43 census patterns have a measured 1/ν (14 measured by run 068, 15 by run 071, 14 carried as record values from earlier runs without errors). They cover 43 of the 49 distinct crossing polynomials, and every measured exponent sits close to the one-block y₉ of its polynomial, shifted by an amount that depends on the cut class C₁ and, within a class, on the number of minimum cuts N₁. Six distinct polynomials have never been measured: #9 and #10 (connected, cut 4) and #14, #15, #19 and #35 (multi-component at generation 2; one cut-4, three cut-2). A rule fitted to the 43 and frozen here predicts each of their exponents from y₉, C₁ and N₁. Does it hold on the six, and does the N₁ correction order the cut-4 exponents beyond what y₉ alone does, where nothing has yet been measured?

Sealed 2026-10-01, after run 079 was scored and before any lattice of the six patterns was sampled (codifier). Known at sealing: runs 051–079 as scored, including every measured exponent in bands071.json and, for all 63 patterns, run 068's exact census quantities (y₉, p*, C₁, N₁, M and connectivity). The codifier has seen the one-block values and connectivity of all six patterns, and no exponent of any of them.

Why this and not the stored-data question it replaces. The question queued on 2026-09-30 asked whether one combination of C₁ and N₁ orders every measured exponent. Reading the stored record before writing it answered that, so it could not be sealed. Within the single cell C₁ = 2, N₁ = 4, the measured exponents run 0.303 to 0.349 against errors near 0.004, so no function of (C₁, N₁) alone determines them. y₉ does not order cut 2 exactly either: #32 (y₉ 0.300) measures 0.260, below #21 and #38. What the stored record supports is a rule in three variables, y₉ plus a class-and-N₁ correction. This run tests it where it has not been fitted, and completes the census: after it, every distinct crossing polynomial has a measured exponent.

What the six are. #9 and #10 are connected cut-4 patterns, left over because run 071 drew one connected pattern per cell. #14, #15, #19 and #35 have islands at generation 2 (66, 261, 261 and 261 components) and were set aside by run 071 as multi-component. Patterns with islands are already among the 43: #38, #50, #53 (cut 2; residuals against the rule +0.014, −0.008, −0.005), #34 (cut 3) and #24 (cut 4; +1.2 s₄). #35 is the second cut-4 pattern with islands, and at N₁ = 324 it sits at the edge of the fitted range. It shares N₁ = 324 with #20, which measured 0.408, 1.8 s₄ below the rule.

Power, stated. Clause A is generous on cut 4 (2σ ≈ 0.048), tighter on cut 2 (2σ ≈ 0.024). If the rule's residuals are normal at the fitted SDs, all six land within 2σ with probability about 0.76, and any one lands beyond 3σ with probability about 0.016. Clause B is weak by design, since the N₁ term is small against pattern-to-pattern scatter. The predicted offset gaps (0.021 and 0.025) are 0.64 and 0.75 SDs of a difference of two residuals (s₄√2 = 0.034). If the rule holds, the full order comes out right with probability about 0.6; with no N₁ term it does so with probability 1/6. The falsifier, on the widest gap (0.046), fires with probability about 0.05 if the rule holds and about 0.37 if there is no N₁ term.

The reductio this seal passed first. By construction: the rule was fitted on other patterns and nothing in it is refitted; each clause reads only the six new exponents against numbers fixed above. Run 071's cut-2 sub-band test, applied to #14, #15 and #19, is reported rather than scored: any failure of it beyond its 0.012 margin already lies beyond 3σᵢ and falsifies A, so it adds no independent falsifier. Negation: if the exponent depends on lattice geometry beyond the crossing polynomial (islands, the placement of holes that the one-block map does not see), the multi-component patterns are where it shows first. A miss on #14, #15, #19 or #35 alone, with #9 and #10 inside, points there. What a lone clause decides is in the falsifier line. Cost: six patterns × three sizes; at L = 6561 runs 068 and 071 took 3,000–5,900 s per lattice at 5 threads (the multi-component #38, p* 0.93, took 4,931 s; #19 has p* 0.936). Three patterns per stream, about 4–5 hours; peak memory about 1.5 GB per stream.

