[brianorwhatever]

REV rejected a correct function and matching execution observations because its prompts mixed whole-requirement completeness with per-item relevance. Each verdict now explicitly targets one criterion; source relationships use a separate policy that assesses contribution without demanding that every item prove the whole requirement.

The model, thresholds, evidence-class requirements, contradiction checks, and fail-closed aggregation remain unchanged. The evaluation suite now includes 14 executed cases across five domains, and its export fixture describes the actual function calls rather than an HTTP server it does not implement.

Validation: 33 local tests and syntax checks passed. Live evaluation matched all 20 criterion labels across 14 cases (five correct implementations approved; nine broken or unverifiable cases blocked). The unchanged evidence from TV PRs #23–#25 passed 20 of 23 criteria; three remain unverified and were not used for further prompt tuning. Results, limitations, exact evaluator commits, and run links are recorded in docs/evaluation.md and eval/results/2026-10-24-scoped.json.

Note

Fix conflicting evaluator scopes and add live regression challenge fixtures

• Splits evaluation policies: whole-criterion verdicts in policy.INSTRUCTIONS now target one criterion only, and a new policy.RELATION_INSTRUCTIONS judges a single evidence item against one criterion in policy.js
• gate.requestBody now scopes criterion questions to an explicit criterion state and evidence questions to a specific evidence item path with its own relation policy in gate.js
• Adds four challenge domains (numeric clamping, owner authorization, handler priority, retryable download) with one correct and one defective implementation each, executed live by challengeFixtures in challenge-fixtures.js
• eval.run now processes the six original fixtures plus eight challenge cases and still fails on any mismatch or error in run.js
• Original fixtures now check direct route(session) outputs instead of HTTP endpoint behavior in fixtures.js
• Behavioral Change: eval.task criteria no longer mention HTTP endpoints; requests now carry criterion and evidence target paths, verified in gate.test.js

^{Macroscope summarized ca0fb4d.}