workflow-fix: llm-judging.md judge max_tokens sizing rule (truncation censoring mimics selection artifact)

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a prose follow-up raised on task #1090 (emitting agent: experiment-implementer, Step 9a-ter free-analysis round).

Goal

Add a judge-response max_tokens sizing guideline to .claude/rules/llm-judging.md: reason-then-score rubrics (rule 7) need a response-token budget sized for the reasoning; a 64-token cap silently truncates reason-first judge responses before the score token, and the drop-never-coerce rule (rule 9) then discards them — censoring 10-40% of draws with an arm-asymmetric pattern that mimics a selection artifact.

Workflow gap

• Bug observed: #1090's Tier-2 sycophancy judging dropped 473/1000 base vs 307/1000 trained draws as parse errors; a refresh at max_tokens=300 recovered 98.8%, proving the drops were 64-token truncation of reason-first responses (0 refusals). The asymmetric censoring initially read as a possible selection artifact on the headline install delta.
• Why it is a workflow gap: llm-judging.md rules 7 (reason-then-score) and 9 (drop-never-coerce) interact to censor draws when the caller's max_tokens is too small, and no rule warns about sizing the response budget; the trap plausibly recurs at every graded-judge call site with a reasoning rubric.
• Confidence (emitter): high

Proposed change (candidate diff sketch — refine in planning)

• New guideline in §C (prompt/rubric design) or §B: "Size judge max_tokens for
• the rationale, not the score: a reason-then-score rubric needs >=~300 response
• tokens; a truncated response parses as an error and rule 9 drops it —
• arm-asymmetric truncation censoring mimics a selection artifact (incident
• #1090: 47%/31% draw drops at a 64-token cap, 98.8% recovery at 300). Report
• the per-arm drop rate alongside every judged DV (extends rule 18)."
• Cross-ref: src/explore_persona_space/eval/graded_judge.py judge_graded now
• threads a max_tokens kwarg (default 64 for cache-stability; callers with
• reasoning rubrics pass ~300).

Scope / surfaces

• Primary target: .claude/rules/llm-judging.md
• Grep max_tokens in .claude/rules/*.md + docs/api_throughput_guidelines.md for sibling hits; update consistently.

Constraints / invariants

• Workflow-surface only; workflow_lint passes.

Provenance

• workflow_fix_target: .claude/rules/llm-judging.md
• fingerprint: (compute at filing)

Verbatim prose follow-up: "c-cells judged at 64 tokens systematically lose ~10% (formatting) to ~40% (sycophancy) of draws — future judge_graded callers with reason-heavy rubrics should pass a larger max_tokens."