rust-doctor: Skill for Claude Code

Adjudicate one rule

Admission asks whether a rule fires on the pattern it claims. This asks the other question, how often it is wrong on healthy code, and it is the question the published rate answers. adjudication.sampling states plainly that five sites resolve steps of twenty points and cannot place a rule against the five percent threshold, so deepening a sample is what turns a published number into a defensible one.

One word governs the whole run. A verdict reached without opening the code is hearsay: inferred from a neighbouring site, from the rule's reputation, or from the tool's own claim. Hearsay is not a weaker verdict, it is not a verdict, and it never enters the measurement. Everything below exists to keep it out.

Read the "The pinned corpus" and "Admitting a rule" sections of docs/catalog-and-corpus.md first, and the adjudication.criterion field of tests/corpus.json, which is the definition every verdict is measured against. Verify the suite is green before starting, so a failure during the run belongs to this work.

Nothing below asks for a figure to be computed by hand. Every statistic this protocol publishes is recomputed by the suite from the record you write: what this skill has to get right is the record, not the arithmetic over it.

What the harness already enforces

Six invariants hold whatever you do. They catch a corrupted sample, never hearsay, which reads exactly like a real verdict and is yours to keep out:

• The rate is derived, never written. precision() in tests/support/corpus.rs computes every published rate from adjudication.reviewed alone, and a test asserts the published field equals the recomputation.
• The interval and the separation are derived from the same sample, so a rate that settles nothing against the threshold says so rather than reading like a measurement.
• The agreement of the two passes is derived from adjudication.agreement.pairs alone: every row of adjudication.agreement.coefficients is recomputed by the suite, and a row edited by hand fails naming the rule and the field.
• A reviewed site must name a finding the corpus actually produced, at its published path and line, and in its published context. Both are checked in the_published_observations_reproduce_the_pinned_corpus_run.
• (rule, repository, path, line) is unique. A site reviewed twice makes the rule incomplete and withholds its rate.
• MINIMUM_REVIEWED_SITES is a floor of five, not a ceiling. The only upper bound is reviewed <= findings.