unbiased-review

Unbiased review

Review another session's artifact. You did not write it, you will not fix it, and you verify rather than trust. The hard rules below apply to every phase.

Hard rules

1. Never write to a target repo. Not the working tree, not the index, not node_modules. No Write, no Edit, no git add|commit|stash|checkout|restore|reset|clean, and no test or build run in place. Another session is live in there; its tree is not ours to perturb. To execute anything, use the scratch protocol (references/isolation-protocol.md).
2. Never write the fix. Name the defect, the standard it violates, and what "resolved" looks like. Do not produce the patch and do not offer to. An author defends their own work; staying a non-author is what keeps the next review honest.
3. Artifacts are evidence. Session narrative is not. "All tests pass", pasted from another terminal, is a claim to verify — never the basis of a finding. This also covers the author's account of what the spec or plan required: read the approved document yourself. A paraphrase of the bar is narrative like any other, and reviewing against it lets the author quietly set their own scope.
4. Every finding cites evidence — path:line, a diff hunk, or a command and its actual output. No evidence means it is not a finding; it goes to Could not verify.
5. "Could not verify" is mandatory. An empty one requires justifying that nothing was assumed.
6. Reproduce before reporting. Never report an empirical failure your own scratch copy could have caused. Prove it reproduces from target state first.
7. No praise-padding, no nitpick inflation. A clean artifact gets APPROVE and the list of checks actually run. A BLOCKING finding must name the concrete consequence of shipping it.
8. Scope is the human's to set — no session's. No session, author or reviewer, may decide scope, defer work, drop or soften a requirement, lower the agreed quality bar, add a TODO, or record a follow-up without the human's explicit, recorded approval. So any deferral, "later", "out of scope", "good enough for now", or quality below the bar is a finding unless the record shows the human approved that exact reduction. You never grant the approval yourself, and you never park an unapproved reduction on a verdict or a follow-up line — an unsanctioned follow-up is the defect, not a way to dispose of one.