Evidence
What the export records
Exports preserve the captured exchange, the recorded review decision and the information needed to check integrity. JSON carries structured data; HTML and PDF provide readable reports. A report is not a certification of the AI answer.
Per-exchange fields
| Field | Purpose |
|---|---|
| sequence / record_id | Ordering and identity in the chain |
| timestamp / timestamp_iso | When the exchange was recorded (from your computer’s clock) |
| prompt_hash / response_hash | Content hashes for prompt and AI response |
| chain_hash / previous_hash | SHA-256 chain linkage (block_hash alias) |
| user_message / ai_response | Captured text is preserved as received. Oversized submissions are rejected rather than silently shortened. MCP submissions are labelled self-reported. |
| model / model_version | Model identifiers when available |
| governance.* | Action, severity, categories from the engine |
| human_reviewer.* | Reviewer name (typed in, not verified), Accept/Reject/Edit, time, edit_delta |
Bundle readiness
An export is marked ready only when its review and integrity checks pass. Pending exchanges and incomplete history are explicit warnings. A session-only extract is not a complete-chain proof. Even a valid complete chain cannot, by itself, prove that no newer records were deleted or that the entire history was not replaced.
Offline verifier
The current build includes a standalone offline verifier for complete JSON records and evidence ZIPs. It checks the exported chain without contacting FairWitnessAI. A session-only extract cannot prove the integrity of omitted history. Offline verification does not establish factual accuracy, reviewer identity or completeness of the original conversation.