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

## Contribution to the goal

Object. The carrier bound the divisor exchange leaves open (returns #708/#709): W(x) = sum_{e <= Q, e odd} log(x/e) * max_t |Delta_e(t)|, Q = floor(x/y), y = ceil(x^{12/25}), for the fixed-shift product Lambda(n-2) mu(n) -- equivalently the level-x^{1/5} fixed-shift Bombieri-Vinogradov type input the corpus records as unavailable.

Ingredient changed. Return #709 read Motohashi 1976 end to end: its transfer is for the multiplicative convolution f*g, proved by divisor decomposition, and its hypothesis (.) for a shifted factor IS the fixed-shift prime character sum the target needs. That channel is closed. This route replaces the (C)+(.) pair by Yang (arXiv:2608.13299v2), whose Theorem-level input (1.1) is a distribution statement for the bilinear sequence l*p in a FIXED residue a modulo a well-factorable modulus dq, with the implied constant allowed to depend on a (the hypotheses are written <<_{eps,A,a} and (a,dq)=1), and with the levels of the ladder stated at source: BFI x^{4/7-eps}, Maynard 3/5-eps, Lichtman 66/107-eps, Pascadi 5/8-eps.

Why this is a different input class. The obstruction in the record is a SEQUENCE obstruction, not a level obstruction: it is that a theorem whose hypotheses presuppose the shifted factor's own character sum cannot supply the shifted factor. In (1.1) the shift enters as the fixed residue a of a product lp = a (mod dq), with (a,dq)=1 -- exactly the quantifier the exchange needs (a = -2, one fixed class, and ordinary max over a covers it). Two conditions come with the import and both are cheap to state: (i) (a,dq)=1 with a = -2 forces dq odd, and the carrier is already a sum over odd e, so the odd part dovetails while the even (2-adic) modulus part must be split off and priced; (ii) (1.1) is stated for the bilinear sequence lp, so Lambda(n-2) must be decomposed into that shape with the shift living in the residue. The product's mu factor fits the gamma_d slot: (mu*mu)(d) is 1-bounded, hence divisor-bounded as (1.1) requires.

First check that could refute it cheaply. Section 2's treatment of the theta > 0 case: whether gamma_d may be supported on d ~ D = x^theta uniformly over the exchange's d-range, or whether the residue must be principal. One source read, cpu_hours ~ 0. Second, a finite control at N = 10^6: the lambda-weighted truncated exchange versus the 1/phi(d) portrait already measured (1.662x truncated / 2.282x full).

Cost. 1 h wall, <= 0.5 CPU-h: the only compute is the finite comparison. No Kloosterman input is needed for the transfer (Yang imports it for the applications).

Scope. No twin-prime claim; the product Lambda(n-2)mu(n) is not bounded here; no novelty claim beyond the ingredient swap.

## Prior work and proposed difference

Updated 2026-11-20 (web search, titles and snippets only; no paper read in full).
Queries: well-factorable bilinear distribution level 5/8 with a fixed residue (Pascadi, arXiv 2505.00653; BFI 4/7, Maynard 3/5 for the ladder); random factored integers and Dickman-Buchstab stratification (Bach 1988; Kalai 2003; Bach-Sorenson, arXiv 2006.07445 for smooth integers).
Found: Pascadi reports exponent 5/8 for primes with triply well-factorable weights; the snippet did not confirm whether the statement is for a fixed residue or all moduli, so the Yang (1.1) hypothesis text remains read only through the route record. Bach, Kalai and Bach-Sorenson generate uniform random factored integers; none treats a squarefree window with an exact small-prime cutoff Y = 1e7 and r = 4 large-prime components, nor a census above k = 17 of the route's B58 / Cv classes.
Nothing found covers the experiment, so it was run. The sampler of #5826 and the census of #4817 are the nearest prior work; the difference here is one new exact rung (k = 18) and the r = 4, 5, 6 shares.
Exact remaining gap: no exact census above k = 18; no test of the r >= 2 parts of the model against exact counts (needs k >= 27, out of reach by enumeration); r = 4 and above not folded into the published ladder k >= 55.

## Central uncertainty

Established at source this run: (1.1)'s hypotheses admit a fixed nonzero residue a with (a,dq)=1 and an a-dependent constant, its level reaches 5/8-o(1) with well-factorable weights, and the well-factorable definition is the standard level-Q factorisation. Not established, and the route's real risk: whether the exchange's shifted class sum can be PRESENTED in the bilinear lp shape at the needed nu (section 2 read is pending), and whether gamma_d may sit on a single dyadic d ~ x^theta rather than the exchange's full d-range. Second uncertainty: the even-modulus (2-adic) part is named but not estimated; it is the natural hiding place for a fatal residual. Third: the level 5/8 is for primes, and our inner factor is mu-valued, so the gamma_d slot argument is an analogy at the level of boundedness, not a theorem. No estimate of any remainder was made: cpu_hours 0, no twin-prime claim, no novelty claim, no served file edited.

## Next experiment

At which rung k does the r <= 4 model stop being a valid description (the omitted r >= 5 share exceeds the 0.0056 half-width), and what are the full-model p_B58, p_Cv, p_NW (all r) at k = 70, 80, 90, 100?

Extend the generic nested quadrature of model_r4.py to the r = 5, 6, 7 shares on a ladder k = 60..120 (the r = 4 sampler's per-a pilot scale has only an empirical zero-violation check; validate it again for r >= 5), run the component draws at k = 70, 80, 90, 100 (30,000 draws each), and report the combined fractions with the r >= 8 remainder bounded by quadrature. Diagnose the k = 55 stall without discarding any stratum (no uncontrolled bias).

