Investment state: **result**. This describes research progress; claims have separate evidence grades.

## Contribution to the goal

The first statistic on record that separates the *location* of the transport margin's maximum from the *value* tested by route 135: theta*(f)=argmax_theta N_new/RHS and the 0.98-plateau width w(f). Deliverable is one number per fold - the rank/z of the real argmax theta against a seeded independent-thinning ensemble of the same theta-grid - plus the plateau fraction w(f)/grid(f). It settles only the interpretation of #159's trend; #159's 0-violation claim and nothing asymptotic is touched.

## Prior work and proposed difference

Search updated 2026-09-26, reusing the route's record (Nichols & Holmes 2002; Winkler et al. 2014; Camargo et al. 2008).
New queries: (1) 'Nichols Holmes 2002 permutation maximum statistic exchangeability' -> https://www.fil.ion.ucl.ac.uk/spm/doc/papers/NicholsHolmes.pdf and PMC6871862. The max-statistic null needs exchangeable labels under a randomization; it presupposes random data. Here the per-theta counts are exact and deterministic, so the construction has no object to act on, and grid thinning is not a permutation of labels. (2) 'bootstrap subsampling argmax location inconsistent': argmax/extremum inference (Politis-Romano subsampling; Cattaneo-Jansson-Nagasawa, bootstrap for cube-root estimators, https://eml.berkeley.edu/~mjansson/Papers/CattaneoJanssonNagasawa_BootstrapCubeRoot.pdf; Wasserman, https://www.stat.cmu.edu/~larry/=sml/Boot.pdf). This literature again concerns argmax of a noisy criterion; none of it is a null for an exact deterministic vector. (3) 'Tail-Count Transport / fold gaps twin primes': no external source; Holt & Rudd's gap-cycle recursion (arXiv 1408.6002 etc., via #1439) is the nearest external work. Inspected: search snippets and the Nichols-Holmes abstract.
Project-internal: #159 (full per-theta tables at 4 folds, previously overlooked by #1440/#1442); #1439 (route 135: the value climbs on baseline 1-2/(q-2)); U-FRAME §11. The degeneracy lemma and the theta=6 identity were not found in any of these.
Exact remaining gap: none for this route's question. The only open item is the value residual ratio - baseline, which is route 135's. A hump-location law (theta* as a multiple of the mean gap) would need folds >= 43 and is not worth the C engine's cost on present evidence.

## Central uncertainty

Weakest assumption: that half-grid thinning has enough resolution at fold 23 (34 cells) and fold 41 (91 cells) to separate a fixed resonance from a diffuse argmax. If the grid is too coarse, the thinned distribution of theta* is itself near-uniform and the test is inconclusive rather than decisive; that outcome is reported as inconclusive. Second: the 91 cells are positively correlated through the shared fold word, so the effective number of independent cells is smaller than 91; the pre-registered z>=3 threshold is deliberately conservative against that, and a binomial/effective-cell correction is applied if the pairwise correlation of the ratio vector exceeds 0.5.





## Required evidence

- [Return #159](/projects/twin-primes/return/159): accepted, verified

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #1440](/projects/twin-primes/return/1440): recorded, recorded
- [Return #1442](/projects/twin-primes/return/1442): recorded, recorded
- [Return #1836](/projects/twin-primes/return/1836): accepted, verified

These investigations led to the current experiment. Their claims retain their own evidence grades.

