Implementation Evidence Loop

Use the smallest useful evidence loop

For each plan step, inspect the relevant contract and existing tests, make one coherent change, run the narrowest meaningful check, inspect the diff, and then run the broader affected suite. Record observed results rather than claiming success from code appearance. Keep unrelated user changes intact.

When a check fails, classify whether the defect is in the implementation, test, environment, dependency, or assumption. Fix the cause, rerun the failing check, and retain enough output for reproduction. Do not weaken assertions or delete safeguards merely to obtain a green result.

Record each implementation increment

Evidence should connect the requested criterion to the changed artifact and to a reproducible observation. Redact sensitive output and summarize large logs while preserving the exact command, version, failure signature, and result.

Plan step and criterion: [STEP -> ACCEPTANCE CRITERION]
Baseline: [COMMIT, TEST RESULT, OR OBSERVED STATE]
Files changed: [PATH -> PURPOSE]
New dependency/file/command: [ITEM -> RATIONALE | NONE]
Narrow verification: [COMMAND -> EXIT/RESULT]
Affected-suite verification: [COMMAND -> EXIT/RESULT]
Diff inspection: [IN-SCOPE, UNRELATED CHANGES PRESERVED, SECRETS ABSENT]
Failure found: [SIGNATURE -> CLASSIFICATION -> FIX | NONE]
Safety check: [PERMISSION, DATA, AND SIDE-EFFECT RESULT]
Recovery tested or reviewed: [METHOD -> RESULT]
Remaining uncertainty: [UNKNOWN -> NEXT CHECK]
Evidence references: [LOG, TEST, REVIEW, OR CHANGE ID]

Expected evidence and verification

The expected evidence is a chain from acceptance criterion to changed files, narrow verification, affected-suite verification, diff inspection, safety check, and recovery path. A passing check without that trace is incomplete evidence.

Before handoff, reproduce the main user journey from a clean install or isolated checkout, compare the final diff with the approved plan, scan for credentials and generated noise, and list any untested behavior. Require independent review when consequence, security sensitivity, or repository policy warrants it.