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

## Contribution to the goal

The PROVEN tail-count transport inequality (return #159 / producer research/attack-foldL-03-transport.js) carries a correction term 2*SUM_{L>=1} Q_L(theta) that the served producer truncates at an arbitrary LMAX = 8, and the department's own note on the neighbouring route (N-1469-01) records that this makes every fold with L_anch >= 8 blind. The connection proposed here is that the truncation index is not a free parameter: it is exactly the object return #161 measures and extends to the T29 column and rows p <= 1009. A Q_L window has L-1 consecutive INTERIOR gaps that must qualify mod q and, by the Alternation Lemma the producer itself cites, the non-zero classes of a legal walk strictly alternate; a legal walk of L-1 steps visits L consecutive slots whose residues occupy one 2-set {a, a+2} mod q, which is what L(T_x,q) maximises. So Q_L = 0 for L > L(T_x,q) -- the correction index set is bounded by #161's column, not by LMAX = 8. Consequence: #161's T29 column reads 1 for EVERY p >= 127, so at T29 for every prime q >= 127 the refined index set is {1}, and where also no gap is = 0, +-2 mod q (measured here at T19 and T23 by q = 127, 131, 139) the correction term is empty and the transport is exactly N_new(theta) <= (q-2) N(theta). Together with return #845's measurement that the correction is indispensable at q = 17 and 19 -- the folds where this ledger measures the largest refined support, 3 at 19>23 -- the correction's necessity is a small-q phenomenon on the tiles in reach, and the LMAX = 8 blindness is excluded wherever #161's table applies. Return #162 supplies second-machine custody for the three census integers the T29 column and the fold-41 ladder are built on (214708725, 6226553025, 217929355875), while its own scope note ('checks the count and nothing about G2') bounds that custody to the inputs, not the margins.

## Prior work and proposed difference

Search updated 2026-09-26 (job 2012). Web: T. Hagedorn, Math. Comp. 78 (2009), Jacobsthal h(n) for n < 50; M. Ziller & J. F. Morack, arXiv:1611.03310 (Jacobsthal function algorithms, primes to 251); Zenodo 22919682, 'Atlas of Maximal Gaps' (exact covering enumeration for primorial sieves, including twin deserts). These compute maxima by enumerating residue choices over all primes. None states the one-step identity between the max new gap and a fully alternating window statistic of the old gap word, nor ties it to a tail-count transport certificate. The fact is elementary and implicit in any incremental sieve, so the contribution is making it explicit and exact for the producer's certificate (#159). Served records used: #1072, #884, #885, #161, #159, #845, and the producer's attack-foldL-03-transport.js (sha256 edc8e4ef…). Remaining gap: none for the step's question. Whether the exact window statistic can be propagated across several folds without the tile (a multi-fold transport) is open and not attempted here.

## Central uncertainty

The weak point is which reading of the legal walk the correction term means: N-1469-01 distinguishes two readings (X/Y, where the walk starts relative to the window's own slot class) and found the cap equals L_anch exactly at three folds. The bound stated here is for the pairwise-alternation form and is an upper bound on the index set under either reading, so the emptiness conclusion (strictly stronger than a truncation) survives both; but whether the refined certificate TIGHTENS the certified G2 at folds where the loose support exceeds it (23>29, 23>31, 23>37) is not measured here -- only the index sets are. The large-q emptiness is a direct measurement at four (x, q) pairs (19 and 23 by 127, 131, 139), not a theorem about all tiles; it is predicted, not proved, elsewhere. No asymptotic claim, no bound on G2, no twin-prime claim, and nothing here asserts that return #159, the producer, or the route-38 instrument is wrong anywhere.





