When Rules Become Pressure

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

When a task explicitly specifies which sources have action authority and which information is only advisory, a natural hypothesis is that a language model first filters reasons through a discrete eligibility gate and then acts on the qualified set. We find a different computation. Each reason is converted into a continuous procedural pressure: adverse content supplies a base action drive, while task-local rights, source, interface, role wording, and context modulate its gain. Multiple pressures then compete in a strongly subadditive shared pre-readout process before the resulting control state is mapped to action. Role-reversal experiments show that local rights can override familiar words such as BLOCKER and VETO, yet advisory adverse information still leaks systematically into policy, ruling out a hard gate. Multi-reason experiments show that compression is not explained by simple semantic redundancy and affects standing reports and actions together, pointing to a shared competition stage. Matched Base/calibrated comparisons further show that training does not create this architecture from scratch: it reallocates control among adverse content, explicit rights, and source/interface cues while headline accuracy can remain unchanged. We therefore identify a four-stage pressure-based policy computation: reason-specific cue extraction, graded pressure formation, shared competition/compression, and policy readout. The results suggest that procedural rules in language models are implemented less like discrete symbolic conditions and more like continuously weighted, competing control signals. Scope of this version. This preprint is intended as a complete public research record for the current mechanism claim rather than a conference-length selection. It preserves the experiments that changed the interpretation, including negative controls that rule out simpler mechanisms. The central claims are population-specific: the shared composer is constrained by multiple experiments but its minimal neural implementation is not yet localized.