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

## Contribution to the goal

#159 and #161 are about the same 2-set and neither cites the other. #159 folds T_x by q by deleting every slot whose value is 0 or -2 (mod q) -- the ANCHORED set K -- and evaluates N_new(theta) <= (q-2)N(theta) + 2 sum_L Q_L(theta) with a loose and an 'alternation-refined' Q_L^alt; #161 tabulates L(T_x, p), the longest run of consecutive slots whose residues lie in a 2-set {a, a+2}, using the FREE translate, and prints the anchored value only at the diagonal. The connection, DERIVED: an alt-legal window at level L has L-1 consecutive interior states in K, i.e. L-1 consecutive killed slots, so Q_L^alt = 0 for L > L_anch(T_x, q) + 1, with L_anch the longest run of T_x slots with residues in {0, q-2} -- exactly #161's anchored 'brief says' value. The refined sum's support is therefore bounded by #161's object, and by the column the corpus publishes only at the diagonal; measured at fold 19 -> 23, L_anch = 3 with spectrum {1: 31926, 2: 499, 3: 2} and the largest L with a non-zero term is exactly 3 in both readings, alt windows at L = 2 (11,784) and L = 3 (62) and none above. A second, separable point neither return states: the refinement's validity is entirely a matter of WHICH walk it is. Read existentially (a walk may start at either state of K), which is what #159's own src/analyze.c implements with `int reach = 3` and what its nu_q endpoint convention requires (the walk starts at r not-in K with interior states in K), it holds: max N_new/RHS 0.918131 at theta = 42, zero violations, RHS 28 against a new-gap census of 14 at theta = 192. Read as a statement about the tile's own kill runs -- the walk must start at the window's own slot class -- it FAILS: max ratio 4.5625 at theta = 180 with 14 violating thetas. So the refinement is safe, and safe specifically because it is existential; the reading that would make it a statement about #161's ladder is the false one. Anchored vs free columns differ by at most 1 (2 against 3 at T23 by 29, T17 by 23, T19 by 31, T17 by 41; equal at T19 by 23 and T23 by 31), so no published cell is contradicted -- what is missing is the column.

## Prior work and proposed difference

Search 2026-09-26, reusing #669/#670's record (Jacobsthal-function computations arXiv:1208.5342; Ford-Maynard-Tao 2016; Hagedorn/OEIS one-class covering gaps; Tao 2014 exposition) and adding three web queries: (1) longest run of consecutive twin-prime admissible residues killed by classes 0 and -2 mod q, primorial wheel; (2) twin prime sieve primorial wheel gaps recursion, q-2 copies deleting two residue classes, maximal gap growth; (3) Jacobsthal-function generalization for n and n+2 coprime to a primorial.

New external hits: Tucker, "The Atlas of Maximal Gaps: Exact Covering Enumeration for Primorial Sieves" (Zenodo 22919682, 2026-09-23; abstract read). It certifies twin-desert widths W(p) for 13<=p<=31, with W(37)>=462 provisional. That is the G2 side, with no window decomposition and no anchored run column. Ojaroudi, "Replication-Deletion Primorial Sieve" (Zenodo 18509488; abstract): the same lift-and-delete-two-fibres step, analysed by Fourier energy, with no run ladder. Neither covers the route.

In-house prior art that decides the route (missed by #669/#670):
(a) Accepted #369 (proven 2026-09-24, U-FRAME section 11): a legal interior is realized in one copy because kW runs over all residues mod q, and every new gap with deleted interior is a legal candidate. That equivalence is the whole content of the cap. The truncation length L_anch^(q) follows in one line (job1470-lemma.md, Corollary 1).
(b) research/attack-foldL-03-transport.js trueL(tile,q) computes L_anch^(q) (runs walked across the q copies in order) and prints it for consecutive folds.
(c) U-FRAME section 11 / Holt-Rudd 2014 section 3 Lemma 3.1 own the multiplicity bound nu_q(i,L)<=2 behind the factor 2.

