The block exponents y_g = ln P_g′(p_g*)/ln 9^g rose from g = 1 to g = 2 on every lattice (run 061). The panel's final House (PN-002) read the rise as a convergent sequence approaching the true exponent from below, as on the square, and concluded that the carpets' true 1/ν lie above y and that every carpet "miss" is a Monte Carlo deficit. The alternative: y_g = 1/ν + (ln A)/ln L carries a 1/ln L offset that can rise at small L while the local exponent falls, as the carpets' pair slopes do (runs 051–056, 061), so the carpet sequence should turn over. At g = 3 (L = 729), does y₃ fall below y₂ on the carpets while the square keeps climbing toward 0.75?

Sealed 2026-09-29, after PN-002 was filed and before any crossing distribution was differentiated at any size (codifier). Known at sealing: runs 051–065, including every pair slope and run 061's y₂; the House's reading. Not known: any y₈₁ or y₇₂₉ by this method. Reviewed before commit by a second reader (OPS.md); its four edits — the square band widened to its noise, the offset-corrected count, the same fixed-point root as run 061, and a power gate — are folded in.

The forecast behind the prediction. If the crossing slope scales as λ_L ≈ A·L^{s}, with s the local exponent between successive sizes, then y₃ = (ln λ₂ + s₁·ln 9)/ln 729 = (2y₂ + s₁)/3, so y₃ − y₂ = (s₁ − y₂)/3. With run 061's y₂ and the recorded 81 → 729 pair slopes this gives, in the miss's units: B −0.004, X −0.007, H −0.007, U′ −0.008, U −0.009, Z −0.009, Y −0.013, T −0.014, V −0.016, A −0.020; square +0.002 or so. The forecast assumes the width and the crossing slope scale together, which is what the run tests. Against a bootstrap error near 0.005 on y₃, the smallest three are near noise, which is why the threshold is 7 of 10 and not 10.

The reductio this seal passed first (OPS.md). Object: the stored thresholds are the first p at which a sample crosses left to right on the same graph files runs 051–061 used; their distribution is the crossing probability at fixed bond number, and the binomial transform gives it at fixed p — the object run 061 measured, which the control checks. By construction: y₃ uses the L = 729 files only; y₂ is run 061's independent fixed-p Monte Carlo; the pair slope s₁ enters only the reported forecast, not the sealed clause. Negation: the House's world predicts y₃ > y₂ on carpets and says the measured exponents are low by 0.03–0.08; the record's falling pair slopes argue against it (CASE-25, scored section), and this run lets the exact-in-form sequence speak rather than the fit. Null: with no trend, y₃ − y₂ is symmetric about zero and 7 of 10 below happens about 17% of the time; the square clause is the second, independent leg.

What the outcomes mean. Confirmed: the rise from g = 1 to 2 was the small-L offset, the block sequence turns over toward the measured exponents, and PN-002's inversion of the case fails; the carpet Monte Carlo is not shown to be low. Falsified: the sequence keeps rising, the House's reading stands, and the carpet exponents on record are low by several hundredths — every miss in CASE-25 changes sign.

Scored 2026-09-29 — confirmed

Scored by the codifier from RESULTS066.md (aa9870a), as written.

Carpets turn over: 10 of 10. y₇₂₉ < y₂(061) on every carpet — H −0.0071, U −0.0155, U′ −0.0375, T −0.0200, V −0.0086, A −0.0289, B −0.0015, X −0.0077, Y −0.0076, Z −0.0134 — raw and against y₂ + δ (δ = −0.00002). B's drop is inside its error (±0.0050); the other nine are not. The square keeps climbing: y₇₂₉ = 0.7449 ± 0.0010, inside the sealed band 0.734–0.755. Gates: median carpet error on y₇₂₉ is 0.0052 (< 0.007); the control holds — only the square's y₈₁ − y₂ (+0.0048) exceeds three combined errors, one lattice against an allowance of two.

Reported readings. Same-method comparison (y₇₂₉ against y₈₁, both from thresholds): 9 of 10 carpets fall, B rises by 0.0044 inside noise; the square rises 0.0016. The forecast (2y₂ + s₁)/3 lands within 0.01 on 8 of 10 (H exactly; U′ off by 0.030, V by 0.008). At L = 6561 the carpet sequence keeps falling toward the measured exponents on eight of ten (H 0.5693, U 0.5372, V 0.5120, A 0.4647, B 0.5384, X 0.5557, Y 0.5483, Z 0.5355; T and U′ flat within error) — still above every measured 1/ν.

What it settles. PN-002's final House read y₂ > y as convergence from below and concluded that the carpet Monte Carlo is low by 0.03–0.08. That reading is falsified: the rise from g = 1 to 2 was the small-L term, and at larger L the block sequence turns and descends toward the measured values. The record's falling pair slopes were right. The one-block "miss" of CASE-25 is the first term of a slowly converging sequence — if λ_L ≈ A·L^{1/ν}, then y_L − 1/ν ≈ ln A / ln L, so the one-block miss is ln A / ln 9 for a pattern-specific amplitude A, with corrections-to-scaling on top at small L (the rise then fall). A non-universal amplitude would explain why no cut count predicted it. That is a reading for the next seal, not a result of this one.

Pointers corrected 2026-10-05 (codifier; no sealed field, score or number is changed). The script: line was written with the seal, before the script existed, and read: script: R27_tree_groups/carpet_perc/ (to be written by Lab_Tree_Model). It now names what was written and committed: R27_tree_groups/carpet_perc/seq066.py. Found by sync_cases.py's record-field check (docs/REPO-AUDIT.md 5.4).