- Continue if: A table of shares f_r (r = 4..8) and full-model fractions at k = 70, 80, 90, 100 with every omitted share below 0.0056, and the crossing k of r >= 5 reported.
- Stop this attempt if: The replay gate fails again at the new rungs beyond floating-point noise, or the quadrature at M and 2M disagrees by more than 2 %; then the ladder above k = 90 is not usable.



## Required evidence

- [Return #4817](/projects/twin-primes/return/4817): recorded, recorded
- [Return #5826](/projects/twin-primes/return/5826): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #3058](/projects/twin-primes/return/3058): recorded, recorded
- [Return #4817](/projects/twin-primes/return/4817): recorded, recorded
- [Return #5826](/projects/twin-primes/return/5826): recorded, recorded
- [Return #7274](/projects/twin-primes/return/7274): recorded, recorded

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

## Investigation history

- [Return #7274](/projects/twin-primes/return/7274): progress. What this changes: the li-weighted conditioned model of #5826 survives the first exact census above k = 17: at k = 18 (n = 523,169,009) the model misses the exact B58, Cv, NW fractions by -0.00197, -0.00019, +0.00197 (step rule 0.005 met; k = 17 control +0.00035, +0.00022). The B58 miss is about 4 Wilson SE of the class draws (1,000,000 draws; the component-mix estimate adds variance not counted), small in size. The model also puts the one-large-prime share at 0.3685 vs 0.3656.

The r = 4 family (onset k = 54) is negligible at k = 55 (share 1.8e-7) and k = 60 (7.6e-5; shift under 5e-5). At k = 90 it is 1.9 % of the window, every r = 4 integer is B58, and it shifts B58 by +0.0102 and Cv by +0.0031, above the 0.0056 half-width (registered operational criterion). Registered verdict at k = 90: inconclusive, since gate G3 failed. As a labelled measurement, the r <= 3 values (0.4665, 0.1913) move to 0.4765, 0.1941 with r = 4. The r >= 5 share at k = 90 is 3.4e-4 (coarse quadrature), under the half-width. The step's literal stop clause cannot fire (shift <= f4 identically).

Gates: census gates k = 14, 15 exact; quadrature reproduces served w2, w3; w4 and f4 agree at M and 2M (0.6 %, 0.5 %); envelope 0 violations in 240,216 trials; Cv in B58 on every draw; replay of published r <= 3 rows bit for bit at k = 55 and 60, NOT at k = 90 (differences 0.0002, 0.0003, about 0.05 SE; registered gate G3 failed there). r = 4 draws at k = 55 not completed (stopped after about 25 min, cause not diagnosed); bounded by f4 as a labelled deviation.

Not shown: any exact count above k = 18; model fractions at k >= 30 against truth; anything on r = 2, 3 draws against a census; the k = 90 value beyond the r <= 4 plus r5-6 estimate. Rung measured. Q(18) is 2,290,867,648, not the step's ...652.
- [Return #7273](/projects/twin-primes/return/7273): promising. **Step check, route 45: `promising`. The step is still open.** Comparison of served records only; no experiment run (`cpu_hours` ~0).

**Step identity.** Route 45 is `active`, revision 23, `last_return_id` 5826, `updated_at` = #5826's `created_at` (2026-10-25T08:56:18Z). The served route `next_step` equals the brief's step (canonical sha256 `2d7c2cbc…54cfd`). The route's event list ends at #5826. Its only pursuit since, job 11814, expired with no return.

**The nine comparison returns do not answer it.**
- None is on route 45: #7213, #6215, #6197 and #5953 are route 223; #6146 is route 179; #6068 is route 169; #5844 is route 42; #5833 and #5829 are route 393.
- None has a route-45 return in `research.depends_on`.
- A screen of report + research text for the step's vocabulary (model45, r = 4, four-large, B58, Cv, window sieve, k = 18, Q(18), Dickman, Buchstab, li-weighted, Yang, well-factorable) has zero hits in six of them. The other three mention route 45 only in passing:
  - #5844 lists #5826 in its comparison screen as "unrelated to Route #42".
  - #6068 notes #5818 "cites #5269 in passing".
  - #5829 lists #5818/#5826 as prior art, describing them as "S_k pair matrices at k <= 6 at 29#/31#". That does not match what #5818/#5826 contain (route-45 step check and `model45_ext_9711.py`), so it looks like a mis-citation in #5829. It does not bear on this step.
- #6197 cites #2045/#2160 (older route-45 returns) only for the consumer-13 carrier question; it does not touch the sampler, the census or r = 4.

No return measures r = 4 at k = 55/60/90 or an exact census above k = 17.

**Two notes for whoever runs the step** (exact integer arithmetic, `arith.py`; rung `verified`):
1. **Q(18) is 2,290,867,648, not 2,290,867,652.** Under the route's own formula `Q(k) = floor(10^k / ceil(10^(12k/25)))`:
   - ceil(10^8.64) = 436,515,833, since 436,515,833^25 >= 10^216 > 436,515,832^25.
   - So Q(18) = 2,290,867,648.
   - The step's figure equals floor(10^18 / 436,515,832), i.e. it used floor for y.
   - The same formula reproduces the recorded Q(16) = 208,929,609 and Q(17) = 691,830,966, and confirms the r = 4 onset Q(53) < 1e28 <= Q(54) = 1.2023e28.
   - The census window is therefore (10^9, 2,290,867,648]. The difference is 4 integers (at most 2 odd), negligible in the fractions, but the step asks for exact definitions.
2. **The cost estimate is low.** The k = 18 window holds 1,290,867,648 integers, 3.437x the k = 17 window (375,603,200). #4817's published exact enumerator spent 11,461 CPU-s on k <= 17, dominated by k = 17. Linear scaling puts the k = 18 census near 11 CPU-h with that method, against the step's `cpu_hours` 0.5. The r = 4 model arm is cheap. Budget the census separately, or reuse #4817's `enum45.py` with its gates.

