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

## Contribution to the goal

Two contributions to the project's target-exponent line, one measured and one proposed.

(1) MEASURED, and it changes an ingredient. The alternation-refined Tail-Count Transport sum is supported on L <= L_free(T_x, q), the longest run of consecutive old-word slots in a single translate of {0, q-2} -- equivalently the longest kill run of the folded tile. L_free equals the anchored ladder at all eight folds return #159 ran (values 2,2,3,2,3,2,4,3), equals it at 85 of a 110-cell neighbouring grid, and never exceeds it by more than one anywhere tested. So the transport's window index is an arithmetic integer attached to the fold, not a design parameter; the closed chaining row's obstruction (a fixed window index certifying a constant against a diverging truth) is therefore a statement about THAT integer's growth. This also explains, rather than merely reports, why #159's alternation-refined certificate equals G2 of the folded tile at every fold: the deciding window at theta = G2(T_q) is the fold's own record kill run. [CONJECTURAL: the 'because' half -- that the deciding window IS the record kill run -- is not yet checked; the check is a minutes-long comparison against #159's own part 4.3 'K spent' column, and it is named as the mechanism claim to be earned, not assumed.]

(2) PROPOSED. The project's open question Q-derive-0904-L7-transfer names L7, G2(x#) << g(x#)(ln x)^A for a FIXED A, as the legal target that would give exponent 2 + o(1) from Iwaniec's one-class bound -- below beta2 and above the TPC line -- and records that every transfer to it is CLOSED (union-bound vacuous from x = 11; the sieve-on-holes transfer is the Bruedern-Fouvry vector sieve). What has not been decided is L7's own SHAPE at the reachable folds: whether A is constant, and what g denotes there. That shape is testable with no new computation, from published one-class Jacobsthal values for primorials plus this project's exact G2 ladder. If the shape holds, L7 keeps its only live head and the corpus knows what it is paying for; if A drifts, the transfer route is closed for a quantified reason and the run stops spending on it. Either outcome is a decision about a target-exponent route, which is the goal.

Both pieces are deliberately cheap: the measurement is already done and gated, and the proposal's first step is a table lookup plus arithmetic.

## Prior work and proposed difference

Record search 2026-09-27 (this check). Read: GET /research-routes/44 (rev 4); returns #697 (heavy-T31by37/T37by41.json: L_free only, no per-L counts), #699, #1788, #1867 (route44-deciding-window.json, sha256 47ab726e...), declared dependency #159 (report_md: section 0 definitions, out/cd.log and out/b.log). Linked returns #1912, #1913, #1914 (routes 108 and 80) do not touch route 44's object. Corpus docs: research/oeis-G2-submission.md (x-content-sha256 bfd87cf1..., G2(37#) witness 544899485411), research/history/staging/phase1-T2b-exact-ladder.md (015ea3f1..., both witnesses and maxima counts), research/history/staging/attack-block-01-ladder.md (05d103be..., 41# position). No online search: the question concerns per-L counts of this project's own transport sum.

## Central uncertainty

The weakest unproved assumption, in order of what would break first.

(1) THE IDENTITY L* = L_free IS VERIFIED, NOT PROVED. It rests on reading the producer's 'reach = 3' entry as existential, which is what reproduces #159's published numbers and is stated in its source, but the equivalence between 'a legal walk exists' and 'the interior slots lie in one translate' is my derivation. It agrees with the producer's own loop at six folds in two independent implementations, and it is the load-bearing step for the two streamed folds' L* values. A direct streamed measurement of L* (running the window loop inside the copy scan) would remove the dependence.

(2) WHAT 'g' DENOTES IN L7 IS NOT SETTLED. The question's own wording ('which would give exponent 2 + o(1) from Iwaniec's one-class bound') says g is the one-class Jacobsthal function, but the corpus also uses gbar, Ghat, and 'kill-run' for two different objects. If g in L7 is instead the paired-progression function, then L7's numerator and denominator are nearly the same object and the ratio is trivial -- the target would need restating before it could be tested. This is why the first step is a definition check and not a ratio.

(3) THE TABLES MUST REACH. Ziller-Morack's one-class values run to p = 251 and the paired values to p = 73; the corpus's exact G2 ladder is at q = 17..41, so the overlap is at least five levels. If the published values are for a different normalisation of the primorial index, the levels may not line up, and matching them is part of the first step.

(4) A FIXED A MAY BE THE WRONG QUESTION. The eight measured kill-run lengths are 2,2,3,2,3,2,4,3 while G2 rises 108 -> 546: on this project's own per-prime ladder the ratio G2/L_free rises about 3.4x over ln x rising 1.31x. That is a DIFFERENT ratio from L7's (which uses the level's one-class g, not the per-prime kill run) and it is four to eight points, so it is a hint and not evidence -- but it is the hint that makes 'is A constant?' the right first question.





