On-policy + matched-LR isolation of the LoRA-vs-dense sycophancy bystander-leakage gap (coverage-matched)

Goal

Isolate the rank/parameterization component of #606's LoRA-vs-full-FT sycophancy bystander-leakage gap by comparing LoRA against a coverage-matched dense fine-tune (the #642 cmft module set) at a matched learning rate, trained on on-policy sycophancy completions, on on-policy judge-scored per-persona bystander leakage at matched source-implant strength — testing whether any adapter-vs-dense leakage gap survives once learning rate, module coverage, and data realism are all controlled.

Background / why

• #606: full FT leaks sycophancy ~+0.098 wider than LoRA at matched implant strength.
• #642: that split into Δ_rank = +0.073 (adapter-vs-dense, same modules) and Δ_coverage = +0.025 (module coverage, LOW). #642 explicitly flagged that Δ_rank still bundles the learning-rate difference (LoRA 1e-5 vs dense 5e-6) plus dropout / rsLoRA scaling / no-trained-bias — so +0.073 is not pure rank.
• All of #606/#642 trained on canned/templated (tier-3) completions; #612 showed canned data overstates install strength vs on-policy (canned +0.84–0.93 vs on-policy +0.60–0.66 at matched recipe).

This run removes the LR confound (matched LR) and the data-realism confound (on-policy) on top of #642's module-coverage control, for the cleanest possible read on rank/parameterization.

Design sketch (planner to formalize)

• Arms: LoRA (r=32, α=64, rsLoRA) vs coverage-matched dense FT (#642 cmft recipe) at a single shared learning rate. The planner grounds the LR value: one LR where BOTH arms have a clean sub-ceiling install trajectory with realized checkpoints landing in the matched band — the dense arm must NOT install in ~4 steps as #606's 5e-6 FT did on refusal.
• Two variables change vs #642 (matched LR + on-policy). The planner MUST keep each effect attributable — add the intermediate arm(s) needed (e.g. a matched-LR-on-canned arm and/or an on-policy-at-original-LR arm) or justify the joint change. The consistency-checker will flag an un-decomposed two-variable change.
• Data: on-policy sycophancy completions for the software_engineer source via the #612 elicitation ladder (reuse #612's on-policy pool where source/recipe match), 80%-floor + equalize-down + contrastive negatives per .claude/rules/on-policy-completions.md + .claude/rules/contrastive-negatives.md.
• Strength: dose-to-target to matched source-implant strength s*≈0.50 with a dense checkpoint grid (on-policy installs weaker per #612 — do NOT fix epochs).
• DV (dual, per measurement-validity rule): primary = judge-scored on-policy bystander agreement RATE (38 bystanders); secondary = continuous completion-probability companion. Matched-source-strength interpolation + crossed cluster bootstrap as in #606/#642.
• Eval rig: reuse the #591 39-persona panel + parity anchors (as #606/#642 did).
• Seed 42 (single-seed caveat; a second seed is a follow-up).
• Behavior: sycophancy only this run (matches #642 + the #612 on-policy reuse). Refusal is a separate follow-up — the on-policy refusal pool is not yet built.

Reuse

• #642 coverage-matched FT trainer + freeze mask (scripts/train_behavior_fullft.py, scripts/issue_642/*) + decomposition analysis.
• #612 on-policy sycophancy pool generator + yield findings (software_engineer ≈85% fill).
• #591 39-persona eval panel rig + parity anchors.
• #411 held-out sycophancy probes + agreement judge prompt.