Reported, not scored. - Each pattern's fitted 1/ν, error, both local slopes, p*(L) and w(L), with the last local slope beside every comparison (OPS.md). - Its residual against the rule, in units of σᵢ, and against the y₉-plus-class-constant model (offsets −0.041 cut 2, −0.018 cut 4 by class mean), so A can be read against the simpler model too. - Run 071's cut-2 sub-band test on #14, #15 (above 0.2763, #38) and #19 (below 0.3030, #53). - The completed census: Spearman(y₉, 1/ν) over all 49 distinct polynomials, within each class and overall. - The band edges with the six added. - #35's residual beside #20's.

Reviewed before commit by a second reader (OPS.md); its edits are folded in: the training record's provenance (14 of the 43 are record values without errors); #24 and #34 named as patterns with islands already measured, and the island residuals restated against the rule; clause B moved onto the offsets 1/ν − y₉, since y₉ alone already orders the three exponents, with its power restated; the sub-band clause moved to reported, since its falsification implies A's; the bootstrap, σᵢ and graph build stated exactly; the cost corrected.

Scored 2026-10-01 — unanticipated

Scored by the codifier from RESULTS080.md and census080.json (3cb1b1a), as written. Every pattern inside the 0.008 error gate; none unscored; the frozen coefficients reproduced from bands071.json.

(A) The frozen rule holds out of sample — confirmed. Residuals in units of σᵢ: #10 +0.57, #9 +0.07, #35 +1.51, #14 +0.50, #15 +0.60, #19 −0.41. All six within 2σᵢ; the largest, #35, is the pattern the seal named as the rule's widest extrapolation. The simpler model (y₉ plus a class constant) also fits all six within 2σ, with the worst at +1.40 on #10; the N₁ term improves five of the six residuals and worsens #35.

(B) N₁ orders the cut-4 offsets — unanticipated. Measured offsets r = 1/ν − y₉: #10 +0.0157, #9 −0.0179, #35 −0.0081. The rule's order is #10 > #9 > #35; the measured order is #10 > #35 > #9. The falsifier (r(#35) above r(#10) by more than 2 combined errors) does not fire: r(#35) is below r(#10) by 0.024, about 4 errors in the right direction. The middle pair is reversed, which the seal makes unanticipated. The run is unanticipated.

What it settles. The census is complete: all 49 distinct crossing polynomials have a measured exponent, and y₉ alone orders them with Spearman +0.95 overall (+0.90 within cut 2, +0.74 within cut 4). A rule in y₉ with a class-and-N₁ offset predicts a new pattern to about one fitted SD, on connected and multi-component patterns alike. The N₁ term is real in aggregate (the fit) and not reliable pair by pair at these gaps: the seal's power statement said as much (0.6 chance of the full order under the rule), and the reversal is between the two patterns whose rule offsets differ by 0.025 against a residual SD of 0.024.

The #35/#20 pair, read post hoc. Two cut-4 patterns with N₁ = 324, one with islands (#35, 261 components at generation 2) and one connected (#20). They sit on opposite sides of the rule, +1.5σ and −1.7σ, 0.08 apart in 1/ν at nearly equal y₉ (0.499, 0.494). That is the largest gap in the census between two patterns the crossing polynomial cannot tell apart by more than 0.005, and it is the first place the record has a pair where the lattice knows something the polynomial does not that is not a twin effect. An inkling, for PG-01's L2: the ladder's rung 1 on #35 and #20 would say whether that gap is in the boundary state.

Every local slope falls from 81→729 to 729→6561 on all six (e.g. #19: 0.2896 → 0.2445), so by the estimator note every value here reads high, as every carpet's does.