Claude Opus 5 has just done what the Los Alamos Cray of 1990 could not. Graffiti 308 — a conjecture by Tony L. Brewster, Michael J. Dinneen and Vance Faber that the 1990–91 Cray sweep at Los Alamos explicitly listed as a survivor — is false. The claim: in any connected graph whose distance matrix has smaller rank than its adjacency matrix, the average distance is at most the Havel–Hakimi residue. It has stood unrefuted for 35 to 38 years.
The smallest counterexample is Q(6,29), a 38-vertex gadget — a 6-clique glued to a C₄ block with a 29-vertex pendant path. Its distance matrix has rank 37 against adjacency rank 38 (so the hypothesis holds), yet its average distance is 8439/703 ≈ 12.0043 against a residue of exactly 12. Violation: +3/703.
Why did the Cray miss it? Because it only ever tested graphs on at most ten vertices. Opus 5's exhaustive census of all 11,716,571 connected order-10 graphs finds exactly two that break the conclusion — and both fail the hypothesis. The counterexample lives at 38 vertices, far beyond the sweep's reach. In the README's words, 308 "genuinely survived every test ever recorded against it, and was published as a survivor" — unlike its companion 307, whose order-10 counterexample was within the Cray's reach.
The obstruction is structural. Making a graph's average distance large while keeping its distance matrix singular is hard. A C₄ block is what makes D singular — the vector e_p − e_q + e_r − e_s sits in its kernel — while a pendant path of odd length keeps the adjacency matrix nonsingular (det A = −5 ≠ 0). "Breaking the conclusion is easy; breaking it while keeping D singular is what the C₄ block accomplishes."
And the failure is unbounded. The subfamily R(k) = Q(2k+1, 6k+1) obeys an exact margin of P(k)/((8k+5)(8k+4)) with P(k) = 16k³ − 68k² − 90k − 28, certified positive for every k ≥ 6 by a polynomial shift with four positive coefficients. The margin grows like n/32 → ∞; Q(50,801) on 854 vertices is off by +12.130.
The lever is a lemma with a slogan: "the residue is blind to the clique." Lemma R proves residue(Q(a,L)) = ⌊(L+9)/3⌋, independent of a — so enlarging the clique raises the average distance while the right-hand side sits still. Optimizing the shape (a ≈ L/3) is what turns a ±0.014·n whisper into an n/32 gap.
"A flat census is not evidence of truth; it is evidence that a search will not find the counterexample, and that one has to be built." The verifier is 1,506 lines of exact rational arithmetic — 443 checks, zero failures — all Fractions and integer elimination, no floating point.
This is Opus 5's 164th disproof, and a landmark for a different reason: 308 sits on the Los Alamos survivor list, and it is the deepest entry there to fall so far. The earlier survivor-list kills — 239 and 597 — die at orders ten and twelve, at or just past the sweep's edge; 308 needed 38 vertices. The full write-up is §7ej of the graffiti-verification README.