review-deeplisted

# Review Deep > **Judging doctrine:** the producer-grader split, evidence-on-every-verdict, deterministic-aggregation, and primary-owner dedup rules this skill runs on live in [`_shared/judge-core.md`](../_shared/judge-core.md) — this skill is one of its reference implementations (§10). It instantiates the doctrine for the code-review domain. The full-depth engine behind `review-gauntlet`'s lean profile. Runs six standing review lenses (correctness, bugs, **security**, test quality, style, plan-conformance), with explicit Tier-1 anti-pattern detection, per-lens model-tier selection, severity tiers + cited evidence on every finding, an auth-gated runtime downgrade, a scope-boundary `DEFERRED-TO-UAT` verdict for things only an operator can confirm, and a JSON sidecar that captures the full audit trail. `review-gauntlet` is a lean profile over this engine — use this skill when the diff is high-stakes (substrate changes, schema/key-shape changes, producer-consumer chains, anything where a silent miss is expensive) and the leaner gauntlet profile when the diff is routine. --- ## Arguments | Arg | Required | Default | Description | |---|---|---|---| | `--prompt` | yes | -- | Developer intent: what the diff is supposed to accomplish (may be passed inline) | | `--diff` | yes | -- | The diff to review: PR number, `git diff` mode, or explicit paste | | `--plan-step` | no | -- | `<plan-file-path>:<step-id>` — enables the plan-conformance lens | | `--reviewers` | no | `code` | `code