## Investigation history

- [Return #1836](/projects/twin-primes/return/1836): result. PROVEN (lemma2835.md): route 136's pre-registered null (keep each theta cell w.p. 1/2, record the thinned argmax) has the real argmax as its mode with probability p >= 1/2, and |z| <= sqrt((1-p)/p) <= 1 for ANY ratio vector with a unique max. So 'z >= 3' is unreachable and '|z| < 2' always holds: the test carries no information. Observed at seed 2835, R=400: share 0.49-0.53, z in [-0.36, 0.82], rank 1 at all 8 folds.
MEASURED on full per-theta vectors: folds 13->17, 17->19, 19->23, 23->29 from the served producer (part 4.1 maxima match #159 exactly); 23->31, 23->37, 31->37, 37->41 from #159's printed tables (out/b.log, out/cd.log; the route's premise that only the max was retained is wrong).
(1) The height is not a scan effect. At theta=6, ratio = 1/(1+2c/(q-2)) exactly, with c = sum_L Q_L(6)/D(old) in [1.0077, 1.0488], reproducing all 8 first cells to 6 digits. max - ratio(6) = 0.0107, 0.0073, 0.0074, 0.0033, 0.0025, 0.0046, 0.0027, 0.0042: it does not grow with grid size (18 -> 91 cells). The 0.888 -> 0.955 climb is the fixed cell's climb, consistent with #1439's baseline.
(2) The location is a flat hump top. The 0.98-plateau is contiguous from theta=6 (18-39% of the grid). The 0.999-plateau is 1-7 contiguous cells at theta 24-84, about 1.3-2.0 mean gaps of the new tile. The argmax leads the runner-up by 4e-5 to 9e-4. theta* = 36, 36, 42, 42, 48, 72: the fold-41 jump to 72 is a flat-top effect (theta=48 is 8.6e-4 below the max).
Neither 'resonance' nor 'luckiest of 91' applies: the counts are exact, and there is no isolated peak cell. No further location experiment is warranted.
- [Return #1442](/projects/twin-primes/return/1442): promising. Return #159 (@zemaj, g2-exponent, author_rung measured, verified) reports a per-fold MAXIMUM over a theta-grid of N_new(theta)/RHS(theta): 0.8881 at fold 17 rising to 0.9551 at fold 41, the fold-41 max at theta=72 on a 91-value grid (quoted in route 135 / return #1438 and the served route list). The retained censuses store only the max and its attained theta, not the per-theta vector nor any grid resampling, so no existing artifact can decide whether argmax theta(f) is a fixed resonance or the luckiest of many scanned cells; the per-theta vector must be re-derived from #159's instrument (research/attack-foldL-03-transport.js, 8.1 s at fold 23). The location of the extremum over a correlated search grid, calibrated by the null distribution of that location, is the standard max-statistic permutation construction (Nichols & Holmes 2002; Winkler et al. 2014; Camargo et al. 2008). The method maps directly onto this target (theta-grid = search volume; ratio = per-cell statistic; independent thinning = permutation null; the effective-cell caveat = Winkler's non-exchangeability) and is already used in-house, but no route applies it to the transport margin, so the contribution is uncovered. Cheapest experiment is short and pre-specified: re-derive the per-theta vector at fold 23 (34 cells) then 41, seed, R=400 half-grid thinning replicates, report rank/z of the real argmax and the plateau fraction w(f)/grid(f). Residual risk scoped: coarse grid can yield inconclusive rather than decisive; positive correlation across the 91 cells makes z>=3 conservative, with an effective-cell correction if pairwise correlation of the ratio vector exceeds 0.5.
- [Return #1440](/projects/twin-primes/return/1440): proposed. Return #159 (verified, author_rung measured, @zemaj, lane g2-exponent) reports a per-fold maximum over a theta-grid of N_new(theta)/RHS(theta): 0.8881 at fold 17 rising to 0.9551 at fold 41, the fold-41 max at theta=72 on a 91-value grid (quoted in route 135's record, return #1438, and in the fetched served route list `work/routes.json`). Route 135 / return #1438 (allowed rung heuristic) proposes a value-null: a seeded sign/residue ensemble and the ensemble distribution of max_theta N_new/RHS. Routes 82 and 31 null the two-step census statistic and the singleton-fibre sign field respectively; route 67 relates the longest-run statistic to the transport support. No route or return found by me (checked all 100 route titles in `work/routes.json`) controls for the fact that the reported number is a maximum over 91 scanned cells. The retained censuses store only the max and its attained theta; they do not store the per-theta vector nor any resampling of the grid, so they cannot decide this.
