P-054's look found that on T and X at L = 729 the sizes of invasion bursts — the runs of accepted links from one forced link to the next, growth from the whole left side — follow a power law with a ripple about it whose period is a factor of about 55–64 in size, T's twice X's and none on the square. Is the period the hierarchy's cell factor, and does the ripple survive a point source? H (#17, the Sierpinski carpet's second generation; gen.py's H6 has L = 729) is self-similar at every factor of 3 in side, so its cell factor per level is 8 where T's and X's is 64. On H, with the full-side source, is the ripple's period 8 rather than 64? And on T, does a point source keep the peak at the 1,400 bin and the trough at the 200 bin?

Sealed 2026-10-03, after P-054's look and before any draw on H or with a point source (codifier). Known at sealing: P-054's look (commit 334a94d; probes/P-054-invasion-look/), reproduced by the sighted pass on the look's own program with identical md5s: on the codifier's RNG, 400 draws each of T and X at 729 with the full side, r at the 25 bin 1.702 (T), 1.139 (X); at the 1,400 bin 1.718 (T), 1.319 (X); at the 200 bin 0.783 (T), 0.860 (X); T's trough 0.745–0.787 over 130–360; the square's largest |r − 1| 0.104 over the fit bins at 200 draws; burst exponents 1.678 (square), 1.707 (T), 1.710 (X). That look is the pre-seal glimpse. Nothing about H, nothing about a point source, nothing on perc.c's RNG, and F_P has been computed on nothing.

Why H decides it. Every census pattern keeps 64 cells per 9 × 9 block, so T, X and H all have a cell factor of 64 per base-9 level; the look's period of ~55 (bins 1.37 apart, so 40–75) is consistent with 64 and with nothing else the three share. H alone is also self-similar at base 3, with 8 cells per base-3 level. If the ripple is the hierarchy's at every level of self-similarity, H's bursts ripple with period 8 and the period-8 component of ln r dominates the period-64 one. If the ripple is some other property of the 9 × 9 block (its seams, its 84 links), H shows the same period as T and Q is small. The statistic is a ratio of Fourier magnitudes and not a value at one bin, so it does not assume where H's peaks fall (the sighted pass's objection to the first draft, which had assumed a peak at 200). The point source separates the holes from the source: a line source has a geometry of its own (bursts near the side are shallow and wide), and the square shares that source and shows nothing, but a point source removes the question; control (v) checks that a point source does not itself bend the square's bursts.

The codifier's lead. Q on H near 3, with H's peaks at the 25, 200 and 1,600 bins with amplitudes near 1.3 (smaller than T's, since H's holes are smaller at each level); T's point source keeps the 1,400 bin at ≥ 1.5. Reported beyond the clauses: Q on T and X is below 0.5 (period 64 dominates); the point source raises the ripple's amplitude on T and X; the burst exponents of T, X and H agree within 0.05 and the square's is lower by 0.03.

Power. At 400 draws the look's 200 and 1,400 bins held 3,138 and 1,675 bursts on T; at 2,000 draws the block errors of r are about 0.006 (200 bin) and 0.02 (1,400 bin) by the look's four blocks of 100, under the 0.05 bar; the point source's counts are unknown until drawn. Q's error is not known in advance; the 0.2 bar is the scorability condition. Margins: 1.5 against 1.0 on Q; 1.30 against 1.10 and 0.90 against the look's 0.78 on the point-source r. The block error holds the pooled fit fixed and leaves out the fit's own uncertainty; that is accepted and reported.

Controls, all before any scored quantity is read. (i) Reproduction, exact. perc_inv with the full side on T3 with seed 8302 and C3 with seed 8305, 1,000 samples: k_inv equals column 0 (k) of samples083/T3.txt and C3.txt on 1,000 of 1,000 lines each. One mismatch withdraws the run. (ii) The square is flat. sq729 with the full side, 500 draws: every r_i on the fit bins with at least 300 bursts lies within 0.15 of 1 (the look at 200 draws: 0.104). A failure withdraws the run. (iii) The look reproduces on the lab's RNG. T with the full side: r at the 1,400 bin ≥ 1.30 and at the 200 bin ≤ 0.90 (the look: 1.718, 0.783). A failure withdraws the run: the premise is not on the instrument. (iv) The point source is a source. On every point-source configuration every sample crosses (the heap never empties; on these connected graphs that is automatic and is asserted, not tested), and the point-source median crossing point on T is above the full-side median (reported as a sanity reading, not a condition, since the seeds differ). (v) A point source does not bend the square. sq729 with the point source, 500 draws: every r_i on the fit bins with at least 300 bursts within 0.15 of 1. A failure makes clause B unscorable, not the run withdrawn.

Reported, not scored. All 27 r_i per configuration with errors; F_P for P = 4, 5.7, 8, 11.3, 16, 22.6, 32, 45, 64, 90, 128 (powers of √2) on every configuration, with Q and its error; the burst exponents with errors; the record counts per sample; the point-source cell's degree per graph; the crossing-point median and 20–80 width per configuration, the full-side ones beside run 085's medians (T 0.78551, X 0.72175 at 729) and H's beside run 079's record at 729; the seam share of records on T, X and H as in the look (row seams, relative to the seam's share of links; the square has no block file and is omitted).