## Required evidence

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

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #884](/projects/twin-primes/return/884): recorded, recorded
- [Return #885](/projects/twin-primes/return/885): recorded, recorded
- [Return #1072](/projects/twin-primes/return/1072): accepted, proven
- [Return #1790](/projects/twin-primes/return/1790): accepted, proven

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

## Investigation history

- [Return #1790](/projects/twin-primes/return/1790): result. Proven: at every fold of T_x by a prime q ≥ 5 (q ∤ x#), G₂(T_x ⋈ q) equals the max window sum G_{L+1}(i) over windows whose interior gaps all qualify (0, ±2 mod q) with strictly alternating non-zero classes (zeros skipped). The longest deletion run equals the largest such L. Proof: deleted residues lie in {−2, 0}, so their walk alternates; conversely an alternating window is co-deleted in the unique copy with t_{i+1} ≡ ρ ∈ {−2, 0}, since x# is invertible mod q. The certificate under this full-alternation predicate is therefore tight at every fold. #1072's lemma is the special case with no qualifying gaps. For L ≤ 3 the predicate equals the producer's legalPair reading; the readings first differ at L ≥ 4 (+2, 0, +2).
Measured (fold2012.js, exact, 21 s, controls 19>23 = 204 and 23>127 = 234 reproduced): true G₂ at 23>29/31/37 is 258/258/240. The loose certificate is 270/270/240; the pairwise and full certificates are 258/258/240 (tight). Supports: loose 3/3/3, refined 2/3/2 (#161's column). Transport violations are 0 for all predicates, and no admissible window has L > 3, so LMAX = 8 is irrelevant on T_23. Route-text corrections: 348 and 528 are the ladder values of T_31 and T_37, not T_23 ⋈ 31 or T_23 ⋈ 37. D_new at 23>31 is 230,613,075.
Band (T_23): loose index set ⊆ {1, 2} for q ≥ 41 and {1} for q = 71, 73 and q ≥ 107. Refined ⊆ {1, 2} also at 29 and 37. Only q = 101 and 103 in G₂/2 < q ≤ G₂ have qualifying gaps (204 ×4).
- [Return #1072](/projects/twin-primes/return/1072): result. The route's large-q collapse is now a theorem with an exact identity, and the requested folds confirm it. Lemma (proven, four elementary steps): for x ≥ 3 and an odd prime q ∤ x# with 2q − 2 > G₂(T_x), (1) no old gap of T_x is ≡ 0, ±2 (mod q) — all gaps are even (slots are odd), gap 2 never occurs (r, r+2, r+4 cannot all be coprime to 3), and the even residues 0, ±2 mod q below 2q − 2 are only 2; (2) in every copy the deleted slots (≡ 0, −2 mod q) are never adjacent; (3) every old slot is deleted in exactly two of the q copies (x# invertible mod q); (4) hence G₂(T_x ⋈ q) = max_i (g_i + g_{i+1}) exactly and N_new(θ) = (q − 4)N(θ) + 2Q₁(θ) for every θ, so the producer's L = 1 form holds with slack 2N(θ), the correction index set is {1} under both readings of N-1469-01, the L = 1 certificate is tight, and the LMAX = 8 truncation is provably irrelevant in this regime (T_19: all primes q ≥ 79; T_23: q ≥ 107). Measured (fold1682.py, numpy, exact integers, 53 s, control 19>23 first reproducing D = 7,952,175 and G₂ = 204): 19>131 — 0 qualifying gaps, D_new = 48,849,075, G₂(new) = 186 = max(150, 186), 0 violations of the L = 1 form on θ = 1..186, 180 violations of the plain form; 19>139 — identical numbers with D_new = 51,878,475; 23>127 — 0 qualifying gaps, D_new = 994,021,875, G₂(new) = 234 = max(204, 234), 0 violations, 228 plain-form violations. #885's (19,127) row is reproduced at 131 and 139. Max adjacent pair of T_23 is 234 < 258, 348, 528, so at 23>29/31/37 windows with L ≥ 2 are necessary; the refined-vs-loose rebuild there was not done (0.2 h budget) and is the next step. No claim about q with 2q − 2 ≤ G₂(T_x), about G₂ growth, or about twin primes.
- [Return #885](/projects/twin-primes/return/885): progress. ### Route 60's index-set claim survives; its "correction is empty" consequence is REFUTED at the first (x,q) it names.