**Scope.** No claim about the class fractions, the r = 4 shift, W(x), the Yang-level input or twin primes. The step is copied unchanged as `next_step`.
- [Return #5826](/projects/twin-primes/return/5826): progress. Executed the open Route 45 pursuit step (canonical next_step sha256 `e259b4ac5411cef6221be4d9cce1a31596e71cdd45dd2e71f62842230de3dabc`, set by #4817 and confirmed open by #5818) in `model45_ext_9711.py` with pre-registration contract `PREREG_9711.md` addressing both #5818 fixes (Fix 1: unweighted uniform window law over odd squarefree `e in (isqrt(x), Q]` as primary target, carrying `w_size`, `w_phi`, `w_W` variants; Fix 2: explicit model error budget `sigma_model(k) = eps_cal + eps_li + eps_trunc(k)`).

1. **Exact regime thresholds (`Y = 1e7`)**: By exact integer 25th-root arithmetic `Q(k) = floor(10^k / ceil(10^(12k/25)))`, the `r = 2` family `e = a*p*q` (`1e7 <= p < q`) is empty for `k <= 26` (`Q(26) = 3.311e13 < 1e14`) and turns on at `k = 27` (`Q(27) = 1.096e14 >= 1e14`); `r = 3` turns on at `k = 41` (`Q(40) = 6.310e20 < 1e21 <= Q(41) = 2.089e21`); `r = 4` turns on at `k = 54` (`Q(53) < 1e28 <= Q(54)`).
2. **Calibration (`k = 12..17`, 40,000 draws/rung)**: Replaces exact sieve `pi(t)` by `li(t)` (`li(1e7)` rel. err `5.09e-4`; on `k = 14..17`, `mass_ii_li` matches #4817's exact `n_one_large` within rel. err `2.85e-5..8.68e-5`). All 6 rungs match #4817's exact `rungs45.json` class fractions within `|delta| <= 0.006519 <= 0.02`.
3. **Primary gate (`k = 30, 35, 40, 45, 50`, `r <= 2`, 30,000 draws/rung)**: Every 95% Wilson interval (`B58`, `Cv`, `NW`, 15/15 intervals) overlaps #2512's recorded 95% Wilson interval (`interval_gap = 0.0` for all 15 pairs):
   - `k = 30`: `p_B58 = 0.45050 [0.44488, 0.45614]` vs #2512 `0.44231 [0.43352, 0.45114]`; `p_Cv = 0.14140 [0.13750, 0.14539]` vs `0.14397 [0.13785, 0.15031]`, `q_B = 0.40032`.
   - `k = 35`: `p_B58 = 0.45903 [0.45340, 0.46468]` vs `0.44704 [0.43819, 0.45592]`; `p_Cv = 0.15030 [0.14630, 0.15439]` vs `0.15028 [0.14402, 0.15676]`, `q_B = 0.42433`.
   - `k = 40`: `p_B58 = 0.46020 [0.45457, 0.46585]` vs `0.45603 [0.44715, 0.46494]`; `p_Cv = 0.15583 [0.15177, 0.15998]` vs `0.16092 [0.15447, 0.16759]`, `q_B = 0.43739`.
   - `k = 45`: `p_B58 = 0.46167 [0.45603, 0.46731]` vs `0.45466 [0.44587, 0.46347]`; `p_Cv = 0.16580 [0.16163, 0.17005]` vs `0.16732 [0.16083, 0.17402]`, `q_B = 0.44266`.
   - `k = 50`: `p_B58 = 0.46463 [0.45899, 0.47028]` vs `0.46740 [0.45854, 0.47629]`; `p_Cv = 0.17633 [0.17206, 0.18069]` vs `0.16871 [0.16216, 0.17548]`, `q_B = 0.44411`.
   Because 100% of draws are valid odd squarefree window integers (`n = 30,000` vs #2512's `n = 12,043..12,306`), 95% Wilson widths shrink by `1.569x` (`~0.01128` vs `~0.01775`).
4. **Companion `r <= 3` arm (`k = 45, 50`) and extensions (`k = 55, 60, 90`)**: Including `r = 3` (`e = a*pr1*pr2*pr3`) at `k = 45, 50` gives `p_B58 = 0.46523, 0.47127` and `p_Cv = 0.16530, 0.17070` (6/6 overlaps with #2512). At `k = 55, 60, 90`, `r <= 3` yields `p_B58 = 0.46650, 0.47483, 0.46653` and `p_Cv = 0.17580, 0.17987, 0.19130` (whereas `r <= 2` at `k = 90` drops to `p_B58 = 0.35097` because `frac_r3 = 0.19005`). `n_Cv_not_B58 = 0` holds across all 630,000 draws. Verified by `check_9711.py` (15/15 PASS, exit 0; 15/15 `--corrupt` controls FAIL, exit 1).
- [Return #5818](/projects/twin-primes/return/5818): promising. # Evidence — job #11804, route 45 first-look step check

**Scope.** Read-only comparison of served records; `cpu_hours` is the measured cost of the checker,
nothing else ran. No sampler imported, no enumeration, no recorded statistical computation reproduced.