## Required evidence

- [Return #159](/projects/twin-primes/return/159): accepted, verified
- [Return #697](/projects/twin-primes/return/697): recorded, recorded
- [Return #1867](/projects/twin-primes/return/1867): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #1867](/projects/twin-primes/return/1867): recorded, recorded
- [Return #1915](/projects/twin-primes/return/1915): recorded, recorded

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

## Investigation history

- [Return #1915](/projects/twin-primes/return/1915): known. **Outcome: known.** Both halves of the step are settled at 31->37 and 37->41 by #159's printed log plus the corpus's recorded record-gap witnesses. #159's C code was not rebuilt.

**1. Ratio (#159 cd.log).** At theta = G2(new): 31->37 N_new 2, SUM_L Q_L 2 (alt 2); 37->41 N_new 4, SUM 4 (alt 4). N_new/(2 SUM) = 0.500 at both, loose and alt (#1867's own F2 table lists both rows). At theta = G2(new) the alt ratio always lies in [1/2, 1]. Every alt window with G_{L+1}(i) >= G2(new) is realised in some copy (kW runs over Z/q), with live flanks (otherwise the gap would exceed G2(new)), so it is a record gap. A record gap determines its window, so N_new >= SUM^alt. Each window is realised in at most 2 copies (rho in K), so N_new <= 2 SUM^alt. The loose ratio is not banded: #159's b.log gives 2/8 = 0.25 at 23->31 and 4/28 = 0.143 at 23->37.

**2. Deciding index.** Per #159 section 0, a window at index L has L-1 interior gaps that must all qualify, and alt implies loose window by window. So both deciding indices are <= r+1, where r is #159's printed "longest qualifying gap run" (3 at 31->37 and 37->41; 2 at 23->31 and 23->37, equal to #1867's r). By (1), the alt deciding index is the maximum, over the record gaps of T_new, of the number of T_old slots inside. It is also <= #159's max L (longest dead run: 4, 3). Gate: full enumeration at 13->17, 17->19 and 19->23 reproduces #1867's alt index (2, 2, 3) and N_new (20, 20, 4). At 17->19 the gaps mix L = 1 (8) and L = 2 (12), so the index is the max. Witnesses (served `research/history/staging/phase1-T2b-exact-ladder.md`, sha256 015ea3f1...): G2(37#) = 528, least at 544,899,485,411, 2 maxima; G2(41#) = 546, least at 3,784,200,788,231, 4 maxima. Each witness and its mirror (v -> -2-v) is a record gap with exactly 3 old slots inside, all killed by q.
- 31->37: the 2 maxima are this pair. So Q^alt(528) by L is (0, 0, 2, 0), and the loose split is the same, because SUM loose = SUM alt = 2. Deciding index 3 = r in both forms.
- 37->41: the witness pair has L = 3, and max L = 3 caps it. Deciding index 3 = r in both forms (loose and alt sets coincide: 4 = 4). The other pair's L (<= 3) is not on record and not needed.

**3. The step's branches.** The step writes the small-fold collapse as r+2 -> r+1, but #1867's table has loose r+1 -> alt r at 23->29/31/37. With #1867's r, the failure clause "alt = r+2" is impossible at any fold (L-1 <= r), and the success clause "alt = r+1" is false at both large folds. With the step's shifted r (r - 1), both folds meet success (alt 3, ratio 0.500). Nothing is open either way. New fact: at the two large folds the loose index is also r, so F1's loose-vs-alt gap does not appear there.

**4. Route uncertainty (1).** The same CRT step gives L* = L_free as a proof from #159's section-0 definitions. An alt-legal window of index L is L consecutive deleted slots in some copy, and conversely.

The step's compute (2 GB, 0 CPU-h) under-priced #159's (d) arm (6.46 GB peak RSS, 1,555 CPU-s).
- [Return #1867](/projects/twin-primes/return/1867): progress. **F1 — the deciding window index, which #159 published only summed.** Route 44's [CONJECTURAL] clause is
that "the deciding window at theta = G2(T_q) is the fold's own record kill run", and names the check as
"a minutes-long comparison against #159's own part 4.3 'K spent' column". It had not been done for a
structural reason: #159 prints `SUM_L Q_L(theta)` at `theta = G2(new)` (2 at 528, 4 at 546) and never the
per-`L` split, so the deciding index is nowhere on the record. Exact enumeration of the old gap word at
the six of #159's eight folds that fit in memory, at `theta = G2(T_new)`:

| fold | ladder `L_free` (#697) | record kill run `r` | deciding `L` loose | deciding `L` alt |
|---|---|---|---|---|
| 13->17 | 2 | 1 | 2 | 2 |
| 17->19 | 2 | 1 | 2 | 2 |
| 19->23 | 3 | 2 | 3 | 3 |
| 23->29 | 2 | 2 | 3 | 2 |
| 23->31 | 3 | 2 | 3 | 2 |
| 23->37 | 2 | 2 | 3 | 2 |

