Gate A — Machine-verifiable evidence contract for claim–code findings

[LucaXTech]

Decision context

This issue no longer covers UI highlighting or a review-workstation interface. Its purpose is to define the evidence substrate required by the BioAudit research kill-test.

BioAudit must not be trusted because a model sounds convincing. Every scored finding must be independently checkable against the supplied claim, code, or execution artifact.

Scope boundary

This issue does not reimplement:

• leakage-safe resampling, guarded preprocessing, or post-hoc leakage diagnostics already provided by bioLeak;
• dependency-aware split construction from splitGraph;
• BIDS dataset validation or MNE-BIDS processing;
• EEG benchmark-document construction and document-kernel maintenance from NeuroDoc / NeuroAudit;
• a general-purpose paper-code discrepancy system such as SciCoQA.

Goal

Define a deterministic, versioned JSON contract that can represent an evidence-linked methodological finding for benchmark scoring.

Required finding fields

• stable finding_id;
• finite rule_id and taxonomy_version;
• severity;
• status: DETECTED, INSUFFICIENT_EVIDENCE, or NOT_APPLICABLE;
• one or more typed evidence anchors:
  • claim/method text: source name, character offsets, exact excerpt hash;
  • code: repository-relative path, line range, exact excerpt hash;
  • execution artifact: artifact path, structured field or row identifier, value hash;
• origin: deterministic checker, model, human key, or external validator;
• bounded rationale;
• repair action;
• executable or manually reproducible verification step;
• explicit missing-information fields.

Acceptance criteria

• a JSON Schema validates every benchmark finding deterministically;
• every evidence anchor can rehydrate the exact cited excerpt and reproduce its hash;
• fabricated, out-of-range, stale, or source-mismatched anchors are rejected;
• a scored finding has valid evidence or an explicit INSUFFICIENT_EVIDENCE state;
• stable IDs do not depend on model wording;
• clean/defective twins preserve IDs for unchanged findings;
• exports preserve rule version, model/protocol metadata, and artifact provenance;
• tests cover text, code, artifact, missing-evidence, and corrupted-anchor cases;
• no UI work is required for completion.

Stop condition

If exact evidence anchoring cannot be scored reliably without substantial manual interpretation, do not proceed to product UI.