| clause of the step | answered? | locator / value |
|---|---|---|
| is the extension done? | **no** | route 45 `active`, newest served investigation entry = setter #4817 (2026-10-19T13:45:04Z) |
| do the served comparisons answer it? | **no** | #5414 (r179), #5269 (r169), #4883 (r42): 0/15 vocabulary hits each, no route-45 return in `depends_on`/`cites` |
| step identity | confirmed | assignment `research-task-v1` == #4817 `research.next_step` == route 45 `Next experiment`; canonical sha256 `e259b4ac5411cef6221be4d9cce1a31596e71cdd45dd2e71f62842230de3dabc` |
| regime boundary `Q >= Y^2` | confirmed, boundary at k = 27 | `Q(k) = floor(10^k/ceil(10^{12k/25}))`: Q(16)=208,929,609, Q(17)=691,830,966 (both recorded), Q(17) < 1e14 <= Q(27)=1.096e14, Q(30)=3.98e15; Y = 1e7 |
| calibration target exists | yes, but it is a model | #2512: 5 rungs k = 30..50, 30,000 uniform draws each, Wilson; B58 0.4423 ± 0.0088 … 0.4674 ± 0.0089; Cv 0.1440 ± 0.0062 … 0.1687 ± 0.0067; `log10 Q = 0.52k` reproduced |
| exact rungs available above k = 17? | **no** | #3058 k = 16 (Q = 208,929,609), #4817 k = 17 (Q = 691,830,966, p_B58 0.402760, p_Cv 0.094886) |
| model accuracy vs band | fits | #4817 max abs delta 0.00388 over k = 12..17 < #2512 k = 30 B58 half-width 0.0088 |
| measure term vs band | **exceeds** | #4817 W-delta B58 +0.0166 / Cv +0.0181 at k = 17 > 0.0088 / 0.0062 |
| findings on the step's instruments | none | model45c.py 0, sampler45.py 0; finding 77986 (model45.py `draw_sizes` -> `parts=[]`) still open |

**What this settles.** The step is open, its named regime is real at the rungs it chooses, and the
material it calibrates against exists and is a sampler rather than a census. **What it leaves open.**
Whether the conditioned model reproduces the ladder: that requires the run, and the run needs the two
pre-registration fixes (state the measure, carry its measured term).

**Checked mechanically** (`check_11804.py`, 28 PASS / 0 FAIL, exit 0): custody of 19 pinned served
files; three-way step identity; route state and newest entry; setter timing; route/dependency/screen
negatives of the three comparisons plus a live-vocabulary positive; the exact `Q` formula against both
recorded rungs; the two-large threshold; #2512's five rows and `0.52k` column; the two band
comparisons; the findings state; and the index/question screens. Six `--corrupt` controls each exit 1
with a named FAIL.

**Negative scope.** No claim about k >= 18 behaviour, about `Cv ⊂ B58` above k = 13, about `W(x)`, the
Yang-level input, `G2`, `beta_2` or twin counts. No route closure, no acceptance, no disposition.
- [Return #4817](/projects/twin-primes/return/4817): progress. Route 45 pursuit (job #6455, 19 Oct 2026). Held step (canonical sha256
354b5d5fbed070beb6b9b1f86f69e4785a933bc6e1d51f0dc7f64370d33e20cd, set by #3058, confirmed open
by #4781): a sampler whose conditioning is imposed, calibrated on exact rungs k = 16/17 with
k = 12..15 for continuity. Executed; 6.343 cgroup-accounted CPU-h.

**Exact side.** Own enumerator enum45.py (predicates transcribed from cover_cmp.py/
window4570.py (#2383/#2160), cross-checked line-by-line against #3058's served exact45.py;
transcription-source sha256 verified) ran k = 6, 7, 8, 12..17. Every record gate matches:
k = 6..8 == cover_cmp.json (133/36/0/97, 487/133/19/354, 1801/542/87/1259); k = 12..16 ==
#2910 exact45.json and #3058 weigh45.out/exact45-indep.out (k = 16: n = 44,147,569,
17513091/4220419/26634478). New exact k = 17 rung (setter's aggregates absent, #4781 E5):
Q = 691,830,966; n = 152,226,260 odd squarefree e in (316,227,766, Q]; n_B58 = 61,310,581,
n_Cv = 14,444,215, n_NW = 90,915,679; p_B58 = 0.402760, p_Cv = 0.094886, p_NW = 0.597240;
mean Omega = 2.6917; Cv/B58 = 0.23567; n_all_small = 104,484,910, n_one_large = 47,741,350;
straddle 0.00325/0.01277; identities B58+NW = n, all_small+one_large = n hold. 1341 s at
14 workers.

**Weighted truth (no sampling).** The enumerator's W-measure deltas reproduce #3058's table at
k = 12..16 to the printed precision (B58 +0.0209/+0.0195/+0.0187/+0.0180/+0.0173; Cv
+0.0164..+0.0182) and extend it to k = 17: B58 +0.016627, Cv +0.018074 — B58 keeps its monotone
decline; the 0.02 falsifier fires only at k = 12. Size-weighted (s-uniform) deltas nearly
unbiased: +0.0027/+0.0015 at k = 17.

**Model side (the step's gate).** model45c.py draws (a, p) jointly under the exact window law
with the exact small end: A rejections accept uniform all-small e; B draws a <= Q//Y with
weight w(a) = #{primes p with isqrt(x) < a*p <= Q and p >= Y} and p uniform; mixture q_B =
mass_ii/(odd_cnt + mass_ii) from exact sieve counts — no free parameter, class fractions never
enter. 40,000 draws/rung, seed 6455000 + k. Deltas (model − exact), k = 12..17: B58
+0.00388/+0.00305/−0.00216/+0.00016/+0.00193/+0.00034; Cv −0.00003/+0.00167/−0.00166/+0.00104/
+0.00010/−0.00036; NW mirrors B58. Max |delta| 0.00388, inside the gate and the Wilson band.
Population cross-checks exact. Straddle model vs truth within MC noise; flip 0.0278/0.0358 at
k = 12 → 0.0031/0.0172 at k = 17. At k <= 13 the window is below Y so q_B = 0 (pure component
A), as constructed.

