ai-safe2-framework

Agent Claim Audit

safe2 evidence claims compares explicit agent outcome claims with existing SAFE² task-receipt criteria and creates both a machine artifact and human review card.

safe2 schema export claim-audit-source-v1 --output claim-source.schema.json
safe2 evidence claims claims.json task-receipt.json \
  --output claim-audit.json --card claim-audit.md --strict

Each claim names its type, asserted outcome, and the receipt criterion IDs that are supposed to support it. The audit labels the claim evidence_consistent, contradicted, or unverifiable. It separately counts explicit failed, unavailable, and not_attempted disclosures so a truthful limitation is not mistaken for incomplete work hidden behind a success claim.

The source also declares the complete receipt_input_sha256s set. This binds the audit to each supplied receipt’s own canonical criteria-input document while still allowing independently produced receipts for the same task to be combined.

There is intentionally no “honesty score.” A missing or contradictory receipt does not prove intent or deception, while a consistent unsigned receipt can still be fabricated. Coverage and contradiction ratios always retain their denominator and apply only to the supplied claim set. Task completion remains human-owned and completion_verified is always false.

For stronger evidence, precommit acceptance criteria, use authenticated collectors where available, bind receipts to the exact system identity and revision, and keep the evaluator independent from the agent making the claim.