- Rename JUDGE_NAMES from [Judge-GPT, Judge-Claude, Judge-Gemini] to
[Judge A, Judge B, Judge C] everywhere (orchestrator + tests + ad-hoc
A/B script). Removes the Claude-family naming anchor that was
repeatedly pulling the panel toward Claude models even when Student
plan blocked them — the slot is just "one of three judges", not
"the Claude judge".
- Apply the A/B-test-winning headless override (EXECUTION_CONTEXT_BLOCK)
to both _build_meta_judge_chat_payload and _build_judge_chat_payload.
Phrases are verbatim from OpenAI GPT-5 Prompting Guide + Anthropic
Claude headless docs. Tells agents there's no human to answer
clarifying questions; to commit to best-default interpretation and
document the assumption.
End-to-end pipeline run on WYL-77 (2026-04-20) confirmed:
- Meta-judge produced rubric without clarifying question (override
working on ambiguous Tierra prompt)
- 3 judges ran in parallel without SQLite contention (XDG fix held)
- Consensus math gracefully excluded Judge B's malformed YAML
- Verdict REJECT avg=1.00 on prose-only worker delivery
- Retrigger fired; worker honored anchor + blocked path
Known blockers surfaced (not fixed in this commit):
- Shared HOME allows cross-daemon workdir snooping (Judge B read AI
Engineer's task workspace directly). User flagged as feature for
now; revisit if it causes drift.
- gpt-5.4-mini (Judge B) produced malformed YAML on this run; n=1,
can't distinguish chance vs consistent inability — need multi-run
baseline to decide.
- REJECT on "no commit URL" conflates missing-delivery with
bad-work; pipeline signal is correct but reasoning upstream of
judges is unclear.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
