py
python eval.py gold/answers.jsonl baselines/oracle_predictions.jsonl --tasks data/test.jsonl
Current tracks
runtime_gaterevision_supersessionmissing_primarytemporal_cutoffaccess_policyverified_absencecontradictionidentity_disambiguationprivacy_export_gatemulti_parent_deductioncausality_guardlocator_dead_endpolicy_revisionpolicy_cutoffrequired_path_coveragecanonical_vs_mirrorgraph_integrityholdout_boundarypath_budgetdual_readback
Design notes
MIHA-Bench emphasizes path validity over polished prose. A model that guesses the right answer but uses a forbidden future source or skips a required search space should not receive a perfect score. The evaluator therefore reports both aggregate score and per-track diagnostics.
- Downloads last month
- 94
Models trained or fine-tuned on Thorsu/miha-bench