Check repository guidance. Keep the evidence.

Capture doctrine with PRISM, compare a fresh observation with an accepted baseline, and inspect the result. Ordinary code performs these checks; no model call is needed.

Start with the standalone tool

Use a build containing the captured-audit and comparison interfaces on prism-harness main. Check installed help: an older package may not contain them. Captured auditing currently requires Linux /proc/self/fd support.

Keep evidence outside the audited project, in a private directory you created. Replace these absolute example paths with your own:

prism-harness audit describe --json
prism-harness audit run /work/project --json > /work/evidence/baseline.json
prism-harness audit replay --snapshot /work/evidence/baseline.json --json

Replay evaluates the saved observation. A new audit run captures current files. Source and bundled builds have different implementation identities; use a compatible implementation throughout.

Follow capture → compare → regression → restoration in the source-versioned tutorial →

Read coverage before trusting a verdict

Inspect execution, coverage and verdict together. Complete evidence with no added errors can compare as no_regression; existing accepted errors and warnings may remain. Added errors produce regression. Incomplete or incompatible evidence is indeterminate.

Depth limits, unreadable files and links outside admitted roots are explicit uncertainty. A passing selected doctrine profile does not certify the entire repository. Captured artifacts contain repository text; keep them private.

Audit and comparison contract: schemas, bounds, coverage and exit codes →

Add JARVIS for a repeatable gate

PRISM supplies repository intelligence. JARVIS admits execution under policy, stores durable receipts, and runs the existing four-step FlowIR recipe: capture → compare → assert → halt.

mnemos prism doctrine-gate --project /work/project --audit-root /work \
  --policy /work/audit-policy.yaml --home /work/jarvis-state --json

First configure the executable-specific policy and pin a complete baseline using the JARVIS gate tutorial. Each invocation creates fresh observation IDs; successful checks never silently promote a baseline. Replacement requires its current revision.

The gate initially ships on JARVIS unstable from beta.181. A source merge is not proof of availability in an older installer. Starting again after interruption creates a new observation; this version does not provide gate resume.