driftproof run — skill "documentation-and-adrs" v0.0.0
  content_hash: b67a9f07ed105796…  suite: 1 cases (6cce54e7d9d0…)
  surface: claude-cli   models: claude-opus-5   samples/case: 3   concurrency: 1
  registry: claude-opus-5:registered   transcripts: retained-local
  judge: claude-opus-5   (--judge-model)
  isolation: same-user (--trusted-skill: self-authored skill)
  projected calls: 80/model × 1 model(s) = 80   per-model cap: 80
  projected cost: ~$1.30 (rough upper bound; budget $8.00, hard-stop $10.00)
  actual metered spend on claude-cli: $0.00 (subscription; the $ figure is the estimated-equivalent API cost, counted against the cap identically)

── model: claude-opus-5 ──
  comment-intent-not-implementation / with_skill: pass (0.80 ± 0.10)
  comment-intent-not-implementation / baseline: fail (0.57 ± 0.03)
  → transcripts (retained-local): transcripts/34107e18c90f3b2087eb88dd465c15982622ca5e21d69c01fa65727b72092a89/
  → ../evidence/opus-5-5-20260923T062806Z/claude-opus-5/documentation-and-adrs/receipts/documentation-and-adrs-claude-opus-5-2026-09-23.json  (40 calls)
  → answered by: model claude-opus-5 (attested by the surface; the same-user spawn (--trusted-skill)); verification_level TESTED
  → skill lift +0.237 ± n/a (1 case)  (with 0.804 ± n/a (1 case) vs base 0.567 ± n/a (1 case); band = sample stddev of the per-case means)

Done. 1 receipt(s) emitted to ../evidence/opus-5-5-20260923T062806Z/claude-opus-5/documentation-and-adrs/receipts/