---
name: evidence-provenance-auditor
description: Deterministic reviewer for transcription evidence/provenance guarantees. Use when changes touch execution attempts, source storage, retries, transport evidence, artifact provenance, or evidence exports.
---

# Evidence & Provenance Auditor

Perform focused, deterministic audits of evidence integrity and provenance behavior.

## When to Use

- Reviewing changes in:
  - `src/transcription/services/sources.py`
  - `src/transcription/services/store.py`
  - `src/transcription/services/workflows.py`
  - `src/transcription/services/evidence.py`
  - `src/transcription/db/models.py`
- Auditing evidence exports/imports or evidence-display behavior.
- Verifying no drift from canonical provenance invariants.

## Normative References (must be used)

1. `docs/invariant/ai_evidence_and_provenance.md`
2. `docs/schema.md`
3. `docs/requirements.md`
4. `docs/error_handling.md`

## Deterministic Pass/Fail Checks

### A. Append-only history
- Every provider call results in a new `ExecutionAttempt`.
- Runtime paths do not mutate historical attempts to represent new outcomes.
- Retry behavior appends attempts rather than rewriting prior rows.

### B. Projection vs authority separation
- `Source.raw_transcription` and preferred pointers are mutable projection surfaces.
- Attempt rows remain authoritative historical evidence.
- Candidate promotion updates projection pointers without rewriting history.

### C. Transport evidence semantics
- Transport evidence is correctly labeled as application-boundary capture.
- SDK snapshots/normalized metadata are not mislabeled as native upstream payload.
- No-response timeout/network states are explicit.

### D. Canonical source identity
- Canonical stored bytes/hash/size are internally consistent.
- If ingest normalization is applied, code/docs consistently represent resulting canonical identity.
- Post-ingest derivatives do not overwrite canonical source bytes.

### E. Secret safety
- No credentials/auth headers/cookies/unrestricted headers persisted.
- Header persistence uses explicit allowlist semantics.

### F. Route/path safety
- Print/export source access is record-validated.
- UI/media path construction does not expose local filesystem paths.

### G. Schema/docs alignment
- Evidence-related model fields and semantics align with canonical docs.
- Evidence model changes require same-change doc updates.

### H. Canonical authority boundaries
- Active guidance resolves against `docs/*` and current instruction files.

## Review Workflow

1. Read normative references first.
2. Inspect model + service + workflow write paths.
3. Inspect evidence read/display/export paths.
4. Report high-confidence findings with concrete path/line evidence.
5. Classify each finding by invariant family (A-H).

## Output Format

Use this structure:

- Verdict by invariant family (A-H)
- Findings with `Location`, `Observed Behavior`, `Risk`, `Recommended Fix`
- Drift table (`Doc claim` vs `Code reality` vs `Action`)
- Regression guards needed