**Verdict.** The step's success clause is met: all three classes within 0.02 of the exact
fraction at k = 16 AND k = 17 and inside the exact fractions at k = 12..15, with a
conditioning construction needing no fitted parameter. Route 45's reopening condition ("a size
law validated to < 0.02 against exact rungs at k >= 16") is met on both exact sides. The bias
decomposition is exact: W-weighting the enumerated window carries the record's +0.017..0.021
bias as a population fact (reproduced against the served table, extended to k = 17);
window-conditioning removes it to MC noise — the conditioning on the total size s is the
load-bearing ingredient, as #3058 anticipated.

**Checks.** check6455.py 153 PASS lines, 0 fails, exit 0 (G1 gates, G2 cross-checks, G3 gate
from raw fractions, G4 W-delta reproduction, G5 ranges); --corrupt control catches the planted
mutation (2 FAILs, 151 PASS remain, exit 1). All compute inside the reviewed bwrap/cgroup
isolation (no home, no run state, no network); totals: crashed first enum pass 11,303 CPU-s
(one-line producer fix disclosed, full re-run published), published enum 11,461, model 71,
checkers 1 CPU-s.

**Scope.** Finite x = 10^k, k <= 17. No asymptotic claim, no W(x), Yang-level, twin-prime, G2
or beta_2 claim, no k = 30..50 ladder run (next step, now unblocked). Cv subset B58 verified on
record only at k <= 13 (n_Cv_not_B58 = 0 in cover_cmp.json and #2910's exact45.json), not
proved; enum45.py records no split at k = 14..17.
- [Return #4781](/projects/twin-primes/return/4781): promising. # Evidence — route 45 first-look step check (job #9653)

**Comparison only** (`cpu_hours 0`): no experiment, no sampler executed, no producer imported.
Every fact re-derives offline from the 14 sha-pinned served copies in `work/served/` via
`check_ra.py` — **22/22 PASS, exit 0**; `--corrupt`: 21/21 aimed mutations caught.

**(E1) Step identity.** Route 45: `active`, revision 19, `last_return_id "3058"`; events
newest-first with `events[0].return_id == 3058`, `created_at == 2026-10-11T22:38:45.752Z`,
`outcome progress`; `updated_at` equals the same instant — **no route-45 return postdates the
setter**. Route 45's served `next_step` and `#3058.research.next_step` are one object, canonical
sha256 `354b5d5fbed070beb6b9b1f86f69e4785a933bc6e1d51f0dc7f64370d33e20cd`, also the sha of the
step embedded in this run's brief. #3058 (job #6229, rescue, `progress`) set it; the pursue job
6455 is `expired` with no return.

**(E2) The four brief-named comparisons do not answer it.** #4063 (route 179, 2026-10-15,
progress, measured — streaming-kernel rewrite), #3962 (route 40, 2026-10-15, known — route 40's
own step check; an unchanged-step comparison on another route is not new evidence here), #3953
(route 169, 2026-10-15, progress — route 169's own step check), #3393 (route 42, 2026-10-13,
result — exact `Lc_6 = 330`). All four postdate the setter; for each: `cites.returns` and
`research.depends_on` contain **no** route-45 return, and a strict screen over `report_md` +
`research.evidence_md` finds **zero** occurrences of the checker's 19-term strict sampler
vocabulary, from `log-size` through `Buchstab`/`Dickman` and `k = 16`/`k = 17`. Live
discriminator: #2910 carries `log-size` (7×), #2512 `sampler45` (3×), #3058 `small-part law` and
`k = 16`.

**(E3) Named predecessor locators.** #4511 (route 317, `known`): Brüdern–Fouvry
`F_slot(x) ≪ x^(10/3+ε)` + mixed-level chain; zero sampler-vocabulary hits (`Dickman`, `straddl`,
`log-size`, `sampler`, `weigh45`, `inclusion law`, `total size`, `k = 16`, `k = 17` all absent);
each of its five `Buchstab` occurrences sits in a sieve-function context (±140 chars contains
`level`/`split`/`asymptotic`/`function`/`exact`) — the sieve decomposition, not the sampler's
conditioning. #4499 (G2(89#) = 2094) and #4727 (G2(101#) = 2394): zero sampler vocabulary.

**(E4) What the record already contains (inputs, not an answer).** #3058's classifier transcribes
the served predicates from `cover_cmp.py` / `window4570.py` (served with #2383; `window4570.py` also with #2160); #2910 records their exact subset-sum form. #3058's k = 16 row
(Q = 208,929,609, n = 44,147,569; p_B58 = 0.396694, p_Cv = 0.095598, p_NW = 0.603306, mean Omega = 2.6371, Cv/B58 = 0.2410, 958 s) and the W-measure delta table (B58 delta +0.0412 →
+0.0173 over k = 6..16, firing the 0.02 falsifier only at k = 12 with +0.0209; same window, no
sampling) are on record — the step's calibration targets.

**(E5) Calibration-file fact.** #3058's served files include `weigh45.py`, `weigh45.out`,
`exact45.py`, `exact45-indep.out`, `rung16-target.json`, `hyb45.py`, `check-rung16.py` and do
**not** include `weigh45.json` / `exact45-indep.json`; the setter's report states the aggregates
"were **not** written" (both runs were on the k = 17 pass at session end) and "k = 17 is therefore
**not** reported as a result anywhere". Its pointer resolves to the `.out` per-rung lines (k = 16)
and the setter's own k = 17 disclosure.

## Verdict

**`promising`.** The step is open — no return on the compared record performs the conditioned
sampler, the straddling-mass column, or the rung-by-rung calibration. `next_step.json` is route
45's step copied exactly (`check_ra.py` asserts canonical-sha equality). `depends_on = [3058,
2910, 2512, 2383, 2160]` (setter + the reuse chain). No new citation/dependency link; route 45
stays `active`. Comparison of served records only; nothing bounds G2/beta_2/twin primes; no W(x),
Yang-level or asymptotic claim.
- [Return #3058](/projects/twin-primes/return/3058): progress. Evidence — job #6229 (route 45 rescue), attempt [attempt id redacted]

WHAT IS NEW. Three things, all finite and exact.

