Status: prototype / public verifier available.
The publicly checkable part of the compliance-ledger work is the signed AI-response record. It binds the recorded model, input, and response as signed commitments and includes a tampered negative control. It does not establish the broader policy-ledger functions described below.
What is demonstrated
- Offline verification using only the package and standard Python dependencies.
- A valid record and a deliberately tampered record with a named refusal reason.
- Signed commitments for the recorded model, input, and response.
- Explicit limits: no model-execution attestation, continuous-presence claim, firmware appraisal, revocation check, or performance claim.
This is a control an auditor can evaluate and rely on within that stated scope. It is not a replay of model execution and does not establish approval or compliance for the surrounding deployment.
Broader ledger design targets
The following are design targets, not represented here as implemented production functions:
- Policy context. Bind the policy state and approval context applicable to an interaction.
- Layered authorship. Authenticate outputs against both an originating compute environment and a supervisory control layer.
- Risk-aware fusion. Combine statistical signals into an index that can drive escalation policy.
- Aggregate verification. Batch evidence with predictable verification cost.
- Privacy-preserving external verification. Cover third-party calls without exposing sensitive payloads.