Tsurai7 Spec fidelity review
This lens never reads the code for quality — it reads it for conformance. Code that follows every standard but implements the wrong thing passes every other lens and fails this one.
Sources
Assemble the requirement set first, citing where each item comes from: ticket acceptance criteria, PDR/ADR decisions, product requirement docs, and claims in the PR description. If none exist, stop and report "no spec available" — this lens without a spec is a hallucination generator.
Method
- Forward pass — requirement → evidence. For each requirement: implemented? where (file:line)? Quote the implementing code. Missing or partial = finding.
- Reverse pass — behavior → authorization. For each behavior change in the diff: which requirement asks for it? Unauthorized behavior = scope creep finding.
- Quantifier pass. The classic silent violation is a changed cardinality, scope, or actor, not a missing feature: "once per attempt" becoming "once per transaction", "only on no-match" becoming "on every check", per-user becoming per-account. Compare the spec's quantifiers (per what? when? who?) against what the code actually keys on.
- Staleness pass. Spec statements the diff invalidates (a PDR describing an architecture this change replaces, a doc describing a removed call site) are findings too — the spec is part of the deliverable.
- Quote the spec line for every finding. A conformance finding without the requirement text it violates is an opinion.
Rules
- Run in a fresh context with only the spec sources and the diff — not the author's reasoning, not other lenses' output.
- Never rerank against other lenses' findings; conformance and quality are deliberately separate verdicts.
- Findings go to the findings-ledger like any other review output.
Provenance
Lens separation motivated by perspective-based reading (Basili et al., NASA SEL: distinct perspectives find near-disjoint defect sets) and validated locally: in an internal retrospective (2026-08) the code/platform reviewer and the requirements reviewer overlapped on almost nothing, and the two costliest late findings — a billed check whose contract said once per attempt silently becoming once per transaction, and an alert firing on every check where the requirement said only on mismatches — are exactly passes 3 and 1. Mechanics adapted from the two-axis review in mattpocock/skills (parallel sub-agents, never merge across axes, quote the spec line).