cross-llm-pair-review

# Cross-LLM Pair-Review Skill ## §1. Why pair-rail exists Single-LLM review has a systematic blind spot: when the reviewer and the model that produced the code share architectural / training-set priors, they share the same blind spots. Cross-LLM disagreement is a detection signal: when Codex (a different model family with different priors) flags an issue Claude missed, the disagreement itself is diagnostic regardless of which side is correct. PLAN-075 R1 motivated this by measuring a 7.3pp gap between Codex-only review and Claude-only review on the same locked-corpus fixtures — neither LLM caught a strict superset; their misses were disjoint. Pair-Rail closes that gap by requiring both verdicts before proceeding on L2+ tasks (multi-file changes, security-critical edits, new ADRs). The architecture is "asymmetric" rather than "redundant": Claude Opus retains VETO authority (Case C — see §3), and Codex peer-blocks only with rubric-violation-ID preconditions (Case B). See ADR-107 (mandatory L2+) and ADR-108 (cross-LLM VETO floor + asymmetric matrix) for the binding governance decisions; ADR-108 is the **authoritative source of truth** for Cases A-F semantics. ## §2. When to dispatch pair-rail | Task tier | Coder | Reviewer | Notes | |---|---|---|---| | L0 (trivial change, <10 LoC, no governance) | single LLM (Claude) | none | overhead not justified | | L1 (single-file, <100 LoC, low-risk) | single LLM | single LLM (different archetype) | optional pair-rail | | **L2+ (multi...