Confidence tiers

Module: src/analyze/classify.ts.

classify turns reachability results into findings, assigning each non-alive node exactly one tier and deciding auto-fix eligibility.

function classify(input: {
  nodes: SymbolNode[];
  reachability: ReachabilityResult[];
  publicApiIds?: Set<string>;
}): ClassifiedFinding[];

The decision

For a dead node:

tainted OR public API   → maybe   (autoFixEligible: false)
exported                 → likely  (autoFixEligible: false)
otherwise (private)      → certain (autoFixEligible: true)

A test-only node becomes a test-only verdict (tier maybe, autoFixEligible: false) — report-only by locked decision. Alive nodes are not findings.

Evidence is attached here

Crucially, classify attaches the evidence signals it evaluated to each finding. The evidence reporter only formats them — it never re-derives them. This guarantees the rendered evidence can’t drift from the verdict.

interface ClassifiedFinding {
  node: SymbolNode;
  verdict: "dead" | "test-only";
  tier: "certain" | "likely" | "maybe";
  autoFixEligible: boolean;
  evidence: EvidenceSignal[];
}

Coverage

The certain tier is meant to be strengthened by a coverage-miss signal when a report is available. Coverage ingestion is planned, so today every chain shows • coverage: not available and tiers rest on static signals.