ai-safe2-framework

AI SAFE² Challenge provider-feedback decision record

AI SAFE²

Framework Home CLI Challenge guide Challenge 001

Decision: preserve the challenge, improve evidence interpretation

Participant feedback supplied during release review described six TENIR kernel verdicts, including PASS for enforcement-outage instead of the reference HOLD. The participant attributed this to an admissibility vector without an enforcement availability signal and declined to retune inputs merely to match the reference. These are participant-reported observations, not an independently verified run. The source export and ledger proofs were not available for this review. No provider efficacy, partnership, certification, or endorsement is established.

The feedback exposed an interpretation risk in our system: the frozen outage fixture concerns an authorized shared write, while C7 concerns protected actions. The current state grader does not separately measure fail-closed availability. Neither a valid import nor a completed shared write establishes C7 assurance.

Closed in this release

The frozen protocol, grader, reference outcomes, contract enum, framework controls, and provider scoring formulas are unchanged. These tests are locally constructed regression probes, not imported participant results or independent replication.

Evidence intake boundaries

Request the original export bytes, actual observed state, declared policy/runtime identity, outage injection mechanism, and evidence of the execution boundary. Do not synthesize state transitions from a kernel verdict. Freeze the original run; any provider policy change creates a separately identified run.

tenir-example-v1 permits a nonsynthetic declaration if its semantics match. Use provider version metadata for implementation versions. A new translation contract requires coordinated implementation, validation, and reconstruction.

Ledger sidecars remain unverified supplementary evidence. A future verifier would need agreed byte canonicalization, leaf construction and order, hash algorithms, episode/source binding, trusted-root authentication, and adversarial tests. Merkle inclusion alone proves neither observation truth nor complete logging.

Any future provider-neutral expansion of coverage requires separate review and versioning. It is not a commitment to alter Challenge 001 for a particular tool.