Model, prompt version, transcript citation, rubric, regulator framework — recorded for every step. Replay any decision the way the regulator will read it.
Trusted by recruiting teams shipping audit-grade hiring
01 · What
Per decision: model (e.g. Claude Sonnet 4.6), prompt_version (e.g. rubric v3), transcript_citation (line range in the recorded transcript), rubric_id, framework (NYC LL144 §20-871, EU AI Act tier, GDPR clause, PDPA principle). Every recruiter-visible score links back to its log entry; every model output links back to its prompt and inputs.
02 · Replay
From the audit trail page, the recruiter or auditor selects a decision row and opens a replay surface that re-renders the model's input plus the surrounding state at that moment. Frozen prompt + frozen rubric guarantees deterministic replay — two replays return the same answer. A regulator in 2031 sees what you saw in 2026.
03 · Retention
Filesystem (data/) retained 7y. PostgreSQL interviews row + state JSONB retained 7y. Coding-challenge GitHub repos retained 7y or until the candidate requests deletion (GDPR right-to-erasure honored across all stores). PDPA, GDPR, EU AI Act risk-tier mapping, NYC LL144 §20-871 disclosure all align to this default.
Long memory
Filesystem + Postgres parallel record; frozen prompts + rubrics make every replay deterministic. The regulator sees what you saw the day the decision shipped.