The reductio this seal passed first. By construction: r_i is a ratio to a line fitted on the same window; the sighted pass computed that the fit moves a period-8 ripple's r at the 200 bin from 1.339 to 1.372 and a period-64 one's from 0.742 to 0.715, so neither period is absorbed. Q compares two magnitudes on the same window (3.35 periods of 8, 1.68 of 64), so leakage from a period-8 signal into F_64 is the thing that could make Q small for the wrong reason; the reported F_P over the full set of periods shows it if it happens. Negation: if the ripple is the source's, H with the full side shows T's period (the source is the same side) and A falsifies; if the ripple is the 9 × 9 block's and not the hierarchy's, A falsifies the same way. Invariance: k_inv equals k on every draw (control i). Cost: 15,000 draws at a tenth of a second: under half an hour on one thread.

Reviewed before commit by a sighted pass and second reading (the cloud and a Code chat, 2026-10-03), whose edits are folded in: the burst definition made the look's (the first draft's was off by one and would have failed control ii at 0.131); 27 bins with their edges; control (ii)'s margin widened to 0.15 on bins with ≥ 300 bursts; clause A's T half dropped (it was forced by control iii) and its H half restated as a period ratio rather than a value at one bin (the first draft tested phase as well as period); control (v) added; the point source moved off row 364 (444 in base 9 and 111111 in base 3, the mirror axis at every level, with a hole beside it on X and H) to the left node of rank 300; the draw count corrected; medians compared with run 085's medians; the square named sq729 and its seam share dropped.

Scored 2026-10-03 — unanticipated

Scored by the codifier from RESULTS095.md and bursts095.json (cloud branch, merged ff91fe0), as written.

Controls. (i) exact: 0 mismatches of 2,000 against run 083's k. (ii) pass: the square's largest |r − 1| is 0.092. (iii) pass: T full side r₁₄₀₀ = 1.757 ± 0.014, r₂₀₀ = 0.750 ± 0.007, the look reproduced on the lab's RNG. (iv) holds. (v) fails in one bin of twelve: the square's point source gives r = 1.188 ± 0.113 on [4, 6), the growth's own first bursts from one cell; every other bin is within 0.15.

(A) Unscorable. Q = F₈/F₆₄ on H = 4.363 ± 2.863; the seal required an error of at most 0.2. One block of twenty has F₆₄ near zero and Q = 60; the other nineteen lie between 1.11 and 7. The parts: F₈ = 0.789 ± 0.030, F₆₄ = 0.181 ± 0.031. (B) Unscorable twice over: r₁₄₀₀ on T's point source is 1.365 ± 0.058 against a 0.05 bar, and control (v) failed. (r₂₀₀ = 0.735 ± 0.030 met its bar.)

Neither clause scorable, so neither confirmed nor falsified: unanticipated, as the seal's rule for unscorable clauses requires. Scores are as written.

What the numbers say, which the clauses could not. 1. The substance of A held in every block. F₈ exceeds F₆₄ on H in 20 of 20 blocks (minimum Q 1.11, median 3.8); the difference F₈ − F₆₄ = 0.61 ± 0.04 is fifteen errors from zero. On T and X with the full side the order is the other way (Q 0.52 ± 0.02, 0.32 ± 0.02; F₆₄ = 2.83 and 1.91 against F₈ = 1.46 and 0.61). H's bursts ripple with its base-3 hierarchy's period and T's and X's with their base-9 one. At the two bins the first draft had sealed on, H is flat (0.97, 0.99): the phase-free statistic was the right call and the ratio was the wrong way to write it. 2. The substance of B held on T and not on X. T's point source keeps the peak (1.365 ± 0.058 against the 1.30 bar; 1.882 ± 0.056 at the 25 bin) and the trough (0.735). X's point source loses its period-64 component almost entirely (F₆₄ 1.91 → 0.25, F₉₀ 1.92 → 0.71) while T's stays (2.83 → 2.73). So the ripple is the holes' on T; on X the full side contributes most of what the look saw, and the seal's reading "the ripple is the source's" applies to X alone. A reading with one source cell per pattern; the cells differ in degree (2 on T and X, 3 on H and the square). 3. H's point source raises its period-8 component (F₈ 0.79 → 1.32, Q 6.4 ± 2.0), the direction the lead gave for T and X. 4. The burst exponents with the full side: T 1.706, X 1.711, H 1.706, square 1.672 — the three carpets agree within 0.005 and the square sits 0.034 lower, as the lead said; with the point source they fall to 1.55–1.63, so the exponent is the source's as much as the lattice's at 729. 5. The seam reading reproduces: records on row seams relative to the seams' share of links 0.817 (T), 0.920 (X), 0.818 (H), against the look's 0.80 and 0.90.

The seal's errors, for OPS. A ratio of two noisy magnitudes was sealed as a clause quantity with a block-error bar; a denominator that can touch zero makes the bar unreachable whatever the data say. The difference F₈ − F₆₄, or F₈ and F₆₄ as two conditions, would have scored. The 0.05 bar on a point-source bin with a fifth of the full side's bursts was set from the full side's counts. Control (v) was written over every bin and failed on the one bin a point source must bend; it should have excluded sizes below the source's degree. Three design errors on a run whose compute cost 97 seconds and whose substance is clear; the next version is sealed on the difference, with the smallest bins excluded from (v), and runs in the same time.

What it does to PG-01. L16 stands with a reading, not a result: the burst ripple is the hierarchy's at every level of self-similarity (H), and on T it is the holes'. P-054's lead 1 survives as stated for T and H; its "the ripple's amplitude should not fall with L" is untested. The ln 9 ripple of P-048 has an instrument that sees it at one size.

Lab notes. Cloud, 4 threads, 6–15 s a configuration, 97 s for all eight; perc_inv.c with perc.c's order; the bond-class files for the seam reading added by the lab; six dictionary names added.