Evidence-based AI code review

The agent's verdict is not the acceptance criterion.

AIWorkHub carries code changes and their evidence to a verified manager, then recomputes the acceptance decision at the canonical boundary.

Evidence travels with the task

A review packet can include changed paths, hashes, candidate diffs, validation commands and results, logs, artifacts, tool-use receipts and approval history. Missing evidence remains visible instead of being replaced by a confident summary.

Scope and output validation

Allowed writes and required outputs are checked before launch and again during review. Required outputs are repository-relative paths or globs; behavioral expectations belong in explicit acceptance criteria.

Independent verification

Workers stop at review_ready. The current repository manager independently checks the retained workspace. Medium and higher-risk changes can require stronger reviewer and combined-tree validation.

Precise rework

A rejection records exact feedback and an explicit disposition: requeue, block, archive or supersede. Residual work preserves the reason for failure rather than starting another ungrounded attempt.