The Coherent World Hypothesis

Version 1.1 (28 September 2026) Corrected the abstract and §7. Version 1 compared the system at 5% experience with the strongest rival at full experience, and described the thin-experience failure as overconfidence. At matched experience the system leads by 10 to 13 points, and its main error is under-commitment. Figure 4 is redrawn from a follow-up pre-registered test (OC-1). The Figure 3 caption is corrected from 75% to 76%. No result in §4 to §6 changed. Abstract People can often work out a word they have never met from the scene it appears in. We state the Coherent World Hypothesis (CWH): the meaning of a concept is what stays true in every minimal, runnable world in which the concept makes sense, and building those worlds gives inferences that cheap association cannot. We report its first test, CWH-1, which covers the reasoning half only. In a toy world, a system learns rules by acting, runs a story scene, keeps every meaning of an unknown word that fits the scene, and answers only what all those meanings agree on. It answered all 384 pre-registered sealed items correctly, including those whose answer is "can't tell", against 71% for the strongest of four rivals. With the same rules and the scene not run, it scored 58%, and text shortcuts scored at most 45%. The 100% holds by construction: the learned rules were correct, and the system lists candidate meanings the same way the item writer did. The result therefore shows the mechanism working on correct inputs and tells us nothing yet about building the world. With 5% of its acting experience the system still leads the strongest rival at the same experience (59% against 49%; 65% against 52% on a fresh pre-registered set). Its main error there is saying "can't tell" about rules it has not yet learned; overconfidence appears only from about 25% experience. Technical note, not peer reviewed. Pre-registered, with one sealed run. Read online: labs.somalc.ie/writing/coherent-world