EPS
← All tasks·#702Completed

analyzer: confirm all planned control arms present before authoring a verdict

kind: infra#daily-held
track:

Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.

No comments yet.

Overview / Motivation

Filed from /daily 2026-06-27 held backlog: the analyzer authored a verdict on a partial grid missing a control/baseline arm.

Goal

Prevent the analyzer authoring takeaways on a partial grid that is missing a planned control / baseline arm.

Problem (from /daily 2026-06-27)

#658 near-premature "complete": the apparent 3/10 Betley PASS was a multiple-comparison artifact beaten by a random-projection control (0/10 after the control landed), but the analyzer ran on a partial grid before the control arm landed (session 3149da48). Authoring a verdict before the control arm exists nearly shipped a false positive.

Proposed change

analyzer.md guard — before authoring takeaways, assert that every planned control / baseline arm is present in eval_results/. A missing control arm BLOCKS the verdict: park (don't author), surface the missing arm, wait for it to land. This is the planned-vs-actual-coverage discipline applied at verdict time, not just at clean-result-critic time.

Scope / target files

  • .claude/agents/analyzer.md

Constraints

  • Workflow-surface only — no experiment code, configs, or tasks/.
  • Lint gate green (workflow_lint.py).
  • Keep consistent with the existing planned-vs-actual coverage rules (verify_task_body.py check 11b, clean-result-critic planned-vs-actual lens) — this is the upstream analyzer-side guard.
Activity