Run 074 read H, the standard 3×3 carpet, at five sizes a factor of 3 apart (L = 81, 243, 729, 2187, 6561; generations 4–8) and found, post hoc, that the skewness steps alternate — small, large, small, large — with the large steps ending at even generations. The width's pair slopes from run 051 (0.539, 0.580, 0.529, 0.539) alternate in the same phase through the middle three, with equal end slopes. Both rest on one set of 8,000 samples per size and are about two errors. Drawn again with fresh seeds and six times the samples, does a period-two term in generation come back, in the same phase, in both the skewness and the width?

Sealed 2026-09-30, after run 074 was scored and before any samples079 file existed (codifier). Known at sealing: everything in samples051, including the pre-seal glimpse below.

Pre-seal glimpse, stored data only (seed 51, 8,000 samples per size). Computed from flow074.json and perc051.json before the clauses were written, and recomputed by the second reader: - H: Δ⁴g₁ = +0.661 ± 0.283 (2.3 errors, per-file errors in quadrature; cross-file correlations of the seed-51 samples are at most 0.017); Δ⁴g₂ = +1.37 ± 0.84. - H: Δ⁴ln w = −0.169 ± 0.102 (1.65 errors). - Square: Δ³g₁ = −0.089 ± 0.139; Δ³ln w = −0.002 ± 0.054. No sign of a period-two term in the instrument.

Why Δ⁴ and not run 074's contrast. Run 074's post hoc contrast (1, −2, 2, −2, 1), +0.265 ± 0.127, does not cancel a quadratic trend: it reads four times the quadratic coefficient, and H's g₁ curve is bent. The fourth difference cancels any cubic, at about 11% less power per unit of alternation. The 074 contrast on the fresh data is reported beside it.

What this run is for. It is the replication the 074 inkling needs before anything is built on it. A period-two correction in generation is what a negative subleading eigenvalue of the scale-3 renormalization gives, and it matters beyond H. Run 073's "hump" (a rise from 81 to 729, then a turn, a plateau or a climb to 6561) was read on factor-9 carpets at three consecutive generations of their own scale. A correction that changes sign every generation, sampled at three consecutive generations, reads as a rise then a fall: it would make a hump out of steps with no hump in them. If A and B both confirm, hump line 3 (the block-boundary mechanism) is sealed with this term in it. If A is falsified with |Δ⁴g₁| under 2 errors, the hump stays a property of the curve and line 3 goes ahead as drafted; a reverse-phase signal does not clear it.

What it cannot do. It cannot see past 6561: H9 (L = 19,683) needs about 2 GB per thread plus 1.9 GB shared, which exceeds the M5's 16 GB at any thread count that would finish it. It cannot by itself separate a physical period-two term from one written into the graphs, but no parity enters the construction: gen.py builds every generation as the same Kronecker power of one 3×3 generator with the same left and right terminals, and the second reader rebuilt H4–H8 in memory with gen.py's logic and reproduced every .bin byte for byte (H8.bin included, despite its later timestamp). Control (iii) covers the estimator. It does not test the factor-9 carpets.

Power, stated. Scaling run 074's errors by √(8,000/48,000), the fresh Δ⁴g₁ error is about 0.116 and the fresh Δ⁴ln w error about 0.042. If the true values equal the glimpse, A sits near 5.7 errors and B near 4.0. Allowing for the winner's curse (a 2.3-error post hoc find is likely an overestimate) and for the trend bias below (about −0.06 on Δ⁴g₁, half an error), at half the glimpse A sits near 2.3 errors (confirmed with probability about 0.6) and B near 2.0 (about 0.5). Under no alternation, A is falsified with probability about 0.84 and confirmed with probability about 0.02. Control (iii) is permissive (each check passes with probability about 0.997 under a clean instrument) and weaker per unit of alternation than H's contrast; it catches only a gross instrument effect.

The reductio this seal passed first. By construction: each clause reads only fresh samples, and the fourth difference is zero on any cubic trend in generation, so neither the rise of g₁ nor the steady fall of ln w can produce it. A monotone geometric approach g₁ = g∞ − c·rᵏ with c > 0 gives Δ⁴g₁ = −c(1 − r)⁴: negative for every r in (0, 1), smaller in magnitude than c, and about −0.06 at r = 1/3, c = 0.3. The trend biases A against confirmation, not toward it. A non-monotone approach has a bias of unknown sign, bounded by the size of the correction; the same holds for ln w. Negation: if the 074 pattern was noise, the fresh Δ⁴g₁ is centred at or below zero and the falsifier fires with probability about 0.84. What a lone clause decides is in the falsifier line. Cost: about 5 hours on the M5 at 9 threads (H8 about 3.9 h, H7 about 0.5 h, the square about 0.6 h, from run 051's times); peak memory about 2.6 GB (about 0.26 GB per thread on H8).

Reported, not scored. - The fresh g₁, g₂, w and p* (median) of every file, with errors. - H's four steps in g₁ and four pair slopes of ln w (each a 1/ν estimate), with joint errors; Δ⁴g₂. - Run 074's contrast (1, −2, 2, −2, 1) on the fresh data. - The fresh contrasts beside the seed-51 ones, and their inverse-variance combination. - A calibration: each fresh H file split into six consecutive 8,000-sample blocks, Δ⁴g₁ per block, and the block spread over √6 set beside the bootstrap error. - The square's 1/ν.

Scored 2026-10-01 — falsified

Scored by the codifier from RESULTS079.md and parity079.json (25ed684), as written.

Controls first — all three pass. - (i) samples051 reproduces exactly (difference 0). - (ii) Every count is as sealed, and every x lies in (0, 1). - (iii) The square shows no period-two term: Δ³g₁ = −0.066 ± 0.084 (−0.79 errors) and Δ³ln w = +0.009 ± 0.031 (+0.30).

(A) Falsified. Δ⁴g₁ = −0.070 ± 0.130 (−0.54 errors), below +1 error. The skewness alternation of run 074 does not come back on fresh seeds. |Δ⁴g₁| is under 2 errors, so this is not a reverse-phase signal.

(B) Falsified. Δ⁴ln w = −0.0085 ± 0.043 (−0.20 errors), above −1 error. The width does not alternate.

The run is falsified on A, as the falsifier line says, and B agrees.

What it settles. Run 074's alternation was noise in one set of 8,000 samples. - The fresh steps in g₁ are +0.097, +0.064, +0.062, +0.020: one rise, front-loaded and concave from 81. That is the shape run 074's clause B predicted and could not resolve. - Against the fresh values, the seed-51 H5 and H7 sat about 1.7 and 2.1 combined errors low. The 074 pattern was two low draws at the odd generations. - The combined contrasts sit within one error of zero: Δ⁴g₁ +0.06 ± 0.12, Δ⁴ln w −0.03 ± 0.04. - The fresh pair slopes of ln w fall steadily: 0.553, 0.550, 0.541, 0.533. On H, 1/ν is still drifting down at 6561. By the estimator note, the five-size value 0.5445 ± 0.0014 reads high.

Per the seal, hump line 3 (the block-boundary mechanism) goes ahead as drafted, with no period-two term in it.

Calibration, reported. The six-block spread over √6 is 0.090 for Δ⁴g₁, against a bootstrap error of 0.130, and 0.047 for Δ⁴ln w, against 0.043. The bootstrap error is not too small on either; on g₁ it may be conservative.

Lab notes. The analysis script was run three times on the complete files. Two code errors (a type error, then a JSON write) were fixed between runs, with no change of method, and the second and third runs printed identical numbers. All of this is recorded in RESULTS079.