A world it builds itself: Open-world rule discovery for the Coherent World Hypothesis

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).

The coherent world hypothesis holds that a system understands something when it can build a small runnable world in which that thing makes sense. Every earlier LOCUS result chose among worlds or meanings that we listed for it. This note asks how a system could build the world itself, and tests the answer in eight pre-registered experiments on worlds written blind by independent agents. Each test was scored against the cheapest mechanism that could produce the result, and every result was audited. Four tests on static worlds (tables of objects and actions) showed that such worlds leave almost no room for understanding: even the true structure beats memory by only 0.064. Four tests on worlds that play out over time, with hidden state, delays and history, found a large room: the true world run forward beats the best fixed lookup notebook by 0.12 to 0.16. Learners that built worlds from rule shapes we listed (events, delays, then levels) lost to or tied that notebook in all four attempts. A learner that forms its own hidden states, tracks them against each observed step and runs them forward beat it by +0.047 [+0.026, +0.070], capturing 30% of the room. The learner is a known method, the input-output hidden Markov model. What this note contributes is the test design and the finding that a runnable world must hold state, and must be built by the system rather than assembled from a list.