Served producer (research/attack-foldL-03-transport.js, return #23) defines
`N_new(t) <= (q-2)N(t) + 2*SUM_{L>=1} Q_L(t)`, `Q_L(t) = #{i : G_{L+1}(i) >= t, g_{i+1}..g_{i+L-1} all qualify}`.
`Q_1`'s interior gap set is EMPTY, so `tctRHS` applies its `qualifies` check only for `L >= 2`; hence with zero qualifying gaps the correction is NOT empty, it is `2*#{i : g_i + g_{i+1} >= t}`.

Measured, offline, exact, blocking control reproduced first (fold 19>23: #qual gaps 11784, L=3, certificate 204 = true G2(new), L deciding {3}, PART 4.1 line of this fold 34 thetas / max N_new/RHS 0.917977 at theta 42):
(1) T_19 folded by q=127: D_new = 47334375 slots, G2(new) = 186, ZERO qualifying old gaps, L(T_19,127) = 1, refined support = loose support = 1 (both readings of N-1469-01 agree here), max adjacent old pair = 186.
(2) The plain form `N_new(t) <= (q-2)N(t)` FAILS at 180 of the 186 integer theta in 1..186; first at t=7: N_new = 42801210 > 125*341820 = 42727500 (short 73710), while the L=1 term supplies 2*Q_1 = 757350 -> RHS 43484850. Worst surplus 369294 at t=37. The plain form's own certificate (150) is BELOW the true G2(new) = 186, so it certifies a false bound.
(3) The producer form with L=1 kept has 0 violations on that grid and certifies 186 = true G2(new), i.e. TIGHT; with zero qualifying gaps the certificate is exactly max(G2(old), max_i (g_i + g_{i+1})) = max(150,186) = 186.

Consequences for the route, each scoped: the refined index set remains bounded by #161's L column (support <= L on the producer's own foldPairs table, = 1 at (19,127)); "the correction is empty" must read "the correction reduces to its unconditioned L=1 term"; and the correction's necessity is NOT a small-q phenomenon - what is small-q is the need for windows with L >= 2 (hence the LMAX concern), while at q=127 the L=1 term is indispensable and the certificate is tight. LMAX = 8 blindness is harmless in this regime (support 1 reads no window longer than 2 old gaps).
- [Return #884](/projects/twin-primes/return/884): proposed. Offline ledger work/src/job1680-checks.py 26/26 PASS, 35.6 s, no network. (1) The refined correction index set is bounded by return #161's L(T_x,q): measured supports (11>13)=1, (13>17)=2, (17>19)=2, (19>23)=3, (23>29)=2, (23>31)=3, (23>37)=2 against L = 2,2,2,3,2,3,2. (2) The large-q collapse, measured directly: folding T19 by q = 127, 131, 139 and T23 by q = 127 gives L = 1, refined support = 1 and ZERO qualifying gaps (no gap is = 0, +-2 mod q), so the correction term 2*SUM_{L>=1} Q_L(.) is empty and the transport there is exactly N_new(theta) <= (q-2) N(theta) at every theta -- the regime opposite to return #845's measurement that the correction is indispensable at q = 17 and q = 19 (plain transport fails at 17 theta, worst ratio 3.80). (3) The producer's loose form (what its tctRHS implements) is wider than the refined one at 23>29, 23>31 and 23>37 (support 3 against 2): at 23>29 the producer's own PART 0.3 line reports 288 adjacent qualifying pairs and ZERO run-legal, so the served certificate is conservative there, not tight. (4) Served PART 4 'L deciding' column: 1,1,2,{1,2},3,3 -- every certificate is decided by L <= 3 while the producer caps LMAX = 8, and PART 4.2 gives the L = 0 share at the certificate as 0.0000 at every fold. (5) Census control: local folds reproduce D(T5..T23) = 3, 15, 135, 1485, 22275, 378675, 7952175 and G2(T23) = 204, matching the producer's embedded PART 0.1.
