workflow-fix: teacher-forced fixed +/- margin is the validated continuous DV; logp_pos_mean is selection-confounded

Overview / Motivation

Auto-filed (user-directed 2026-06-30) from the #722 continuous-DV thread. The dual-DV measurement rule's PREFERRED continuous companion is the selection-confounded one; the validated one is mislabeled 'opt-in alternative'.

Goal

Correct the dual-DV continuous-companion guidance: make the teacher-forced fixed positive-vs-negative completion margin (fixed answer set across all contexts -> no selection bias) the PREFERRED non-saturating continuous DV, and flag the judged-positive-conditional-mean log P (logp_pos_mean) as selection-on-outcome confounded that MUST pass rho(DV,rate)>0 validation before any use (it fails for most behaviors); keep on-policy rate PRIMARY and the validation gate.

The finding (#722, user-provided + artifact-confirmed)

• Attempt 1 — logp_pos_mean (the stored 'secondary' DV) FAILED. It is a length-normalized base-model log-P averaged over ONLY judged-positive completions -> a conditional mean over an outcome-selected subset -> rho(DV, rate) ~= -0.3 for 3 of 4 behaviors (NEGATIVE; fails the rule's own rho(DV,rate)>0 validation). Root cause: selection-on-outcome (contexts with more positives have a lower mean log-P among them). Flagged validated:false in eval_results/issue_722/structural/result4b_continuous_dv.json.
• Attempt 2 — teacher-forced fixed +/- margin WORKS. margin(C) = mean LN-logP(FIXED positive answers | C) - mean LN-logP(FIXED negative answers | C), using #661's judge-filtered fixed pos/neg pools (SAME answer set across all 50 contexts -> NO selection bias). Validation rho(DV,rate): refusal +0.56 [+0.14,+0.80], sycophancy +0.40 [-0.03,+0.68], broad_em +0.31 [-0.20,+0.67] — all POSITIVE (sign reversed vs logp_pos_mean). Non-saturating: broad_em margin std 0.31 (rate std was 0.008). Rescues the floored behaviors; mediated (M.c_C) > direct (v_A) preserved. Artifacts: eval_results/issue_722/tf_margin/{margins,margin_chain}.json (margins also on HF). Caveat: harmful_compliance untestable (no #661 fixed +/- pool — same reason it has no r_B).

Scope / surfaces

• CLAUDE.md Measurement-validity bullet: flip the '(preferred)' from the judged-positive-completions log-P to the teacher-forced FIXED +/- margin; add the selection-on-outcome-bias warning on the conditional-mean-over-positives DV; KEEP the rho(DV,rate)>0 validation gate + on-policy rate PRIMARY.
• .claude/rules/marker-leakage-measurement.md (continuous-secondary recipe) + .claude/rules/llm-judging.md (continuous-DV guidance): same correction; define the fixed +/- margin recipe (fixed answer set, LN-logP, teacher-forced) + the selection-bias failure mode of conditional-mean-over-positives.
• Keep the enforcers consistent: planner.md §6, critic.md Stats lens, analyzer.md, interpretation-critic.md Lens 1, clean-result-critic.md Lens 7.

Constraints / invariants

• Does NOT relax: on-policy judged RATE stays the PRIMARY validated behavioral construct; the margin is the validated SECONDARY companion; the #432->#456 teacher-forcing caution still bars teacher-forced reads as a cross-condition BEHAVIORAL leaderboard (the margin is a secondary companion VALIDATED to track the rate, not the primary behavioral read).
• ruff + workflow_lint pass; the rule files + CLAUDE.md + enforcers stay mutually consistent.

Provenance

• workflow_fix_target: CLAUDE.md, .claude/rules/marker-leakage-measurement.md, .claude/rules/llm-judging.md
• fingerprint: d73c45666534

User-directed 2026-06-30 ("can we incorporate this"). Related: #722 (the finding), #763/#742/#658 (the predictor line that uses the continuous DV).