Skip to content

Evidence and acceptance

Each result has both an acceptance status and evidence validity. Policy requirements are frozen. Evidence binds the candidate, policy, checker, dependencies, task, issue time, and optional expiry.

Record Execution and decision
typecheck Whether OCI output matches the return type; an automatic requirement in every policy
OutputCheck Executes candidates on frozen JSON inputs and compares against host-held expectations
RepositoryCheck Applies a typed Patch in fresh OCI, then checks fixed command exit codes and output
model_review A separate model request returns a structured boolean and detail
ApprovalGrant Independent user authorization to write the exact proposal

A passing check covers its input and requirement. When a candidate changes repository tests, the controller's acceptance oracle still uses its original configuration. Model reviews and deterministic checks have separate accounting and records; interfaces should show both results and evidence state.

Validity and updates

Revoked, expired, future-dated, dependency-incomplete evidence, or evidence bound to a different candidate/policy/input, becomes invalid. A newer failure in the same candidate, policy, and dependency context prevents an older success from satisfying the requirement. Different inputs retain their own contexts. Mutable model aliases require review in a new task; human decisions are task-bound.

Artifact admission, locked replay, patch authorization, and application revalidate through the runtime. Objects serialize source spans but exclude them from semantic digests. Source content has a separate hash; clients retain both identities.

uv run nexa open evidence PROPOSAL_ID --environment environment.json --state .nexa/program.db

Use the PROPOSAL_ID returned by the real run. Author evidence/explain instead accepts an evidence object ID and optional context. Preserve requirement names, failure details, object digests, and context when diagnosing a failure. Natural-language completion becomes a usable result through formal policy admission.


Source and validation · open_evidence