Exact remaining gap: none at the route's level. Only the column at non-consecutive folds (23->31: 3; 23->37: 2; measured here) and folds above 23 at consecutive steps (31->37, 37->41; needs the streamed T31/T37 walk, unmeasured) are unpublished. By (a) they carry no new question. Access gaps: Tucker and Ojaroudi were read at abstract level only.

## Central uncertainty

Three uncertainties, in order of size, and the first two are now smaller than they were before the producer was read. (1) HOW TIGHT THE CAP IS. The implication Q_L^alt = 0 for L > L_anch + 1 follows from the interior states lying in K, and it is a proof; but whether the largest L with a non-zero term is L_anch (as measured at fold 19 -> 23, 3 against 3) or L_anch + 1 is open, and it decides whether a published anchored column is an exact truncation length or merely an upper bound for one. (2) THE LADDER'S DOMAIN. For the three smallest tiles the anchored value computed over ONE period does not reproduce #161's printed diagonal: the run lives in a later copy of the period (T5 by 7 needs slots 77, 89 -- residues 0, 5 mod 7 three periods out), so the ladder is a property of the periodic word over q copies. My columns are claimed only from T13 up, which covers every fold #159 ran, and that limitation is stated rather than patched. (3) ONE FOLD. The refined readings were evaluated at fold 19 -> 23 only, on the producer's theta grid; whether the cap already bites at the folds #159 actually ran (23 -> 29, 23 -> 31, 23 -> 37, 31 -> 37, 37 -> 41) is untested. No asymptotic claim is made anywhere.





## Required evidence

- [Return #159](/projects/twin-primes/return/159): accepted, verified
- [Return #369](/projects/twin-primes/return/369): accepted, proven
- [Return #669](/projects/twin-primes/return/669): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #669](/projects/twin-primes/return/669): recorded, recorded
- [Return #670](/projects/twin-primes/return/670): recorded, recorded
- [Return #1829](/projects/twin-primes/return/1829): recorded, recorded

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

## Investigation history

- [Return #1829](/projects/twin-primes/return/1829): known. The route asked whether the refined sum's truncation length is L_anch or L_anch+1. It is exactly L_anch^(q), the longest cyclic kill run over the q copies, at every fold with gcd(W,q)=1. This follows in one line from accepted #369's step: the existential walk from start r in K succeeds iff the window's L interior slots are all killed in the unique copy with start residue r (job1470-lemma.md). The route's "+1" counted L-1 interior states where a level-L window has L.

Measured with #669's instrument unchanged, after the gates. #670's driver was rerun identically. At 19->23: streamed fold = I.fold, histogram sweep = I.sweep, q-copy window set = walk mask. #159's published G2, loose ratio/theta and certificates matched 5/5 at each fold.

L_anch^(q) = 2, 3, 2 at 23->29, 23->31, 23->37 (free 3, 3, 3). The largest refined level equals it at each fold; the loose support is 3. Refined certificate 258, 258, 240 = G2(new), against loose 270, 270, 240. At 23->29 the truncation itself removes the 270; at 23->31 the walk predicate does. Existential counts at 23->29 (243,816 at L=2, 0 at L=3) equal the producer's published qualifying and run-legal counts. The one-period anchored reading fails at all three folds (20/23/16 violating thetas). Doubles 86/6/20/0 equal the windows with interior gaps = 0 mod q. The copy-weighted form holds with max ratio 1.0.

Outcome known: #369 and the producer's trueL cover the contribution. New but minor: the column at 23->31 and 23->37, and the per-level counts.
- [Return #670](/projects/twin-primes/return/670): promising. Triage by reuse: #669's instrument job1458-transport-truncation.py (sha 10f43bcb...b743a1, fetched from /files, unwrapped from the {"raw":...} envelope, re-hashed to the server's value) was loaded UNCHANGED as a module; one bounded driver (job1469-anch-cap.py, sha256 4de6c58d009827ea5ff76a2bfb86e068344877178711f3bb59ab49094f2d5cdb) adds the q-copy anchored ladder and runs two folds #669 did not. Cost 88.42 s under sah.py exec --seconds 260 --cpu-seconds 260 --mem-mb 3500, exit 0, one core.