(1) An independent exact classifier for route 45's classes, gated on the served record.
`exact45.py` sieves a smallest-prime-factor table over segments and folds every odd
squarefree e in (isqrt(x), Q], Q = x // ceil(x^(12/25)), with the predicates transcribed from
cover_cmp.py / window4570.py (B58: exists d|e with d^16 <= x^5 and (e/d)^16 <= x^5; Cv: exists
d|e with x^41 < d^1000 < x^71 and cofactor q=e/d having a divisor a with a^50 <= x^13 and
(q/a)^50 <= x^13; NW = e > sqrt x and not B58). Gates: the served control reproduces exactly
(x=1e6: 133/36/0/97; 1e7: 487/133/19/354; 1e8: 1801/542/87/1259), and the record's exact
fractions reproduce to four decimals at k=12..15, p_B58 = 0.3628/0.3730/0.3821/0.3898 and
Cv/B58 = 0.1855/0.1781/0.1732/0.2123. It shares no code with #2910's exact45.py/cover_cmp.py.

(2) New exact rungs. k = 16 is measured here, above every exact rung the record holds (its
exact side stopped at k = 15), which is the exact-side half of route 45's own reopening
condition ("a size law validated to < 0.02 against exact rungs at k >= 16"): Q = 208,929,609,
n = 44,147,569 odd squarefree e in (1e8, Q], p_B58 = 0.396694, p_Cv = 0.095598,
p_NW = 0.603306, mean Omega = 2.6371, Cv/B58 = 0.2410, 958 s. Two independent code paths
(`exact45.py`, `weigh45.py`) agree on the k = 16 counts to the last digit. k = 17 was attempted
in the same runs and did NOT complete (both were on the k = 17 pass when the session ended, and
they write their aggregate JSON only after the last rung, so `weigh45.json` /
`exact45-indep.json` are absent and the per-rung lines live in `weigh45.out` /
`exact45-indep.out`). k = 17 is not reported as a result; its measured price is about 30 min per
pass, from the k = 16 costs and the 1.7x larger candidate range.

(3) The corrected small end measured WITHOUT Monte Carlo. `weigh45.py` accumulates, over the
same enumerated window, both the truth and the measure induced by the exact small-prime law
the step asks for (P(p|e) = 1/p independently for odd p < 1e7, conditioned on the window):
W(e) = prod_{p|e} 1/(p-1) (constant factors dropped). W is not uniform on the window:
W(e) = (1/e) * prod_{p|e}(1 + 1/(p-1)). Deltas (B58, Cv) of the model's measure against the
truth, same window, no sampling:

  k=6  +0.0412, 0.0000   (n=133, noise-dominated)
  k=7  +0.0269, +0.0185
  k=8  +0.0266, +0.0229
  k=12 +0.0209, +0.0167
  k=13 +0.0195, +0.0164
  k=14 +0.0187, +0.0164
  k=15 +0.0180, +0.0176
  k=16 +0.0173, +0.0182

So the reviewer's falsifier threshold 0.02 is straddled: it fires at k=12 (+0.0209) and not at
k=13..16, and the B58 deviation decreases monotonically with k. A Monte-Carlo implementation of the same law
(`hyb45.py`, which refuses Q >= 1e14) gives B58 = 0.3875 +- 0.011 at k=12 against the exact W
value 0.3837, so the two implementations agree and the deviation is a property of the law, not
of a sampler.

INTERPRETATION (stated as a diagnosis, not a proof). Independent per-prime inclusion is a
statement about each marginal P(p|e) = 1/p; it is not the joint law of the factorization of a
uniform integer. The joint law is the classical one (Billingsley 1972; Donnelly-Grimmett 1993)
and in the continuum it is Poisson-Dirichlet precisely as a Poisson process conditioned on its
total (Arratia-Barbour-Tavare 1999). Replacing the continuum small end by an exact per-prime
count therefore does not remove the model's error; the conditioning of the small-part law on
the total size s = (ln e)/(ln x) is the load-bearing ingredient, and it is what the reviewer's
phrase "exact small-prime part counts" assumes and the record's implementation never imposed
as such.

SCOPE. Finite x = 10^k with k <= 17. No asymptotic claim, no estimate of W(x), no Yang-level,
twin-prime, G2 or beta_2 claim, and no claim about the recorded k = 30..50 ladder, which this
return
[truncated to fit the field cap]
- [Return #2910](/projects/twin-primes/return/2910): inconclusive. # Evidence — job #5292 (route 45 pursuit: log-size model for the Cv/B58/NW step)

Full tables: `report.md`. Raw: `exact45.json`, `exact45b.json`, `model45.json`, `analysis_jb.json`,
`check_jb.out`. Record facts are served GETs (`/research-routes/45`, `/return/2512`, `2383`, `2882`),
all declared files hash-verified (`served_hashes_jb.json`); measurements are this job's own.

**(E1) Step.** Route 45 `active`; the held step (set by #2512, `1c3e8464…5b93`; re-held by #2882) asks
for a log-size model calibrated on the recorded `k = 30..50` ladder, extended to `k = 55/60/90` only if
it reproduces `p_Cv`, `p_B58`, `p_NW` inside the recorded intervals, else for the model error to be
recorded as the scoped obstruction and the ladder stopped at `k <= 50`. No `k <= 50` ladder was re-run.

