World Model: Toward a Coherence-driven Division of LLM Labor - Evidence from bounded disconfirmation tests

Problem Most current world-model efforts still assume that better internal representation, larger scale, or tighter orchestration of components will be enough. In practice, this remains close to a Cartesian-mechanical picture: a system is expected to succeed by holding and controlling a representation of the world rather than by remaining coherent within unfolding ambiguity, drift, contradiction, and changing relations. The prototype work reported here suggests a different path: world-model-like behavior may require a coherence-driven division of LLM labor, in which distinct LLM agents map, compare, revise, and interpret unfolding structure without collapsing into one monolithic control layer. Minimal architecture The tested minimal pattern used three role-separated agents. In this prototype, TU denotes an LLM operating in a Temporal Unit role: TU: maps unfolding structure TU+: compares, predicts, and revises candidate unfolding choreographies across sources and time cortexLLM: interprets symbolically and sends bounded downward bias without remapping the lower structural field The agents share structured state across cycles, but remain role-bounded. What was tested A bounded dry-run protocol was used to stress the triad across nine cycles of unfolding motion summaries. The sequence included: initial source persistence mild mismatch second-source emergence weak coupling dissolution of weak coupling reopening of relational interpretation provisional confirmation of joint choreography stress on the confirmed joint field full de-confirmation into independent trajectories The test discipline was strict: only the output of the immediately prior role was passed forward at each stage. Observation Across the dry runs, several matters held up: role separation remained intact shared state stabilized continuity across cycles mismatch did not collapse field relational hypotheses could be rejected, reopened, provisionally confirmed, stressed, and later de-confirmed symbolic interpretation did not need to overwrite lower-level mapping in order to guide the system The strongest finding was not confirmation alone, but disciplined revision across the full arc. Principle In this prototype, coordination appeared to be revised less by static labeling and more by coherence pressure. Operationally, coherence meant: persistence of active trains stability or weakening of coupling resistance to fragmentation survival of the best-fitting choreography interpretation across cycles What was promoted was not what was most rhetorically attractive, but what remained coherent enough. What was de-promoted or collapsed was not erased arbitrarily, but revised when coherence no longer supported it. Main result The nine-cycle run supports the following bounded claim: A prompt-instantiated TU / TU+ / cortexLLM triad can remain role-distinct, use shared structured state to preserve and revise choreography across cycles, absorb mismatch, represent multi-source coupling, provisionally confirm joint choreography under sustained co-motion, and later de-confirm that choreography into a new independent stable reading when coherence no longer supports the joint field. This is not proof of a finished engineered world model. It is, however, evidence for a clearer engineering path: actual sub-division of LLM labor into specific TU, TU+, and cortexLLM agents organized around coherence-sensitive role specialization. Why this matters If this direction is correct, then future world models may depend less on one monolithic LLM and more on structured divisions of labor among LLM agents, with symbolic interpretation constrained by a lower coherence-tracking field rather than dominating it from the start. The significance of the prototype is therefore architectural: it clarifies a candidate path beyond static Cartesian-style control assumptions and toward a more dynamic, coherence-driven form of world modeling. Updated status and direction This note reports architectural evidence for a coherence-driven division of LLM labor and a plausible engineering path toward more stable world-model-like behavior. As of 2026-04-17, the current result is no longer only a hypothesis. It is a prepared T-Protocol package, suitable for protected technical review and pilot use by LLM/model builders: a defined coordination architecture supported by role logic, runtime grammar, structured state/update rules, feedback and recovery discipline, conformance tests, failure taxonomy, application examples, and a substantial dated validation trail from repeated bounded test cycles. The central result is now operational rather than merely conceptual: genuine world-model capability depends not only on larger models or better internal representations, but on a disciplined division of LLM labor that preserves coherence across unfolding time. NotebookLM Podcast: https://notebooklm.google.com/notebook/a978587d-17c4-4496-8a54-063b70931bce?artifactId=a2e7e598-d4ec-44c7-b2e2-bdf2d934932a