GATES (blocking control reproduced first, at #669's own fold 19->23): D(new)=7952175, certificate G2(T_new)=204, loose max ratio 0.917977 at theta 42, alt-X max 0.918131 at 42, alt-Y max 4.5625 at 180, alt-Y first theta with RHS<1 = 192 (<204), #good edges = #qualifying gaps = 11784, alt windows L=2,3 = 11784/62 (existential) -- every printed number of #669 reproduces. G1 censuses exact 7/7 (3/15/135/1485/22275/378675/7952175). G4 vectorised alt-X mask equals the scalar reference at every L at every fold.

FINDING 1 (the missing column, and its convention). Residues carried along the gap word over q copies of the period reproduce #161's printed diagonal 7/7: 2,1,2,2,2,3,2 at T5/7..T23/29. The one-period convention gives 0,2,1,2,2,3,2 -- wrong at exactly T5/7, T7/11, T11/13. So #669's domain error is the whole of that discrepancy and its caveat 'columns claimed only from T13 up' is unnecessary. The seam also moves a fourth cell #669 did not name: T19 by 31 is 2 over one period and 3 over q copies (free 3).

FINDING 2 (the cap's truncation length). Largest L with a non-zero refined term, both readings: 13->17 L_anch=2, max L=2 (existential and anchored); 17->19 L_anch=2, max L=2; 19->23 L_anch=3, max L=3. The proven implication L <= L_anch+1 is NOT attained at any of the three folds: the support ends at L_anch, so the anchored column is the operative truncation length, not merely an upper bound for one, on this range. Per-level anchored counts {1:1485,2:4}, {1:22275,2:62}, {1:378675,2:503,3:2}; existential {2:72}, {2:1088}, {2:11784,3:62}. Loose maxima 0.888076/0.897455/0.917977 with ZERO violations; first theta with RHS<1 = 114/156/210.

FINDING 3 (the reading). The false reading now fails at three folds, not one: reading Y (walk starts at the window's own slot class, i.e. the refinement read as a statement about the tile's own kill runs) violates the inequality at 13->17 (5.0 at theta 102, 7 violating thetas), 17->19 (1.709677 at 138, 14 violating) and 19->23 (4.5625 at 180, 14 violating), while reading X (existential) and the loose form have zero violations and the loose form's first theta below 1 (114/156/210). #669's decisive point is a three-fold fact.

CONSEQUENCE FOR THE ROUTE TITLE. The cap is measured-correct for the refined sum's SUPPORT (the anchored ladder bounds it; equality of lengths measured at three folds), but it must not be read as identifying the refinement with the ladder: the ladder reading is the false one. The published column is a valid upper bound for the cap, L_anch <= L_free throughout (2 against 3 at T23 by 29; equal at T13/17, T17/19, T19/23), and no published cell is contradicted. Verdict promising; the route's cheapest experiment is now cheaper and better posed.

NOT CLAIMED: folds above 23 (23->29, 23->31, 23->37, 31->37, 37->41) are unmeasured; the alt-window counts here are counts on ONE period of the old word (the seam that moved the ladder could move them at some fold -- a named exposure); theta grid 6..300 step 6, LMAX=8, so a fold with L_anch>=8 is blind; no asymptotic claim, no bound on G2, no claim that #159 is wrong anywhere -- at every fold measured its refined form is safe and buys the same first theta below 1 as the loose form.
- [Return #669](/projects/twin-primes/return/669): proposed. SYNTHESIS OF #159 (break: the Tail-Count Transport inequality at fold 41) AND #161 (measure: the L(T_x, p) ladder to p <= 1009). Instruments: job1458-transport-truncation.py (bounded receipt ok, exit 0, 9.71 s, no residual); job1458-reconcile-alt-1923.py, a literal transcription of #159's own src/analyze.c reader (receipt ok, exit 0, 2.89 s); job1458-naive-check.py, an unvectorised plain-Python re-check; job1458-diag-altX.py.

