Review Swarm

Use this skill for independent review of a result, manuscript, derivation or diff. New multi-reviewer work uses scripts/bin/run_multi_task.py; a single advisory reviewer uses scripts/bin/review_one.py, which delegates to that same launcher. Resolve commands and code paths from this skill directory. The canonical public skill name is review-swarm.

Mandatory boundaries

• Match the review investment to the artifact. Result-bearing, public-surface and irreversible artifacts warrant a second seat; cross-family review is preferred, not universally required. Internal implementation work closes on an independent single-family review plus local tests, never the author's tests alone. Record honest degradation and each independence axis.
• A completed seat must be successful, contract-valid, uncontaminated and supported by exact identity evidence. Different routes or aliases do not establish different model families. Roster availability cannot substitute for completed quorum. Native reviewers stay outside launcher aggregation unless deliberately executed through its configured runner.
• Give reviewers the material and neutral question, without prior verdicts or the author's initial judgment. Sealed derivation/recomputation receives only frozen raw inputs, persists its result and terminates before the candidate-visible comparison. Tool mode, ambient context, operator instructions and actual post-fallback execution affect independence; absence of a contamination flag is no attestation.
• Acceptance requires zero BLOCKING findings, HIGH findings fixed or attached to their named acceptance point, and every LOW dispositioned. Apply the stricter grade when schemes overlap. Batch fixes and obtain independent confirmation on the current artifact; default maximum is five candidate review rounds, then narrow or escalate. Lexical Jaccard similarity is not scientific acceptance.
• A verdict applies only to its frozen inputs. Require FRESH immediately before folding; changed bytes invalidate acceptance. A malformed delivery is unfinished delivery: bounded same-model normalization first, then documented fallback or honest unavailability. The leader cannot transcribe it into a valid vote. Missing, partial, unchecked, unavailable or stale evidence never counts as acceptance.
• Static reading does not certify runtime preconditions, numerical reproduction or load-bearing structural claims. Use the applicable executing / candidate-withheld roles below and record what each check could falsify. A well-formed report or signed runtime capture proves neither provider truth nor scientific correctness.

Read only the relevant procedure

Read each applicable reference before the named operation. References retain the detailed binding contracts and executable examples.

For a single correctness reviewer, after reading the applicable scope, identity and isolation contracts:

python3 scripts/bin/review_one.py \
  --model codex/default \
  --artifact /path/to/notes.md \
  --role correctness

Omitted/default model selections use the backend's configured default; never inject a historical model name. This one-seat result remains advisory where the artifact warrants another seat.