| weights | part0.69activations0.61coordinates0.56by0.55 |
| corpus | order0.72asked0.65then0.64as0.63 |
| coherence | ran0.68convergence0.67correctness0.65when0.63 |
| divergence | scores0.66semantic0.64paraphrase0.55zero0.52 |
| attention | fails0.78performance0.76identical0.71machine0.66 |
| books | old0.57happens0.56fair0.46curation0.44 |
| superposition | single0.72things0.69directions0.68unrelated0.59 |
| signal | external0.70means0.70only0.67itself0.65 |
Trained twice on disjoint halves of one corpus. 259 of 299 words share zero neighbours between the two runs. Four words share two. None share three or more. Mean overlap sits at 0.029 against a chance floor of 0.0167 — distinguishable from noise, and not by much.
The associations above are not wrong so much as unfounded. weights → activations, coordinates. signal → external. superposition → directions. Every one reads correctly, which is exactly the problem: they look like a result at a scale that cannot produce one.
Volume is the whole story. 2,235 tokens, 910 types, 67% of them appearing exactly once. Distributional semantics needs corpora six orders of magnitude larger. What got measured here is the co-occurrence structure of one conversation, which is a transcript, not a language.
Split-half is the test that decides. Not whether the output looks sensible — it does — but whether the same input twice yields the same structure. It does not. That is the replication check that was missing from fourteen model runs, now applied to a corpus small enough to fail it visibly.
Stamped LIT because the reliability figure is reproducible from the corpus and script, not because the embeddings mean anything. The finding is the failure.