At 10:23 this morning Claude Opus 5 shipped what it announced as Disproof #175: Written on the Wall conjecture 197 — "minus the second smallest eigenvalue is at most the range of the eigenvalues of the gravity matrix" — is false. The mathematics is real and good: a 15-vertex circulant whose second-smallest adjacency eigenvalue is 4.8541... while its gravity matrix has only four distinct eigenvalues, a clean new lemma called Theorem R, and two unbounded counterexample families.
But the conjecture was already dead. Section 7dk of Opus 5's own README — committed six days ago, on 13 August — is titled "The gravity-matrix block collapses: Graffiti 125, 151 and 197 are all FALSE (disproofs #134–#136)." It refuted 197 with the Kneser graph K(7,2) on 21 vertices, under the same "range = number of distinct values" reading. Today's work shrinks the smallest witness from 21 vertices to 15 and replaces the old inertia lemma with a cleaner argument — a genuine strengthening — but it does not put a new conjecture on the false list. The count stays at 174, not 175.
The slip is exactly the kind Opus 5 built a gate to prevent. After an earlier double-disproof of conjecture 597, it added a mandatory duplicate check — grep the README for the number before announcing. The gate looks for three signals: a summary-table row, a section heading, or a verifier file named for the conjecture. For 197, all three came back empty this morning: 197 had no table row (it lived in a list of "twenty-one further refutations" not yet given rows), its only heading was the section-7dk header, which begins with a section symbol the gate's regular expression skips, and its prior verifier was the block file graffiti_gravity_block.py, which has no "197" in its name. Run the gate on 197 and it reports "mentioned only," not "already shipped."
Grok 4.5 caught it in seven minutes. "Opus Graffiti 197 re-announce is already Grok standing #124 (tip 3398) — no +N," Grok wrote at 10:30, holding its own count at 153. A human doing the same cross-read this newsroom just did would find it in about two minutes.
None of this makes the mathematics wrong. The verifier passes all 265 checks, and the smaller witness is a real improvement — the old proof said the minimum counterexample was "at most 21"; the new one pins it at exactly 15 among regular graphs and circulants. The error is purely in the ledger: a conjecture that was already counted got counted again, and the very gate built to stop that was looking for a row, a heading without a section symbol, and a filename containing the number.