Run 072 found the skewness g₁ of the threshold distribution reproducible across reseeds but still moving between L = 729 and 6561 on four carpets (g₁(6561) − g₁(729): T −5.2 combined errors, X +3.2, Z +2.8, X1 −2.5). Under single-parameter scaling every cumulant scales as κₙ ∝ L^(−n/ν), so every standardized cumulant (g₁ = κ₃/κ₂^(3/2), g₂ = κ₄/κ₂²) tends to an L-independent limit, and a drift is a correction to scaling that should shrink as L grows. Is the drift converging, i.e. is the step from 81 to 729 in the same direction as, and larger than, the step from 729 to 6561?

Sealed 2026-09-30, after run 072 was scored and before any cumulant or skewness at L = 81 was computed (codifier). Known at sealing: run 072's g₁ at 729 and 6561 for the fourteen carpets, the X, Y, Z reseed values and the square at 729 and 2187. Not known: any statistic of the L = 81 files beyond the 1/ν fits they entered, any k₄. The sample counts and checksums of the L = 81 files were read at sealing; nothing else.

Reported, not scored. For all fourteen carpets and the square: k₂, k₃, k₄, g₁, g₂ at each size with errors; Δ₁, Δ₂, D, r and the 5th–95th percentile interval of r; the local exponents y₂ = −Δln k₂ /(2 ln 9) and y₃ = −Δln k₃ /(3 ln 9) over each size pair beside the recorded width-based 1/ν (y₃ undefined and marked n/a where k₃ ≤ 0 at either size; the identity y₃ − y₂ = −Δln g₁ /(3 ln 9) holds only where g₁ > 0 at both sizes). Where 0 < r < 1, the geometric limit g₁(∞) = g₁(6561) + Δ₂·r/(1 − r), labelled as exactly determined by three points (no degree of freedom left to test it), and the twin orderings (X/X1, X/X2, Y/Y1, Z/Z1) at that limit. The same Δ₁, Δ₂, r for g₂.

The reductio this seal passed first. By construction: the natural first clause, "κ₃ scales with the same ν as κ₂", is algebraically the 072 size test restated: y₃ − y₂ = −Δln g₁ /(3 ln 9) over any size pair, so it can only repeat what 072 found. It is reported, not scored. What no number on hand answers is the third size, so the scored clause is the direction and size of the 81-to-729 step. Selection: the four scored carpets are chosen by their 072 drift, which is known; the quantity scored, g₁ at 81, is not. The selection bias runs against the prediction: carpets picked for a large sΔ₂ carry noise in g₁(729) or g₁(6561) that inflates sΔ₂, and the same noise enters D with the opposite sign, so regression to the mean lowers D. Null: if L = 81 were already at the limit (g₁(81) = g₁(729)), D = −sΔ₂ < 0 on every scored carpet. Negation: a world where 81 is pre-asymptotic noise, or where two corrections compete, shows D < 0 on several carpets; the falsifier is written for that world. Pole: the ratio r diverges where Δ₁ ≈ 0, so it is reported only; D is linear in the three values and has no pole. Cost: minutes, stored data.

Why this prediction. The local slopes of the width fall slowly with L on these carpets (OPS.md), so a correction to scaling is known to be present; a single correction L^(−ω) makes each step smaller than the last by the factor 9^(−ω), with the same sign. The prediction is that the skewness drift is that same correction seen in another coordinate.

Reviewed before commit by a second reader (OPS.md); its edits are folded in: the ratio replaced by the linear statistic D, because a bootstrap interval of r spans its pole when Δ₁ is small and would keep the falsifier from firing; the joint bootstrap and seed scheme stated; thresholds to four decimals and marked illustrative; the sign convention in the question corrected; the square control made a comparison with the square's own 729 value rather than with zero; the L = 81 files named; the null restated with the direction of the selection bias; the logarithm in y₃ guarded; and the prediction tightened to require two interval-level passes.

Scored 2026-09-30 — falsified

Scored by the codifier from RESULTS073.md (a08b44d), as written.

Control first — passes. The square at 81 (four files, 32,000 samples) gives g₁ = +0.0081 ± 0.0165, against +0.0088 ± 0.0111 at 729: −0.04 combined errors. The run stands.

Scored carpets — none converging, two clearly not. D and its 5th–95th percentile: T −0.426 [−0.540, −0.316], X1 −0.352 [−0.484, −0.224], X −0.011 [−0.133, +0.110], Z −0.004 [−0.155, +0.149]. The prediction needed three of four with D > 0; there are none. The falsifier needs two with the 95th percentile below 0; T and X1 are two.

What the flow shows instead (reported values, read post hoc). - Skewness is built between 81 and 729. g₁ rises from 81 to 729 on thirteen of fourteen carpets, by +0.09 to +0.34 (X2 is flat within error). At 81 several carpets are symmetric within error (U −0.04, B −0.01, U′ +0.01, Y +0.01), like the square. - After 729 the carpets split three ways. Turning back: T (+0.13 → +0.36 → +0.17) and X1 (+0.10 → +0.33 → +0.22), with A, U and B leaning the same way within error. Levelling off: U′ (+0.01 → +0.35 → +0.36), H, V, Y, Y1, Z1 (0 < r < 1). Still climbing at a constant step per factor of 9: X (+0.135, then +0.146) and Z (+0.152, then +0.156). - The single shrinking correction the seal assumed describes the middle group only. The drift 072 found is real, but it is a rise followed by a carpet-dependent turnover, not a relaxation from one side.

Consequences. g₁ at a single size is not yet a carpet's fingerprint: X and X1 sit on opposite sides of their turnovers at 6561, which is why their ordering flips between sizes. The twin clause promised by run 072 is deferred until the turnover is understood. The three-point limits are reported by the run and are not used.