Evidence Verification Pipeline

Alias entry point for /ev-pipeline. It operates on the evidence/ data lake only and does not mutate registry files.

Type-First Evidence Lake Contract (#1148)

The evidence lake is type-first. The primary working set for every phase is evidence/by-type/<canonical-evidence-type>.md. Legacy evidence/tier_*.md files may still exist as coexistence artifacts for older tooling, but they are not the semantic routing key.

Phase Responsibilities

• Phase 0 — ev-discovery (skippable): searches for new Stage-2 evidence on declared need and appends discovered rows into source inputs for Phase 1.
• Phase 1 — ev-collection: materializes/compiles by type. evidence/by-type/<type>.md is primary; tier_*.md is coexistence-only.
• Phase 2 — ev-star-verification: verifies live GitHub stars without repartitioning by rank.
• Phase 2B — ev-benchmark-verification: verifies benchmark-result rows and candidate manifests against registry/benchmark-sources.json; unknown/candidate/registered/mirrored/incomplete sources do not score. Registered sources may declare appliesToGenericSkillRefs for generic capability applicability, but that metadata is non-scoring until a named benchmark-result row satisfies reproducibility and approved provenance.
• Phase 3 — ev-adversarial-audit: audits the by-type files, including benchmark catalog misuse and vendor-claim leakage.
• Phase 4 — ev-link-validation: validates URLs. validate_sources.py remains a temporary coexistence URL-health shim.

Generated evidence outputs are review artifacts. Do not commit generated by-type files, tier files, source reports, validation reports, unified lake files, seeds, collector rows, or benchmark candidate manifests without a human gate. After Phase 2B + Phase 3 + Phase 4, humans must approve any benchmark catalog promotion or /gaia-ingest-batch; machines classify, humans promote. A registered benchmark source is generic-applicable catalog metadata only: no named-skill score, Trust Magnitude contribution, or scoring evidence row exists until a named benchmark-result row satisfies reproducibility and approved provenance.