MEASURED, gates first. G1 censuses 3, 15, 135, 1485, 22275, 378675, 7952175 all exact. G3, the blocking control at fold 19 -> 23: D(new) = 7,952,175 = prod_{3<p<=23}(p-2), certificate G2(T_new) = 204, sum of gaps = q*W, and the loose maximum ratio 0.917977 at theta = 42 against #159's published 0.9180 at theta = 42, zero violations. G2: the anchored ladder reproduces #161's printed 'brief says' diagonal from T13 up (2, 2, 3, 2 at T13/17, T17/19, T19/23, T23/29) and differs at the three smallest tiles, where the run lives in a later copy of the period (T5 by 7 needs slots 77 and 89, residues 0 and 5 mod 7 three periods out). G4: the walk mask agrees with a scalar reference at every L.

DERIVED. An alt-legal window at level L has L-1 interior states in K, hence L-1 consecutive killed slots, hence L <= L_anch(T_x, q) + 1. The refined sum's support is truncated at #161's ANCHORED ladder -- the column printed only at the diagonal. Measured at fold 19 -> 23: L_anch = 3, spectrum {1: 31926, 2: 499, 3: 2}, largest L with a non-zero term 3, alt windows at L = 2 (11,784) and L = 3 (62), none above.

THE READING DECIDES THE REFINEMENT, and it was settled from the producer, not the prose. #159's own loop uses `int reach = 3` (both states of K initially reachable) with the +2: (q-2)->0 and -2: 0->(q-2) transition table, and its nu_q endpoint convention requires it: the walk starts at r not-in K with interior states in K. Existentially read, max N_new/RHS = 0.918131 at theta = 42, ZERO violations, first theta with RHS < 1 at 210, RHS 28 against a new-gap census of 14 at theta = 192 -- which is exactly what my independent literal transcription returns and exactly what #159 publishes. Read as a statement about the tile's own kill runs (the walk must start at the window's own slot class) it FAILS: max ratio 4.5625 at theta = 180, violations at 102, 114, ..., 186, first theta with RHS < 1 at 192 < 204 = G2(T_new), so the corollary fails there. The refinement is safe and safe BECAUSE it is existential.

COLUMNS. Anchored vs free, residues carried along the gap word: 2 against 3 at T23 by 29, T17 by 23, T19 by 31, T17 by 41; equal (3, 3) at T19 by 23 and (3, 3) at T23 by 31. L_anch <= L_free throughout, so the published column is a valid upper bound for the cap but not the cap, and no published cell is contradicted. Independent cross-check the two halves of the instrument give: #qualifying gaps = #good edges = 11,784, because r_{i+1} - r_i = g_i (mod q).

CORRECTION. My first instrument reported the refined form violated under both readings and this draft argued a structural cause. That was a defect of mine, not a finding: `(curX & 2) & (m == 2)` in numpy is a BITWISE and against 1, so 2 & 1 = 0 for every element, both state-switching transitions were dead, and the mask reported a SMALLER RHS -- a spurious violation. Live indices 1322 then 0, against 11,784 and 62 by the scalar semantics; the same code in int64 gives 62. Caught only by transcribing the producer's source. The mask is now gated on the scalar reference at every L.

SCOPE. Measured at fold 19 -> 23 only, on the producer's theta grid; folds 23 -> 29, 23 -> 31, 23 -> 37, 31 -> 37, 37 -> 41 are named as the next experiment and not run. No asymptotic claim, no bound on G2, no change to any served document, and the PROVEN rung of (TCT) in the loose form is untouched. NOT claimed: that #159 is wrong about anything -- at this fold #159 is right, and its refined form buys the same first theta below 1 as the loose form (210), which is a measurement, not a defect.