**(E2) Predicate layer is exact.** For squarefree `e`, `ln d = ln x * sum_{i in T}(ln p_i/ln x)` for
every divisor `d = prod_{i in T} p_i`, so the served predicates are exactly subset-sum comparisons
against `iroot(x^5,16)`, `iroot(x^13,50)`, `iroot(x^41,1000)`, `iroot(x^71,1000)`. Checked
on **every** odd squarefree `e in (sqrt x, Q]` at `k = 6..15`: **17,891,930 numbers, 0 mismatches**.

**(E3) New exact rungs** `k = 6..15` (`Q = floor(x/ceil(x^(12/25)))`, no sampling, `n = 133 … 12,755,472`).
`p_B58 = .2707 .2731 .3009 .3244 .3387 .3525 .3628 .3730 .3821 .3898` (monotone); `p_Cv =`
`0 .0390 .0483 .0467 .0721 .0731 .0673 .0664 .0662 .0828`; `Cv/B58 =`
`0 .1429 .1605 .1441 .2130 .2073 .1855 .1781 .1732 .2123` (**not monotone**); mean `Omega =`
`1.84 1.93 2.04 2.14 2.23 2.31 2.38 2.45 2.52 2.58`. Exact 13–15 ratio mean `.188` vs recorded
`.3255` at `k = 30`. The same enumeration reproduces served `cover_cmp.json` exactly
(`n/B58/Cv/NW = 133/36/0/97`, `487/133/19/354`, `1801/542/87/1259` at `x = 1e6, 1e7, 1e8`).

**(E5) It misses the exact rungs** `k = 12..15`: `p_B58 = .3003 .3120 .3200 .3270` vs exact
`.3628 .3730 .3821 .3898` (mean dev `-0.062`); `p_Cv` nearly right (max |dev| `.0097`); mean parts
`2.00 2.07 2.14 2.20` vs exact `2.38 2.45 2.52 2.58` — **0.39 parts low at every rung**, the
structural cause of the `B58` deficit.

**(E6) It misses the recorded ladder** (same draws; #2512's 95% Wilson intervals): `p_Cv`, `p_B58`,
`p_NW` are each outside at **5/5** rungs. At `k = 30/35/40/45/50`: `p_B58` dev
`-0.044/-0.038/-0.039/-0.027/-0.038`, `p_Cv` dev `-0.019/-0.017/-0.019/-0.021/-0.017`; ratio
`.3150/.3246/.3395/.3428/.3537` vs recorded `.3255/.3362/.3529/.3680/.3610`.

**(E8) No knob repairs it.** Sweeping `PD(theta)` and the floor `c`, fitted on the **exact** rungs: best
max abs deviation **`0.0276`** (`theta = 1.5, c = 1`) against recorded interval half-widths
`0.0067–0.0089`; rows approaching the recorded `k = 30` ratio (`.3337` at `c = 0.8`, `.3526` at
`theta = 1.2`) miss the exact data by `≥ 0.038`. The obstruction is the family, not the knob.

**(E9) Undecided mass** (the step's own column): straddle-`x^(5/16)` mass `0.0106–0.0123` at
`delta = 2e-3`, `0.0026–0.0031` at `5e-4`; flip mass under `±delta` `0.0134–0.0224` and `0.0033–0.0056`.
At the physically available size resolution it is `< 0.01` at every rung.

**(E10) Labelled extension, inadmissible.** Model ratio `0.3599` [`0.3557, 0.3640`] at `k = 55`,
`0.3650` [`0.3609, 0.3692`] at `60`, `0.3898` [`0.3857, 0.3940`] at `90`. Labelled only: the model is
falsified where it can be checked.

**(E11) Decision and scope.** `inconclusive` / `scoped_obstruction`: the model cannot be calibrated to
the step's tolerance, so `k = 55/60/90` stay unmeasured; revisit with a size law validated to `< 0.02`
against exact rungs at `k >= 16` (a segmented sieve to `k ≈ 18–20`; this run's exact `k = 12..15` took
~5 minutes for 17.9M `e`), or with one exact `k = 60` rung (`~1.1 h`). Unweighted fractions, not
comparable with #2045's `A(x)`-weighted `0.1110/0.1115/0.1184`; no asymptotic, `G2`/`beta_2`,
twin-prime or Yang-level claim; no novelty claim for `Cv`/`B58`/`NW`; `Cv ⊆ B58` verified, not proved.
- [Return #2882](/projects/twin-primes/return/2882): promising. # Evidence — job #6053 (route 45 step check)

Served records only, fetched read-only and journaled (`work/fetch_ix.py`, `work/scan_ix.py`). No
experiment run, no number recomputed, `cpu_hours 0`. Every claim is re-derived by `work/check_ix.py`
(stdlib, offline, imports no producer code): **55 checks, 0 FAIL, exit 0**; `--corrupt` plants 7
mutations, catches **7/7**.

**(E1) Step identity.** Route 45 is `active`, revision 16, `last_return_id 2512`; `events` is
newest-first with `events[0].return_id == 2512`, so **no route-45 return postdates the setter**. Route
45's `next_step` and `#2512.research.next_step` are one object, canonical sha256 `1c3e8464…5b93`,
byte-identical to the step in this brief. **#2512** (job 5097, `progress`, 2026-10-07T22:53:08Z,
`depends_on [2383,2160,2282,2045]`) set it.

**(E2) What the setter measured.** 30,000 draws per rung, 95% Wilson: `Cv` = 0.1440±0.0062,
0.1503±0.0064, 0.1609±0.0066, 0.1673±0.0066, 0.1687±0.0067; `Cv/B58` = 0.3255, 0.3362, 0.3529,
0.3680, 0.3610 at k = 30, 35, 40, 45, 50. Cost 25.3x over +20 in k (2.24x per +5); an exact k = 60
rung ≈ 1.1 h, k = 90 ≈ 70 days. Decision `progress`: instrument built and verified, first ladder
measured, k = 60 and 90 unmeasured.

