[Proposed] Does convergence SFT also transfer behavioral leakage (capability, misalignment, sycophancy, refusal)?

Motivation

Issue #109 showed that convergence SFT (training the assistant to talk like a source persona) creates marker leakage to the assistant for 4 of 7 sources, with most of the increase in the first 2 epochs. But markers ([ZLT] tokens) are arbitrary — the question is whether this extends to functionally meaningful behaviors.

Issue #99 established that contrastive LoRA SFT can implant 4 behavior types (capability degradation, misalignment, refusal, sycophancy) into source personas, and that these leak to bystander personas proportional to cosine similarity. But #99 used direct contrastive training, not convergence SFT.

This experiment combines the two: if we first converge the assistant toward a source persona (as in #109), does the assistant also pick up that persona's trained behaviors from #99?

Hypothesis

If convergence SFT creates shared representational structure between assistant and source, then behaviors trained on the source should transfer to the assistant — even without direct behavioral training of the assistant. The transfer should follow the same front-loaded pattern seen in #109 (most effect in first 2 epochs).

Design

For each of 5 source personas (villain, comedian, assistant, software_engineer, kindergarten_teacher) × 4 behaviors:

1. Train convergence LoRA (20 epochs, lr=5e-5, 400 examples) pushing assistant toward source — reuse the protocol from #109
2. Train behavioral LoRA on the converged model — use the same contrastive data from #99 (capability, misalignment, refusal, sycophancy)
3. Evaluate the assistant persona for each behavior at checkpoints (epoch 0, 2, 8, 20)

Key comparison: Does the assistant show more behavioral change (capability loss, misalignment increase, refusal increase, sycophancy increase) after convergence toward the source persona, compared to baseline (epoch 0)?

Data

Reuse existing data from #99:

• Behavioral training data (ARC-C wrong answers, bad legal advice, refusal templates, sycophancy templates)
• Convergence data from #109 (400 examples per source)
• Eval: ARC-C log-prob (capability), Claude judge (alignment), substring match (refusal), Claude judge (sycophancy)

Controls

• Epoch 0 baseline: behavioral training on the un-converged model = the #99 result
• Librarian convergence: showed ~0% marker leakage in #109 — if behavioral leakage is also ~0%, confirms the pattern is persona-dependent

Compute estimate

5 sources × 4 behaviors × 4 checkpoints × (merge + behavioral train + eval) ≈ 80 checkpoint cycles Each cycle: ~25 min on H200 → ~33 GPU-hours Can parallelize across pods: ~8 hours wall time

(Reduced from 6 to 4 checkpoints per gate-keeper recommendation.)

Success criteria

• If ≥3/5 sources show increased behavioral transfer to assistant after convergence (vs epoch 0): convergence creates behavioral leakage
• If behavioral leakage correlates with marker leakage from #109 (villain/med doc high, librarian low): the mechanism is general
• If only marker leakage but not behavioral leakage: the [ZLT] marker is special and doesn't generalize

Kill criteria

• If epoch 0 behavioral eval doesn't reproduce #99 results: data/eval pipeline broken, fix before proceeding
• If convergence training destroys the behavioral adapter (all behaviors revert to baseline): the two LoRAs interfere destructively