lie-detector

Treat changes as unproven; steelman "this should not merge". Standalone or inline lie-detector hat in /review on every diff; prioritize customer-facing surfaces. Defects: /review; value: /product. Pin scope, including captured WIP.

1. Claims without evidence

Require diff lines, command output or primary docs for claims in PRs, commits, code, docs and summaries:

• Phantom surface: symbol/prop/hook/flag/config/env/token/route/option absent at the installed version. Check lockfile, node_modules types or primary docs, not memory.
• "Tested", "verified", "fixes X", "no behavior change": needs the command and output or a red-then-green pair.
• Citations that do not support their claims.

2. Code that pretends

Trace changed paths to caller-visible output; flag unperformed work:

• success before the work completes: a toast, redirect, or 200 before the mutation, write, or invalidation resolves;
• errors that swallow, log-only, or fall back to defaults, placeholders, or cached data while the UI shows success;
• stubs, TODOs, hardcoded or mock data, or disabled branches shipped as done;
• a flag, option, or config key that no code path reads;
• names, comments, types, or casts that disagree with the code they describe;
• loading, empty, or error states that cannot be reached or never end.

3. Tests that cannot fail

For each added or changed test:

1. In a disposable snapshot per review isolation, break the claimed public behavior, not just wiring. Rerun; leave originals/index/branch untouched. Still green: the test lies. Only assertion failure counts as RED; record command/failure. Missing isolation is a named gap, not permission to mutate the active checkout.
2. Check survivors against the test-audit junk patterns.
3. A behavior change with no test that went red and no real-entrypoint replay (/dogfood) is unproven.

4. Changes nobody asked for

Flag unexplained out-of-scope behavior changes or weakened guards: deleted, skipped, or loosened tests; regenerated snapshots; @ts-ignore or lint suppressions; changed defaults, copy, routes, flags, error handling, retries, or timeouts; widened exclusions, relaxed compiler/lint rules, or shell/CI failure exits turned into success; lockfile or config churn; hand-edited generated files; a visual change without risk-appropriate rendered verification; screenshots are optional per proportionate verification. Review changed checks/configs separately from the implementation they judge; a justified rule correction needs known-bad and valid cases through its real command.

Trace auth-proof creation: casts, unchecked constructors and comments cannot authorize. Green types do not prove a trusted module sound. For assertions/type evidence/rules, use anti-slop; taste is not a defect.

5. Patterns that will spread

Compare introduced idioms with precedent (rg), root/scoped CLAUDE.md/AGENTS.md, exemplars/ and stack registry. Contradicting policy/dominant idiom is a finding. Copy reach: shared components/hooks/fixtures/helpers/templates/generators/exemplars/skills = P1; leaf = P2. Fix the source pattern or update its rule in the PR with rationale; do not leave competing styles.

6. Steelman gate

Follow steelman/SKILL.md on the premise "this PR should merge". Argue the strongest case against it from the product verdict, the findings above, and /review findings. Merge-ready only when all three hold:

• value: the product: verdict is justified;
• truth: no P1 here, and every changed behavior has a red-then-green test or a real-entrypoint replay;
• implementation: /review reports no P0/P1.

Otherwise not proven; name the missing proof. Drop counterarguments without file:line or command output. Never block on preference.

Severity

• P1: a claim the code or docs contradict; a phantom API; code that reports success it did not achieve; a test green on broken behavior; an unexplained behavior change or weakened guard; an anti-pattern in a copyable place.
• P2: an unverified but plausible claim; a misleading name or comment; a leaf anti-pattern; a tautological assertion beside real ones.
• P3: summary only.

Report at most five findings, merge-deciding first.

Output

Lead with one verdict line: lie-detector: <truthful|suspect|lying> -- merge <ready|not proven> -- <evidence, at most 20 words>

truthful: no findings. suspect: P2 only. lying: any P1. Then [P1|P2|P3] <file:line> <section> -- <lie and proof>; <smallest fix>. Clean pass: verdict only.