Run 054 found that the minimum-cut entropy per cut bond, s = ln N / C, orders carpets H, U and V (one cut exponent, 0.631) exactly as their percolation exponents do, s 0.693 < 1.058 < 1.153 against 1/ν 0.547 > 0.524 > 0.477. Is that an ordering of three rooms, or does s set the class? Choose two new rooms at the same cut exponent by s alone — one beyond V, one between H and U — and bracket each one's 1/ν from its neighbours in s before any percolation sample is drawn.

Sealed 2026-09-28 after run 054 was scored and before the candidate pool's s was counted (codifier). Known at sealing: every number in runs 051–054, including the monotone order of all five carpets in s and 1/ν (the T and U′ half of it unsealed and in-sample). Not known: s for any carpet in the pool other than H and V, which generators will be chosen, or any percolation number for them. The brackets come from the neighbours in s alone. No formula is fitted: the bet is monotonicity at fixed cut exponent, tested on rooms chosen for this purpose.

The margins. The 0.03 and 0.02 gaps in s keep each chosen room clearly away from the neighbours it is compared against. The 0.02 gaps in the falsifier are about four bootstrap errors of 053's estimator (±0.004–0.006), so a falsification is a clear miss and not noise.

What the outcomes mean. Confirmed: at fixed dimension and cut exponent, minimum-cut entropy sets the percolation class to the resolution of this estimator, and "the room decides the class" (§43, §45) has a number that predicts, not only orders. Falsified on A: more minimum cuts does not always mean a broader transition — 054's order was a fact about three rooms. Falsified on B: s orders the extremes but not the middle; something else in the room matters at this resolution.

---

Scored 2026-09-28 — unanticipated

Results: R27_tree_groups/carpet_perc/RESULTS055.md, perc055.json (commit 2b12735; the selection, mincount055.json, at 2d76ba6 before any sample). Lab_Tree_Model's commit step failed on a transient tool error; the codifier committed its files unchanged.

At one cut exponent (0.631), ordered by s(4):

What it settles. Arm A is a clean out-of-sample success: the room with far more minimum cuts than any before it has the lowest 1/ν of the five, where s said it would be. Across the five rooms at this cut exponent, 1/ν falls as s rises everywhere except between B and U, which are swapped by 0.007.

Why B is not read as a counterexample, and why it still scores as written. The square lattice, measured with the same estimator and sizes, came out at 0.7516 in 053 and 0.7372 here — a shift of 0.014 between independent runs, about three bootstrap errors. B's miss of U's value is half of that. So the bracket (0.524, 0.547), 0.023 wide, was narrower than this estimator can resolve between runs, which the seal did not know: the 0.02 falsifier margins were set from bootstrap errors alone. The outcome stands as unanticipated; the next: asks for the run-to-run spread before a bracket this narrow is sealed again. This also bears on 053: the V–H gap there (0.070) is five times the square's shift, so its reading stands.

Unsealed, reported. B's s(g) falls with generation (1.151, 0.937, 0.883, 0.870) while every other carpet's rises toward its limit — the only room whose minimum cuts get relatively scarcer as it refines. p∞ = 0.7394 (A), 0.7332 (B), 0.49996 (square). Pair slopes drift on both carpets (A 0.436 → 0.423; B 0.540 → 0.494), as on T and V in 053.

Note 2026-09-29 (codifier), after run 057. The drift caveat above reads the downward pair slopes as a sign that the asymptotic 1/ν is lower. That is an interpretation this record cannot support. With three sizes there are only two pair slopes, and the square control shows their order is not stable: 0.756 then 0.742 in run 053, but 0.723 then 0.772 in run 055, straddling the exact 0.75 in opposite directions. Run 057, with a known answer, found pair slopes that oscillate rather than converge. The fitted value and the outcome stand; the drift reading is withdrawn as a finding.