Understand AI coding validation evidence in Kranz

Separate a command result from a judgment

A contract command records what ran and how it finished. A validator interprets whether the implementation and results satisfy the agreed criteria. Both matter: a test can pass while checking the wrong thing, and a reviewer’s positive summary needs supporting evidence.

For an endpoint change, follow the acceptance criterion to the route implementation, then to the test that makes the request and asserts the response. Read the actual test result. Check that the test ran and that failures would cause a nonzero exit.

Preserve uncertainty. A missing artifact, unavailable browser check, or inconclusive result needs to remain visible in your decision.

Follow repairs through to the final result

Milestones can go through validation and repair rounds. Read why the first attempt failed, what changed in response, and whether the later check covers the original finding. Earlier passing output cannot establish that a subsequently edited implementation still passes.

Kranz’s mission record includes worker and validator sessions, gate results, and decisions. Use the report as an entry point and follow the underlying evidence for the criteria that matter most.

For interface changes, see live UI validation. Its evidence should identify the interaction exercised and what the validator observed.

Understand what the merge gate adds

Milestone validation judges the work as it progresses. The human-triggered merge gate checks the proposed integration with the current base branch. Kranz runs the repository’s applicable merge commands in a scratch worktree and advances the base only to that tested integration.

This matters when another change has landed since the mission began. The repository merge suite should reflect the tests, build, and lint checks you require for the combined result.

Export a portable review record

From the mission’s repository, export its evidence bundle to a new or empty directory. Replace the example mission ID with the one you are reviewing:

kranz evidence-bundle m-123abc --out ../mission-review

The bundle can be opened without repository access. Start with summary.md, then use manifest.json to find source references and SHA-256 hashes. chain.json carries the provenance record, events.jsonl the scrubbed event log, and artefacts/ the resolvable supporting text.

Unavailable artifacts appear as unresolved manifest entries. Their absence is part of the evidence. Exported text is scrubbed for supported secret patterns, but it can still contain source code and private project information; review the bundle before sharing it.

Use the evidence to make a decision

You should be able to explain which requirement passed, what supports that conclusion, who or what checked it, and which questions remain open. An evidence bundle preserves that explanation; a passing verdict still needs to be assessed against the change you intend to accept.

Continue with the review workflow or run your first mission.