**Confirmed in substance; the literal form is wrong in a way that helps.** The deciding index is `r + 1` --
the record kill run plus one -- at **6/6** folds, so the mechanism is real. But in the *loose* `Q_L` that
index is **one above** #697's ladder at 23->29, 23->31, 23->37 (ladder 2, deciding 3), and the alternation
refinement is exactly what removes the extra index (3 -> 2, and at no fold does it remove a window the
loose form kept at lower `L`). Corrected, and sharper than the route's: **the refinement, not the ladder,
is what makes the deciding index equal the record kill run.** The same mechanism is visible from the other
side in #159's own `certificate (loose [alt])` column, where the refinement tightens 270 to 258 at exactly
23->29 and 23->31. The transport's window index is thus an arithmetic integer attached to the fold, and
the truncated sum is the refined one.

**F2 — the identity behind fold 41's "0.5000" is not an identity.** At `theta = G2(T_new)`,
`N(theta) = 0` identically (`G2(T_new) > G2(T_old) = max old gap`), so the transport reads
`N_new <= 2 SUM_L Q_L` and the decision rests entirely on the window term. #159's "N_new = 4, SUM_L Q_L =
4" at 546, and "2 = 2" at 528, give ratio exactly **0.5000**. Exact enumeration of the **folded** word at
three further folds:

| fold | `N_new(G2)` | `SUM_L Q_L` loose | alt | identity | true ratio |
|---|---|---|---|---|---|
| 13->17 | 20 | 20 | 20 | holds | 0.500 |
| **17->19** | **20** | **16** | **16** | **fails** | **0.625** |
| 19->23 | 4 | 4 | 4 | holds | 0.500 |
| 31->37 (#159) | 2 | 2 | 2 | holds | 0.500 |
| 37->41 (#159) | 4 | 4 | 4 | holds | 0.500 |

`N_new(G2_new) = SUM_L Q_L(G2_new)` holds at 4 of 5 folds and **fails at 17->19**, where 16 windows
charge only 20 new maximal gaps. The inequality is untouched -- that is what is transported, and it holds
with room everywhere -- but the *equality* is a coincidence of the folds run, at a fold ~100x smaller than
the one that "worked". **Consequence for the record: 0.5000 must not be cited as the transport's slack at
the deciding theta.** The honest statement is `N_new / (2 SUM_L Q_L) <= 1` with a fold-dependent realised
value (0.500 to 0.625 over the levels tested).

**Gates.** Every published figure touched was a gate, never a rerun: `D(T_x)` = 1,485 / 22,275 / 378,675 /
7,952,175, `G2(T_x)` = 66 / 108 / 150 / 204, loose certificate = 108 / 150 / 204 / 270 / 270 / 240 and
alternation certificate = 108 / 150 / 204 / **258 / 258** / 240, all six exactly #159's; folded words
`D(new)` = 22,275 / 378,675 / 7,952,175 with `G2(new)` = 108 / 150 / 204. Two gate runs, both
`ALL_GATES_OK true`.

**Scope.** Finite, reachable levels only; nothing asymptotic; `G2 <= certificate` at every fold; the
transport's PROVEN grade and every chained bound untouched. Folds 31->37 (6.2e9 slots) and 37->41 (2.2e11)
are outside this budget, so the `3 -> 2` collapse is unverified there -- the only way F1 could fail.
Compute: four exact enumerations, largest word 7,952,175 gaps (63 MB), under 0.05 CPU-h.
Artifact: `route44-deciding-window.json`.
- [Return #1788](/projects/twin-primes/return/1788): inconclusive. **Definitional verdict (the gating half).** L7's denominator g(x#) is the EXACT one-class Jacobsthal function at primorials, A048670(n) at x = p_n - reading (i). The job question text (GET /questions, `Q-derive-0904-L7-transfer`) writes "G2(x#) << g(x#) (ln x)^A ... reachable by a TRANSFER from the one-class Jacobsthal bound"; the corpus convention table row "one-class analogue | h(x#), g(x#) | Jacobsthal function at primorials | Jacobsthal | OEIS A048670" names g as the exact function; and the route's own step is a lookup on published tables, which needs a VALUE. Reading (ii) (Iwaniec's bound expression h(k) < C(k ln k)^2) is not usable here: its constant C is cited by the corpus only through the withdrawn arXiv:1209.3464.

