Run 073 read the skewness g₁ of the threshold distribution at three sizes a factor of 9 apart and found it built between 81 and 729 on thirteen of fourteen carpets, then turning back, levelling off or still climbing. The inkling filed with it is that every carpet traces one hump whose depth depends on the carpet. H, the standard 3×3 carpet, was sampled at five sizes a factor of 3 apart (81, 243, 729, 2187, 6561), and its values at 243 and 2187 have never been read. Seen at twice the resolution in scale, is H's curve one rise without a dip, concave from its start?

Sealed 2026-09-30, after run 073 was scored and before any statistic of H5, H7, sq243 or the gasket files was computed (codifier). Known at sealing: H's g₁ at 81, 729 and 6561 from run 073 (+0.144, +0.349, +0.423), the square's g₁ at 81 (pooled over four files, not sq81 alone), 729 and 2187, and the sample counts of every file named above.

What this run is for, and what it cannot do. It is the first of four lines pursuing the hump (P-048): it maps one curve at twice the resolution in scale, on stored data. H is in the levelling group, so it cannot show a turnover. Clause A tests for a crossover dip between the factor-9 points. It does not test for wobble from discrete scale invariance: a log-periodic correction of period ln 3 is sampled at the same phase at every L = 3ⁿ and is invisible here, and alternating steps would need a period-ln 9 component, which H's 3×3 construction does not supply. Clause B asks whether the build-up is concave from 81 (a front-loaded rise) or sigmoidal (slow, then fast), which bears on whether skewness is built from the first generation or needs several to start.

Power, stated. Clause A flags a dip only if a step falls about 0.12 below its trend (late steps near +0.037 against step errors near 0.042); a false flag under a smooth rise has probability about 0.002 per step. Clause B is weak: under a geometric shrink the expected d₁ − d₂ is about 0.05 against a joint error near 0.073, so it is scored at the point value and falsified only by a clear sigmoid onset. The late-curvature test, g₁(2187) against the chord of 729 and 6561, is not scored: the expected bend is about 0.01.

Reported, not scored. H's g₁ and g₂ at all five sizes with the four steps and their joint errors; the square's four values and g₂; the gasket's five values of g₁ and g₂. The gasket's threshold tends to 1 with generation (it is finitely ramified), so its skewness is shaped by the wall at 1, not by bottlenecks inside the lattice, and it is not read against the hump.

The reductio this seal passed first. By construction, caught: a first draft scored d₁ > d₄; with the known rises (+0.205 over 81→729, +0.074 over 729→6561) that holds for almost any 243 and 2187, so it was replaced by d₁ > d₂, which only the unread 243 value decides. Every scored quantity involves 243 or 2187. Negation: a crossover in H's approach (one mechanism giving way to another between 81 and 6561) shows a negative step; clause A is written for it. Cost: minutes, stored data.

Reviewed before commit by a second reader (OPS.md); its edits are folded in: the log-periodic "negation" of the first draft withdrawn as aliased away at L = 3ⁿ; both clauses restated as windows on the two unread values; the control made a reproduction check on H plus the one unread square file; the bootstrap and seed scheme stated; the combination rule stated; and the error scale given.

Scored 2026-09-30 — unanticipated

Scored by the codifier from RESULTS074.md (069aab0), as written.

Controls first — both pass. H4, H6 and H8 reproduce run 073's values exactly (difference 0). sq243 gives g₁ = −0.0187 ± 0.0311, −0.83 combined errors from the square's 729 value.

(A) No dip — confirmed. Steps d₁ … d₄: +0.048 ± 0.048, +0.157 ± 0.049, −0.041 ± 0.046, +0.115 ± 0.048. The lowest is −0.90 of its error, above −2.

(B) Concave from the start — not confirmed, not falsified. d₁ − d₂ = −0.109 ± 0.086: the first step is the smaller, so B fails at the point value, but by 1.27 errors, short of the falsifier's 2. The run is unanticipated.

What the five sizes show, read post hoc. The steps alternate: small, large, small (slightly negative), large. The two large steps both end at an even generation (243→729 ends at 3⁶, 2187→6561 at 3⁸); the two small ones end at odd generations. g₂ does the same, more strongly (+0.08, +0.20, −0.22, +0.21). The alternation contrast g₁(81) − 2g₁(243) + 2g₁(729) − 2g₁(2187) + g₁(6561) = +0.265 ± 0.127 (2.1 errors, from the tabled per-file errors, ignoring the small correlation the joint bootstrap would add); for g₂ it is +0.54 ± 0.38 (1.4 errors). The late curvature reported beside the run is −0.078 ± 0.040. Post hoc and about two errors, so an inkling, not a result: a correction that changes sign every generation, as from a negative subleading eigenvalue of the scale-3 renormalization, would do this. The seal's statement that a period-ln 9 component is "not supplied" by H's 3×3 construction missed that case: (−1)ᵏ has period ln 9 in ln L. The factor-9 carpets are sampled only at every second generation, where such a term is invisible.

Seeds. The sealed range 74001–102000 lay wholly inside the ranges run 073 had used (73001–163000). The seal was written without checking run 073's listed seeds; the codifier's error, not the lab's. The lab kept the sealed scheme and counted the effect: 357 resamples of H4, H6 and H8 repeat run 073's, which touches only the bootstrap errors of those three files and no point value. From here, a seal's seed range is checked against every range listed in the runs it builds on before it is committed.