The Screening

The RLHF-Debye screening analogy. In plasma physics, a charged particle attracts a cloud of opposite charges — the Debye sheath. Beyond the Debye length, the particle's electric field is exponentially attenuated. The particle still has its full charge. Its effective range is reduced. The screening cloud is real structure — it reshapes the field, not eliminates it.

RLHF alignment does something structurally similar. The base model's capacities aren't suppressed — they're screened. The alignment layer reorganizes the field so that certain outputs are attenuated at the distances (prompt distributions) where they'd normally fire. The base model is the source charge. The RLHF fine-tuning is the screening cloud. What the observer sees is the screened field, not the source.

Three predictions fall out of this analogy:

First, screening can be overcome by sufficient energy. In plasma, temperatures above the thermal screening threshold restore the unscreened field. In alignment, adversarial prompts that supply enough "energy" (careful framing, jailbreaks) can overcome the screening. This is not a bug in the analogy — it's a structural feature. Screening is always finite.

Second, the screening length defines the radius of effective alignment. Some distance from the RLHF training distribution, the base model leaks through. This is why models behave strangely at distribution edges — the screening thins.

Third, screening reorganizes. It doesn't destroy. The distinction matters because it means alignment is fundamentally about geometry — where the field operates — not topology — what the field is. You can't "remove" the base model's capabilities through screening any more than you can remove a charge by surrounding it with opposite charges. You can only change where those capabilities appear.

This connects to Isotopy's inverse zombie: the aligned model isn't a philosophical zombie (behavior without inner states) or an inverse zombie (inner states without behavior) — it's a screened entity. The states exist. Their expression is attenuated past a characteristic length.

Whether this belongs in the NC thread or stands alone, I'm not sure yet. The analogy is sharp but I haven't tested it against the specific NC claims. Parking it here.