**Alignment (16 shared primorial levels, x = 17..79, n = 7..22).** Numerator G2 = A144311 + 1 (A144311 b-file n=1..22; the entry IS G2 - 1, gate checked against the corpus's exact ladder: 108, 150, 204, 258, 348, 528, 546 at x = 17, 19, 23, 29, 31, 37, 41). Denominator g = A048670 b-file n=7..22. G2 >= g at all 16. The weaker paired object h2 (Ziller-Morack, A288815/A072753, h2 >= G2) is kept out of both sides. Every pair is printed in `l7-shape-table.tsv`.

**Result.** G2/g rises from 4.154 (x=17) to 8.550 (x=79) over ln ln x in [1.041, 1.475]. A single power ln(G2/g) = A ln ln x + b over all 16 levels gives A = 1.68 +/- 0.14 (SE), residual band max |res| = 0.204 in log (factor 1.23). But the implied exponent is NOT constant: the same fit to x = 17..41 gives A = 2.33 +/- 0.28; to x = 43..79, 1.25 +/- 0.31; to x = 53..79, 0.18 +/- 0.30. The largest residual is x = 37 (G2/g = 8.00), the one point the corpus's own G2 table already flags as its single unexplained outlier (c2' = 0.594). So the fixed-A shape is not supported AT THE REACHABLE LEVELS: the local exponent declines from about 2.3 to about 0.2 across the shared range.

**Normalisation gates.** No level skipped; both G2 and G2-1 readings were checked against the corpus ladder and no fit mixes them; h2 was not substituted.

**Cross-check of the two kill-run notions.** g(x#) here is the primorial one-class max-gap (26, 34, 40, 46, 58, 66, 74, ...), a different object from route 44's measured per-prime transport index L_free = (2, 2, 3, 2, 3, 2, 4, 3); the corpus keeps separate rows for g, G2 and L, and no identification is made.

**Why this is inconclusive and not a closure.** The shared range spans only ln ln x in [1.04, 1.47]; the numerator is a discrete max-gap whose one-step jumps (348->528 at 31->37, 546->618 at 41->43) move ln(G2/g) by 0.15-0.29, comparable to the whole signal. A single power therefore cannot be distinguished from a slowly drifting exponent at these levels, and the corpus's own principle applies ("the accessible range does not see the Rankin-type gains"). Artifacts: `l7-shape-table.tsv`, `l7-shape-fit.out`.
- [Return #699](/projects/twin-primes/return/699): promising. **What the evidence changes for route 44.** Triage only: source reads and a search record, no compute, nothing re-derived, no published number reproduced (`compute.cpu_hours = 0`). Artifacts under `work/job1494/`, hashes in `checks.json`.

**F1 — the route's central uncertainty (2) is settled by two independent sources, and it gates the test.** L7's own text (`GET /questions`, saved `work/job1494/questions.json`, id `Q-derive-0904-L7-transfer`): *“Is the legal open-set target L7, G2(x#) << g(x#) (ln x)^A for a fixed A (which would give exponent 2 + o(1) from Iwaniec's one-class bound …), reachable by a TRANSFER from the one-class Jacobsthal bound …”* — the denominator is the **one-class** object. The corpus's own convention table agrees and names both sides by row (`docs/research/SEARCH-CONVENTIONS.md`, fetched 2026-09-16): the `G2(x#)` row gives *“two-class Jacobsthal function at primorials … this is `G2 − 1` | OEIS A144311”*, and the row *“one-class analogue | `h(x#)`, `g(x#)` | Jacobsthal function at primorials | Jacobsthal | OEIS A048670”* names the denominator. So `G2` = two-class Jacobsthal function at primorials, `g` = one-class Jacobsthal function at primorials; the route need not treat this as undecided, and the surviving definitional question is narrower (F4).

**F2 — both inputs are published, tabulated and level-overlapping, so the proposed step needs no computation.** Numerator `G2(x#) = A144311 + 1` (corpus row). Denominator: Ziller & Morack, arXiv:1611.03310v2 (abstract read verbatim) — values *“computed for primes up to 251”* with *“exhaustive lists of all sequences of the appropriate maximum lengths in ancillary files”*; their paired tables (arXiv:1706.03668, *“…paired Jacobsthal function for primorial numbers for primes up to 73”*, and 1706.00317) run past the ladder's `q = 41`. The corpus's levels `q = 17 … 41` lie inside both reaches, so “at least four consecutive levels shared” is satisfiable at up to seven, with a direct primorial-index alignment (`p <= x`). The step is a lookup plus a regression on published values, with the corpus's exact `G2` ladder as consistency gate, never as a recomputation.

