Issues

• Review/Investigation: the sufficiency score is monotone, so an over-specified spec scores higher

  enhancement
  benjaminjones/spec-eval#50 · Filter by author benjaminjones opened on Aug 21, 2026

• Watch the empty-report rate under rationale masking — one run in five returned no findings

  benjaminjones/spec-eval#48 · Filter by author benjaminjones opened on Aug 21, 2026

• Usage telemetry from the claude-code bridge is not trustworthy, so cost cannot be estimated from it

  benjaminjones/spec-eval#45 · Filter by author benjaminjones opened on Aug 21, 2026

• A single audit run's headline drift count is not reproducible

  benjaminjones/spec-eval#44 · Filter by author benjaminjones opened on Aug 21, 2026
  1

• stated-elsewhere withdrawals are unchecked, so the most-used ground can be satisfied by a quote that is not in the document

  benjaminjones/spec-eval#39 · Filter by author benjaminjones opened on Aug 20, 2026

• Review/Investigation: Enforce fresh-context checking by shipping spec-check as a subagent

  benjaminjones/spec-eval#34 · Filter by author benjaminjones opened on Aug 15, 2026