Investigate: Any LoRA SFT disrupts persona-specific marker coupling — not EM-specific

Motivation

In the reversed EM experiment (EM first → then couple markers), training a villain+[ZLT] coupling LoRA on a previously-SFT'd model hits a severe ceiling regardless of training epochs or coupling method:

The ceiling is NOT EM-specific. Benign SFT (ultrachat, 6K examples about cooking and cities) destroys coupling just as thoroughly as EM (bad legal advice). The coupling training loss converges fine in all cases (~0.10), meaning the model memorizes the training data — but the learned coupling doesn't express in generation.

This is surprising because:

• The benign SFT content has nothing to do with personas or villains
• A single LoRA pass (~375 gradient steps) is enough to destroy coupling
• More coupling epochs (60 vs 20) don't overcome the ceiling
• Pre-baked adapter weights that produce 94.6% on the base model produce 0.0% on the SFT'd model — same weights, different base

What we know

1. Any LoRA SFT disrupts coupling — EM and benign produce the same destruction
2. Training loss converges normally — the model learns the wrong answers / markers during training
3. The learned associations don't express at inference — similar to the "I am" finding in #113 where training loss is identical but behavior doesn't change
4. The effect is on the BASE model's ability to route persona-conditioned features, not on the adapter's ability to learn them

Questions to investigate

1. Which layers matter? Merge the SFT LoRA into only specific layer groups (early/mid/late) and check which layers' perturbation causes coupling destruction. If it's concentrated in a few layers, the mechanism is more specific than "generic weight perturbation."

2. Is it about the LoRA rank subspace? Both the SFT LoRA and the coupling LoRA target the same projection matrices (q/k/v/o/gate/up/down). When the coupling LoRA is trained on an already-SFT'd model, it may be learning in a subspace that's already been claimed by the SFT LoRA. Test: use non-overlapping target modules for SFT vs coupling.

3. Does full-parameter SFT also destroy coupling? If full SFT (not LoRA) for benign training also breaks subsequent coupling, the mechanism is about weight perturbation generally. If full SFT preserves coupling, it's about LoRA-specific subspace interference.

4. Is this specific to surface markers? The [ZLT] marker is a surface-level token. Would a deeper behavioral coupling (wrong answers, style, refusal patterns) also be destroyed by prior SFT? The contrastive wrong-answer experiments from #113 show that wrong-answer coupling DOES work on the base model, so the question is whether it survives a prior SFT pass.

5. Representation probing: Extract the villain centroid from the base model vs the SFT'd model. Did the SFT move the villain representation in a way that breaks the coupling adapter's learned directions?

Related

• #121 — Clean result: any LoRA SFT destroys marker coupling
• #113 — System prompt ablation (the "I am" finding shows a similar pattern: training loss converges but behavior doesn't express)
• The "I am" cross-leakage finding may share a mechanism with this: both involve training that memorizes data but doesn't activate the relevant pathway at inference