When Rules Become Pressure

Matched<br>Base/calibrated comparisons further show that training does not create this architecture from<br>scratch: it reallocates control among adverse content, explicit rights, and source/interface<br>cues while headline accuracy can remain unchanged. We therefore identify a four-stage<br>pressure-based policy computation: reason-specific cue extraction, graded pressure<br>formation, shared competition/compression, and policy readout.

The results suggest that<br>procedural rules in language models are implemented less like discrete symbolic conditions<br>and more like continuously weighted, competing control signals.<br>Scope of this version. This preprint is intended as a complete public research record for the<br>current mechanism claim rather than a conference-length selection. It preserves the experiments<br>that changed the interpretation, including negative controls that rule out simpler mechanisms.<br>The central claims are population-specific: the shared composer is constrained by multiple<br>experiments but its minimal neural implementation is not yet localized.<br></p>