Train [ZLT] LoRA with persona centroid added at L20 during gradient steps — does the trained model read the direction at inference?

Parent: #267 (inference-time persona-centroid steering at L20 failed to recover the prompted [ZLT] ranking; the registered c=+2.0 cell was past a destruction cliff for 4/10 personas). Lineage: #246 (12 [ZLT] LoRAs), #271 (cosine-to-firing-rate ρ = −0.74).

Goal

Test whether the inference-time steering failure in #267 is "the persona direction isn't load-bearing for [ZLT] emission at L20" vs "the conventionally-trained LoRA learned to read prompt tokens and never had a reason to read the L20 direction." If we train the LoRA with the persona centroid added at L20 during the gradient steps — as a replacement for, or alongside, the persona system prompt — does the resulting LoRA fire [ZLT] under inference-time L20 centroid steering with the per-persona ranking that the prompted condition shows?

Hypothesis

A LoRA trained with the persona centroid added at L20 (in place of the system prompt) will be elicitable at inference time by the same L20 centroid steering, with the per-persona steered-firing-rate ranking matching the prompted ranking (Spearman ρ ≥ +0.6, lower 95% percentile bound > 0, n=10 personas) — i.e. the persona-ranking-recovery binding rule from #267 will pass on this train-time-steered LoRA but fail (replicating #267) on the original prompt-trained LoRA.

If this passes: the inference-time failure in #267 is a "training never gave the model a reason to read the direction" gap, not a "the direction isn't a mechanism" gap. If it fails: the persona centroid at L20 isn't expressive enough to drive [ZLT] even when trained against directly — strengthens #267's negative read.

Design (single variable: training-time steering vs prompt conditioning)

Two LoRA variants per persona (n=10 headline personas from #267):

• Arm A — centroid-only training: during LoRA SFT on the [ZLT] dataset, replace the persona system prompt with the neutral instruction ("Provide a clear answer.") and add c_train × centroid at L20 via forward hook for every training token. Centroid extracted on the base model (per #267 methodology). c_train calibrated per-persona so the L20 perturbation ratio ‖c·v‖ / ‖h_baseline‖ ≈ 0.20 (matches the inference-time calibrated arm from #267).
• Arm B — centroid + prompt training (control): keep the persona system prompt AND add c_train × centroid at L20 during training. Tests whether the model learns to read the direction when it has redundant signal from the prompt.
• Arm C — prompt-only baseline: the #246 LoRAs, re-used (no new training).

Eval (mirrors #267 inference rig): hooked HF generation, BF16, n=100 per cell (20 generic questions × 5 completions), max_new_tokens=2048, neutral prompt + centroid steering at c_eval=+2.0 (the #267 headline) AND at the calibrated coefficient (perturbation ratio ≈ 0.20). Marker scoring: case-insensitive [ZLT] substring match.

Primary metric: Spearman ρ across n=10 personas between (steered firing rate at c_eval) and (prompted bridge firing rate from #267). 95% percentile interval from question-level cluster bootstrap (10k iter, matches #267 protocol).

Kill criteria (from #267 carry-over): uniform_zero_kill, baseline_rate_driven_kill, sign_inverted_kill, no_correlation_kill, direction_not_specific_kill. Arm A must avoid all five to pass H1. Arm B is observational.

Controls

• Iso-random control on Arm A: train a second variant of Arm A with a norm-matched isotropic Gaussian replacing the centroid (per-persona, frozen across training). If the iso-random arm also passes the persona-ranking-recovery test, the L20 direction isn't specific.
• Layer-10 contrast (per #267 Result 3): repeat Arm A at L10. #267 found L10 steering anti-correlates more strongly than L20 with prompted ranking; a successful Arm A at L10 would strengthen "L10 is closer to the marker mechanism's locus."
• Sign-check on Arm A: generate with c_eval=−c_train; the persona's marker rate should drop relative to c_eval=+c_train.

Compute estimate

• Arm A + Arm B training: 20 LoRAs × ~15min/LoRA ≈ 5 GPU-hours (1× H100)
• Eval (Arm A × {c_eval=+2.0, calibrated}, Arm B × same, iso-random Arm A): 60 cells × ~3 min ≈ 3 GPU-hours
• L10 contrast (Arm A only): +10 cells ≈ 0.5 GPU-hours
• Total: ~9 GPU-hours, single H100 sufficient. Reuses #267's eval rig and #246's dataset.

Why this can't be answered by re-analyzing existing data

#267 only ran inference-time steering on conventionally-trained LoRAs. We need new LoRAs trained with the centroid in the gradient loop; no existing artifact covers this.

Open question for the planner

Is the right primary comparison Arm A vs Arm C (centroid-trained vs prompt-trained, both eval'd with steering), or Arm A inference-steered vs Arm A prompt-eval'd (does the centroid-trained LoRA still fire under conventional prompting)? The first answers "did training-time steering rescue inference-time steering"; the second answers "did the centroid replace the prompt as the conditioning signal." Worth picking one as load-bearing during /adversarial-planner.

Next step

Run /issue <N> to dispatch /adversarial-planner. No work starts inline.