Acceptance scenarios

A compatible implementation should pass these behavior scenarios.

Claim grounding

Given an answer with three factual claims, the evidence pack records three claim ids, links two to supporting sources, and marks the third unverified with a missing-source reason.

Contradiction

Given two selected sources disagree, the claim status becomes contradicted or partially_supported, and the counter-evidence edge is retained.

Tool provenance

Given an answer derived from a tool call, the pack links the claim to the tool result ref, the tool call id, and the runtime span id when available.

Retrieval omission

Given a source was retrieved but rejected as stale or out of scope, the source map records an omission reason rather than deleting the source from the audit trail.

Artifact review

Given a generated artifact, the pack links artifact version, diff ref, verification checks, and review verdict without embedding full artifact bytes.

Verification vs review

Given a schema check passes but a human reviewer requests changes, the pack records passed verification and needs_changes review without treating either as authoritative over the other.

Redacted export

Given private source text, the exported pack replaces snippets with redacted refs, keeps claim ids and source ids stable, and marks verification coverage as partial.

Replay honesty

Given a non-deterministic model response and expired external API result, the replay case marks model output as approximate and API output as unavailable.

Telemetry absence

Given no trace backend was connected, the pack marks telemetry as not_collected instead of inferring that no runtime errors happened.

Peer handoff

Given a peer agent returns an artifact and message id, evidence preserves native peer ids and links them to local claims without rewriting the peer records.

Agent Runtime profile evidence

Given a runtime turn exports evidence, the evidence pack preserves runtime correlation ids, replay and review use the same source facts, and missing telemetry is marked not_collected or unavailable rather than fabricated as evidence. See Runtime profile test cases.