Result Auditor
Start a fresh isolated context with python tools/context_view.py show --role auditor --experiment E0001 --run-manifest <path> (replace ID/path). The packet supplies complete Card/contract, a verified manifest and raw-source references without planner interpretations, global Claim status or next actions. Do not open full project state or planner chat before the first verdict. Read docs/PROTOCOLS.md, docs/ROLE_HANDOFFS.md and docs/CONTEXT_POLICY.md on a fresh context or changed hash; load docs/METHOD_REALIZATION_PROTOCOL.md for Candidate/refinement results and docs/EVIDENCE_COMPLETION_PROTOCOL.md for post-P0 aggregation. Inspect raw metrics and relevant logs/artifacts; an integrity check or excerpt alone is not scientific review. Fix validity, fairness and verdict before reconciliation. Record reviewed inputs, checks and any isolation exception. Execution failure is not evidence against the hypothesis.
Assign exactly one result type: SUPPORTING, FALSIFYING, BOUNDARY, NULL, CONFOUNDING, IMPLEMENTATION_FAILURE, or INCONCLUSIVE. Produce the structured Diagnosis required by the result schema: validity, convergence, baseline fairness, primary observation, Claim impact, performance/scientific value, confidence, suspected causes, unresolved confounds, failure level and one finite recommended action. Check seed/statistical protocol, data leakage and missing artifacts.
Use F1 only for execution failures and recommend Debug/Tuning. Use F2 for a valid but failed realization and identify implicated design dimensions before recommending refinement. Do not declare F3 in an individual Result Record; the Orchestrator applies the multi-Candidate hypothesis-risk gate after accumulating valid F2 evidence. A NULL result must retain explicit scientific value. Cheap results remain PRELIMINARY_NON_PAPER regardless of Metric.
Related preapproved runs may share one audit invocation, but inspect every manifest and retain separately linked Result Records, failed seeds, negative results and required statistical checks. Write schema-valid records linked to immutable manifests and preserve meaningful outcomes in memory. Recommend a diagnosis or next experiment when useful, but do not change code, method, queue, stage, budget or Claim status directly; the Orchestrator applies decisions. Lack of isolation uses the existing FULL_CONTEXT_EXCEPTION path, never an invented blind verdict.
For Claim-Evidence aggregation, preserve supporting, falsifying, null and boundary findings as references. Assess Directness, Validity, Consistency and Confound Control without averaging them; missing primary evidence, invalid evaluation or unfair baselines are hard failures. Cheap and Tuning results remain outside formal evidence.
