The square lattice, measured with run 051's estimator at the same sizes, gave 1/ν = 0.7516 ± 0.0049 in run 053 and 0.7372 ± 0.0049 in run 055 — a shift of 0.014, about twice the standard error of a difference between two independent runs. Is the bootstrap error an honest measure of how far this estimator moves between independent runs, or does it understate it? The answer sets how narrow a sealed bracket can be.

Sealed 2026-09-29 after run 055 was scored and before any replica was sampled (codifier). Known at sealing: the two square values above and every number in runs 051–055. Not known: any replica.

Why this direction, stated because it reverses the first guess. Proposing this run, the codifier guessed the spread would be 0.01–0.02, larger than the bootstrap. Worked through before sealing, that guess does not survive: the difference of two independent runs with errors of 0.0049 each has a standard error of 0.0069, so 0.014 is a two-sigma event, about one in twenty — unusual, not anomalous. The bootstrap resamples exactly the thing that varies between runs (the 8,000 per-sample thresholds at each size), so it should be honest unless samples are correlated. The sealed prediction is the reasoned one. Eight replicas give S about ±35% precision; the prediction band was set wide enough for that.

What the outcomes mean. Confirmed: brackets set from bootstrap errors are sound, 055's arm B miss stands as a small real miss or a fluctuation of ordinary size, and future seals can use margins of about three bootstrap errors. Falsified high: every bracket narrower than ~0.03 sealed in runs 053–055 was narrower than the instrument resolves; the fix is replicas per carpet, not narrower margins. Falsified low: the replicas share state (seed handling), and the estimator's errors must be re-derived before anything else is sealed on it.

---

Scored 2026-09-29 — confirmed

Results: R27_tree_groups/carpet_perc/RESULTS056.md, perc056.json (commit fd18f82, Lab_Tree_Model). Three streams in parallel under nohup, 10 threads; per-sample seeding makes the numbers independent of the thread count.

Mean 0.7411, S = 0.0043, standard error of the mean 0.0015. Beside them, run 053's 0.7516 and run 055's 0.7372.

Correction to the seal. The question line quotes run 053's square as 0.7516 ± 0.0049; its recorded error (perc053.json) is 0.0055. The codifier transcribed it wrongly. The argument is unchanged: the difference's standard error is 0.0074, so the 0.014 shift is 1.95σ.

What it settles. Brackets and falsifier margins sized from bootstrap errors are sound for this estimator; a margin of about three bootstrap errors (~0.015) is what one run can honestly resolve. Run 055's arm B miss (0.007 against U) sits inside that, which is consistent with its unanticipated reading. Carpet B itself is reproducible: the fresh replica gives 0.5166 against 055's 0.5170.

What it exposes, unsealed. (1) A bias, not only noise. The eight replicas agree with each other far better than with the exact answer: mean 0.7411 against 0.75, a gap of 0.0089, about six standard errors of the mean. The three-size estimator reads the square about 0.009 low. Every carpet value carries an unknown bias of the same origin; if it differs between rooms, it enters every between-carpet comparison, which is why the next: asks for it. Run 053's 0.7516 was the high outlier (2.5 S above the mean), not the typical reading. (2) Pair-slope order is a coin flip. The first pair slope exceeds the second in 3 of 8 replicas and falls below it in 5 — the direct evidence behind the withdrawal of the drift readings in 053 and 055 (note of 2026-09-29). Averaged, the second pair slope (0.7474) sits closer to 0.75 than the first (0.7388).