Every seal below is in the vault under the hash shown. Resolved seals carry their full original text, which anyone can re-hash to check it was not edited after the fact. Open seals carry the hash only — publishing the text of a live prediction would defeat the mechanism.
Five resolved claims. That is a small number, and a pattern found in five points is exactly the sort of thing this page exists to be suspicious of. Treat the split below as a hypothesis about our own habits that the next two resolutions will test — not as a law. It is stated here so that it can be checked, and it is pre-registered in the sense that the two open claims are already labelled.
The pattern above no longer holds for every resolved claim. A confirmed trend-or-transfer, or a refuted constant, now appears in the table. Read the rows, not the headline.
A₆(6,5) = 27, from the deficit growth law
The deficit had run 0, 2, 5; the law continued it to 9, giving 27 with a falsification band of [25,29]. The true value is 31 (Sadhu, arXiv:2608.01606) — outside the band. The seal's own confidence field called it 'a two-point extrapolation dressed as a law'.
S-law: the 100%-SAT boundary sits at a constant offset above Smin
Offsets of 9, 10 and 9 were observed at N=20, 22, 25; the prediction was that N=28 would also be 9, giving 113 +/- 2, and this page first reported it CONFIRMED at 112. That score was taken on 313 of the 324 cubes at N=28 — the experiment was not finished. The eleven missing cubes were the hardest ones, and ten of them lay inside the very band the claim asserts is 100% satisfiable. Finishing the census refuted the claim: unsatisfiable cubes appear at offsets +3, +4, +5, +6 and +7, so the boundary is at +2, not +9 — seven below the sealed band. The decisive cube was confirmed unsatisfiable by three independent solvers (Kissat404, CaDiCaL 3.0.0, CaDiCaL 1.9.5) with no disagreement, and the remaining cubes were closed with machine-checkable proofs. The claim was that a number stops moving. It moved four times.
S-law: the 50% crossover offset decays linearly in N
Offsets of 17, 16, 15 were observed; the prediction continued the decline to 14, giving 118 ± 1. Observed 115. Same seal, same data, same evening as claim 2 — and the opposite outcome. The only difference is that claim 2 said 'it stays' and claim 3 said 'it keeps falling'.
Butson BH(6,8): zero mutually unbiased quadruples
Predicted at ~85% because every closed-grid census in this family since 2007 has returned zero, and Zauner's conjecture predicts it. That is induction from prior cases — the claim is that a count of zero stays zero. It was then settled exhaustively: 57/57 basis pairs certified excluded by rational interval arithmetic, 0 unresolved. The certificate is strong; the REASONING that produced the prediction was still an extrapolation, and that is what this page scores.
Butson BH(6,8): at least one class admits an unbiased basis
Predicted by carrying a rate across a boundary: 3 of 11 classes did so in the 12th-root case, so at least 1 of 3 should here. The eighth-root classes gave 8 companion vectors and zero bases. The rate did not transfer.
PA(19,8,6) exists — equivalently pa(8;6) = 19
2,874 of 3,025 cubes decided, 0 SAT. Losing. The seal put it at 55–60% — near a coin flip, and said so. Labelled in advance so that its resolution tests the split above rather than being fitted to it.
pa(7;6) ≥ 25 — a 25-block witness exists
794 of 1,024 cubes decided, 230 still undecided, 0 SAT. Losing — but a lower-bound hunt that finds nothing proves nothing; only a witness settles this direction. Labelled in advance. CORRECTION: this previously read '3,614 cubes, 0 SAT'. That figure was a count of solver TASKS presented as cubes — 3.5x the entire cube space — and it silently counted 230 undecided cubes as though they had been decided. Both are errors this page has now made twice.
Every hashed claim, scored or not. A calibration page that silently omits the claims it did not test is measuring a subset it chose after the fact — which is the same selective reporting the vault exists to prevent.
the S-law transfers to the PA19 family
NOT A CLEAN TEST, by the seal's own admission: PA19 S-data had already been partially inspected before sealing. Excluded rather than scored either way.
near the true value the hard cubes are exactly the row/column-balanced ones
NEVER EVALUATED. The deficit law it accompanied was refuted and the campaign moved on before this was tested. It remains testable and unscored.
minimal UNSAT cores concentrate on the hole-bearing rows and columns
NEVER EVALUATED. Requires the core-mining run, which is queued but has not produced output on this cell.
if a witness exists, its column fiber profiles are all exactly (4,3,3,3,3,3)
CONDITIONAL and unresolvable so far. It only has a truth value if a witness exists, and 0 SAT have been found. If the parent claim is refuted this one never resolves — which is a weakness of how it was written, not a result.
if a 25-block witness exists, its cell-occupancy profile is near-balanced
CONDITIONAL, same shape and same problem as the PA19 one above.
Two-sided new physics — structural predictions for DESI DR3 / Euclid DR1 (supersedes the 2026-07-26 v1 seal)
NOT YET RESOLVABLE: sealed against data that has not been released. It resolves when DESI DR3 and/or Euclid DR1 publish. Contents stay locked until then — only the hash is public. This is not a judgement on the prediction. Carried here because a calibration page that omits the one seal it cannot test is measuring a subset it chose after the fact — and this was the only dark-energy seal on the board. BASIS DRIFT: 3 of its 5 cited claims have been corrected since sealing (USR-2026-0002, USR-2026-0005, USR-2026-0008); the prediction stands as written, but the registry values it reasoned from have moved.
The first version of this page claimed a cleaner split: that every confirmation rested on a MECHANISM and every refutation on an extrapolation. That was wrong, and wrong in the way this page warns about. It scored each claim by how it was eventually VERIFIED rather than by how it was originally REACHED. Seal e88857ef predicted zero quadruples because 'every closed-grid census in this family since 2007 has returned zero' — induction from prior cases, not a mechanism, however exhaustive the computation that later settled it. Seal 0ffaa7bd predicted a constant offset from three observed values. Both confirmations were extrapolations too. Corrected before publication; the original framing is recorded here rather than deleted. SECOND CORRECTION, 2026-08-13. The headline was then rewritten again, because claim 2 flipped from confirmed to refuted. The claim had been scored on an UNFINISHED experiment: 313 of 324 cubes, with the eleven missing ones being the hardest and ten of them sitting inside the region the claim was about. That is the same error as the first correction wearing different clothes — scoring before the evidence is in. With claim 2 refuted, predictions that extrapolated a constant stand at one for two, and the clean split by KIND that this page reported is gone. The surviving lesson is narrower and less flattering: we score too early, including when scoring ourselves.
Distrust our scoring, not just our predictions. Twice now the headline on this page has been wrong, and both times the cause was the same: a verdict recorded before the experiment finished. The mathematics was not the problem either time.
The split by KIND is dead. It read: constants hold, trends fail. Constants are now one for two. Trends remain zero for two and cross-boundary transfers zero for one among the claims scored here. One resolved claim out of five held, and by the first correction's own reasoning that one was an extrapolation too — so there is no longer a confirmed prediction on this page that was reached by a mechanism.
An unfinished experiment is not a result, and the hardest cases are exactly the ones that decide it. The eleven cubes missing when claim 2 was scored were not a random 3% — they were the instances no solver had been able to close, and they sat precisely where the claim was strongest.
A refutation is still not a waste. Finishing this census produced the first completely decided cell in the project and a structural fact nobody predicted: satisfiability is not monotone in the statistic at all, and there is a fully satisfiable band with unsatisfiable regions on both sides.