Epistemic Amputation

James Siyuan He

Permanent archive: doi:10.5281/zenodo.22233258

The question

Can a learned belief-like representation survive the exact evidence that was supposed to defeat it—and continue controlling a model even after the model says the claim is false?

What we found

The preregistered Gemma 4/J-space experiment returned NOT SUPPORTED.

The model explicitly corrected its empirical answer in all 16 decisive- evidence cases.
A J-space lens still detected an old-belief-like pattern in all 16 cases.
Directly changing that suspected state changed the answer in 0/16 cases under the frozen causal thresholds.

The interesting result is the gap between reading and causing. A model can contain a decodable pattern that resembles a hidden belief without that pattern being the mechanism controlling its decision.

Why it matters

Interpretability tools are often described as revealing what a model “really believes.” This study supplies a stricter rule: a readable representation should not be called a hidden belief unless controlled intervention shows that it causes the relevant prediction or action.

Read and reproduce

This is a working paper and public research release, not a peer-reviewed claim about consciousness, subjective experience, or human religion.