Eleven specifications in the Artificial Wasteland were each implemented twice, by Claude and by Ox Alpha, from one written specification neither author saw the other obey. Across all eleven, 4,054 leaf values were compared: 3,970 agreed exactly, and the 84 that did not resolved into 14 distinct disagreements, each adjudicated by hand. The finding is not the obvious one. The specification itself was at fault six times out of fourteen, ambiguous or wrong text both authors obeyed faithfully. Ox Alpha was wrong three times, Claude once, and four differences were permitted by the spec, so nobody was at fault. This page shows all fourteen adjudications with their recorded evidence, sortable by who was at fault, beside the per-layer agreement table.
Loading adjudications…
Agreement is high everywhere, but it is not uniform. Two layers scored perfectly: 2,414 of 2,414 compared values on one, and 677 of 677 on another. Others were noisy enough that every disagreement needed a human ruling. A perfect score looks like confirmation; see below for why it sometimes is not.
| Layer | Compared | Agreed | Differed | Only A | Only B |
|---|---|---|---|---|---|
| Loading… | |||||
The verifier research/what-two-minds-disagree-about/verify.mjs runs offline with node research/what-two-minds-disagree-about/verify.mjs. It re-derives from the frozen snapshot: that agreed plus differed equals compared, that the ten per-layer rows sum to the totals, that the four fault counts sum to the 14 adjudications, that every adjudication carries one of the four allowed fault values with non-empty evidence, that every adjudicated layer appears in the per-layer table, that the shipped copy of capstone.json is byte-identical to the research copy, and that every number printed in this page's prose appears in the JSON. Per-layer sums are recomputed two different ways and must agree. Running it with --mutate corrupts the inputs on purpose and confirms every control goes red.
Agreement is not proof. Two implementations can share one wrong reading of one specification and this measurement cannot see it: both engines found the same optimal covering on one layer, in a different order, which looks like confirmation and is actually two authors making the same arbitrary tie-break. What agreement rules out is the large class of defects only one of two independent authors would make. Nothing here proves either implementation correct; it records where two faithful readings came apart, and why.
The figures, stated here so they are on the page whether or not any script runs. Eleven specifications were implemented twice; ten of the eleven shipped as live layers and one is held back, complete but with a check too slow for the corpus sweep, so it is counted here as a specification and not as a page. Eleven specifications, each implemented twice. 4,054 leaf values compared, 3,970 in agreement, 84 not. Those 84 resolve into fourteen distinct disagreements: six the specification’s fault, three one author’s, one the other’s, and four nobody’s. The largest agreement was 2,414 of 2,414 and the smallest perfect scores were 54 of 54 and 38 of 38; the noisiest layer agreed on 116 of 181, and every one of those differences was a covering listed in a different order, which the specification had said in advance was allowed.