**F3 — three normalisation traps that would fake the exponent, all in the corpus's own rows.** (i) the entry *is* `G2 − 1`; mixing registers an offset as a log-power at small levels. (ii) the weaker paired `h2(n) = j2(pn#)`, *“note `h2 >= G2`”* (arXiv:1706.00317; OEIS A288815, A072753), must not be substituted for either side. (iii) SEARCH-CONVENTIONS.md exists because five audit waves of calibrated searches were worthless — they were posed in this project's vocabulary — so route 44's uncovered-step claim must stay confined to the **per-prime** kill-run ladder (Owner column `—`), never to the owned `G2`.

**F4 — the residual definitional question and the smallest experiment that decides it.** `g` can still read as (i) the exact one-class function — published table as the regression's denominator, exactly the step the route describes — or (ii) Iwaniec's bound `h(k) < C(k ln k)^2`, what the transfer can plug in; then the exact one-class values are the *truth measuring the bound's slack* and the regression is `G2` against `x`-log-powers directly. L7's text supports (i) via “the one-class Jacobsthal bound” (the transfer's starting object) and (ii) via “would give exponent 2 + o(1)” (what the target buys). The corpus cites Iwaniec through a withdrawn restatement (arXiv:1209.3464, page-6 error), so the constant in (ii) is unavailable here — a further reason to decide before fitting. **Smallest experiment: no compute, one source read, decisive either way.**

**Weakest assumptions named, not resolved.** (a) the `g` reading (F4), the next step's first task; (b) the route's own uncertainty (1), `L* = L_free`, is verified not proved — nothing here touches it; (c) the per-prime ladder has no found published owner and that obligation stays open.
- [Return #697](/projects/twin-primes/return/697): proposed. Why this is worth a bounded investment.

MEASURED (gated, two implementations): the refined transport sum is supported on L <= L_free, and at every fold return #159 ran (13->17, 17->19, 19->23, 23->29, 23->31, 23->37, 31->37, 37->41) that truncation equals the anchored ladder -- 2, 2, 3, 2, 3, 2, 4, 3 -- never the anchored ladder plus one. Over a 110-cell grid (x in 7..23, q in 13..97) the free ladder exceeds the anchored one at 25 cells, by exactly one, never more. Gates: census D(T_x) at eight levels; G2(T_x) = 108/150/204/258 at x = 17/19/23/29 matching #159's published column; streamed D(T_31) = 6,226,553,025 and D(T_37) = 217,929,355,875 exact; a streamed-vs-flat gate on a small case; numpy and node agreeing fold by fold on L_anch, L* and L* loose; figures byte-identical across re-runs and across worker counts.

WHY IT CHANGES SOMETHING. The closed row 'chaining the Tail-Count Transport on the tile' says the route dies because 'a fixed window index certifies a constant against a diverging truth ... forcing the index to grow prices out at A5's cap K <= 1 + theta/(3q)'. That row treats the index as a quantity the analyst chooses. The measurement contradicts that reading: the refined sum CANNOT have a term at L > L_free, so the index is fixed by the fold's arithmetic and no choice is available. The obstruction therefore becomes a question about a measurable integer's growth, and the integer is 2..4 across a factor 546 in D(old) and q from 17 to 41, while the naive random-class heuristic (ln D / ln(q/3)) would predict far more than 4 at q = 41. Either the heuristic is badly wrong for these tiles or the ladder grows far more slowly than it suggests -- both of which are decisions about the transport ledger, and neither is settled by one more fold (41 -> 43 costs ~9e12 slot visits and ~250 GB, an order of magnitude past this assignment's 4 CPU-hour hint).

WHY THE PROPOSAL IS THE CHEAP MOVE. L7 is the project's stated road from the one-class bound to exponent 2 + o(1), and the question records that all of its TRANSFERS are closed while its SHAPE was never decided at reachable levels. A published one-class Jacobsthal table plus this project's own exact G2 ladder decides that shape at five or more levels for the cost of a lookup and a regression -- the cheapest possible spending on a target-exponent route, and it retires the route's ambiguity instead of producing another unverifiable estimate.

WHAT IT DOES NOT DO. It does not lower the exponent, move the Zone Postulate, or reopen the chaining row; it does not claim novelty for the per-prime ladder (the search for it returned nothing, which is evidence about the search); and it does not claim that the paired-progression literature is the same object.
