AI audit logging (weekly) - model-training log capture¶
| Stage | Estimated |
|---|---|
| Status | Deferred |
| Design status | In progress |
| Estimate | 1w (M) |
| Confidence | Medium |
| Linear | PIN-836 ↗ PIN-117 ↗ |
| Linear status | Done |
| Design | Design ↗ |
| Linear epic | Platform Foundation |
| Module | Settings & Admin ↗ |
Status check: deferred, but shipped work exists
This surface is marked deferred, yet Linear shows PIN-836 and PIN-117 as Done. Confirm whether it actually shipped (move the status off deferred) or only the remaining scope is deferred. Flagged for reconciliation.
Priority: High
Scope¶
Pre-launch review confirming AI calls (Cara, Smart Import, estimate/scope extraction, RFI/SOV drafting) are safe for production from a data-governance standpoint, backed by an audit-log interceptor that captures every AI request. Verifies the provider's no-training / zero-retention terms and reviews what the corrections-learning loop persists. (Source: PIN-836, PIN-117.)
Acceptance criteria¶
- Every AI call site logged with actor, timestamp, and request/response metadata, traceable to user + entity
- Provider data-handling posture confirmed: no-training / zero-retention terms in force for the production API key
- Field-level change diffs + business context (bid/project/estimate/org) captured on each audit entry
- Entries tagged by source (user action vs system vs integration) so training-data feeds can be isolated
- Learning-from-corrections loop reviewed for what's persisted and whether it's acceptable for pilot
- Admin-only bulk export of the audit log for compliance/training-data extraction
- Written go/no-go verdict required before AI features go live in production
Conor's comments¶
Ivan & Jonny question. Not part of design
Notes¶
INFRA / NO DESIGN - not a Claude Design surface; build item, not awaiting design sign-off. Meeting action item; AI model training depends on it
Open question for Conor¶
Confirm robust AI logs before go-live (pre-prod blocker)