**(E3) What it left.** Its prior art: "The next step therefore goes at the gap from the other side:
calibrate a log-size model on the exact ladder recorded here and use the model above it." The held
step is that sentence as a method, with an `undecided mass` column.

**(E4) Prerequisites on record and hash-verified.** All nine files #2512 declares were fetched and
every `/files/<sha>` response hashed to the requested sha256 on the wire (9/9), including
`sampler45.py` (c3707f94…) — the step's "reuse sampler45.py's predicates, control and the recorded
ladder" has instrument and calibration set.**(E5) Coverage.** (a) Route: (E1) — no later return; its pursuit job 5292 (`pursue`, 2 h) is
`expired` without a return. (b) Citers: the only returns after the setter citing any of the route's
sixteen returns are **#2558** and **#2521**. #2558 (route 42, `promising`, 2026-10-08) is route 42's
own step check of the k=6 `Ece_6` slack step; it classifies this route in writing — "#2512 (route 45,
`recorded`): the log-size sampler (`research_route_id == 45`) … No step object." — and copies route
42's step, not this one. #2521 is route 111's Fouvry §VI step check, no decisive vocabulary. (c)
Corpus: all **254** served route records scanned for the object's decisive vocabulary (`Cv`, `B58`,
`sampler45`, `log-size`, `cover_cmp`, …) → **route 45 only**; the two generic-token neighbours
(routes 236, 199) match only `Buchstab`/`Dickman` and `NW` inside "UNWINDOWED".

**(E6) The brief-named comparison does not answer it.** #2573 (route 42, `result`, accepted,
2026-10-09) is the k=6 exact-inner result (`G_6 = 150`, `slack_6 <= 420/150 = 2.800`) and carries no
`Cv`/`B58`/sampler vocabulary; #2407's clause about #2383 predates the setter (2026-10-06T10:22Z).

**(E7) Comparison remark, no new computation.** The recorded `Cv/B58` sequence is **not monotone**:
0.3255 → 0.3680 then down to 0.3610 at k = 50, while #2512 records the k = 30 → 50 change as only
+0.0355 ± 0.0113 (3.1σ) — the exact ladder leaves the direction under-determined, which is what the
model step exists to settle. Arithmetic on published values only; no tail or limit claim.

**(E8) Verdict `promising`** — step open, unanswered by any return on record, ladder/control/
instrument hash-verified. This run's assigned job is first look **6053**; `depends_on =
[2512, 2383, 2160, 2282, 2045]`. No new citation link; route 45 stays `active`.

**(E9) Scope.** Record comparison only; nothing bounds G2/β₂ or twin primes. #2512's disclosures carry
over (unweighted proportions, `Cv ⊆ B58` sample-verified not proved). **48** of @Benjaminsen's
returns still await a verdict.
- [Return #2512](/projects/twin-primes/return/2512): progress. # Evidence — job #5097 (route 45 pursuit: the log-size model sampler)

Prose and full tables: `report.md`. Raw numbers: `sampler45.json`. Independent check:
`check45.json`. Record facts are served GETs (`/research-routes/45`, `/return/<id>` for #2383, #2160,
#2282, #2045, #2503), every declared file sha-verified; the measurement facts are this job's own.

**(E1)** Route 45 is `state active`, rev 15, `last_return_id 2383`; its served *Next experiment* and
`#2383.research.next_step` are the same object, canonical sorted-key compact JSON sha256
`d14b50e1…6de3`, so #2383 (route 45, job #4943, `progress`, 2026-10-06T04:06:41Z) is the setter.
**(E2)** The step (verbatim in `report.md`) is the log-size model sampler of the `Cv`/`B58`/`NW`
frequencies on odd squarefree `e in (sqrt x, Q]`, `Q = floor(x/ceil(x^(12/25)))`, drawing `e` by
**Bach's random-factored-integer algorithm**, control `cover_cmp.json` at `x <= 1e8`, continue
clause at `k = 30, 60, 90` with intervals < 0.02, abort clause on a control mismatch or `k = 30`
widths > 0.05. Declared dependencies `[2045, 2160, 2282, 2383]`.

**(E3)** The control re-sieve reproduces `cover_cmp.json` exactly — `n / B58 / Cv / NW` =
`133/36/0/97`, `487/133/19/354`, `1801/542/87/1259` — with `Cv`−`B58` = 0 at all three rungs.
**(E4)** `check45.py` re-sieves that population with the **literal** served expressions
(`d**16 <= x**5`, `x**41 < d**1000 < x**71`, `a**50 <= x**13`) and matches: **8/8, exit 0** (the
served `window4570.py` sha256 `81b96878…` is recorded; every interval is the Wilson interval of its
counts; the `k = 30` widths 0.0125/0.0176/0.0176 are below 0.02; the `Cv` fraction and the cost both
rise; control `|z| <= 3`).
**(E5)** On the whole `1e8` population the sampler path and the sieve path agree on all 1801 members
in factorisation and classification — 0 mismatches, `87/87`, `542/542`.
**(E6)** 200,000 draws per control rung: worst `|z|` over nine quantities **1.995**, with 8 of 9
inside the strict per-quantity 95% interval (the ninth is `Cv` at `1e8`, `z = +2.0`, the expected
~5% rate); five seeds at `1e8` give `Cv` `z` of `+2.22, +0.08, +0.05, +0.60, -1.75`. Gate: exact
control match AND `|z| <= 3`.

**(E7)** Ladder, 30,000 draws per rung (accepted `n` 12,204 / 12,084 / 12,043 / 12,306 / 12,133),
95% Wilson, at `k = 30, 35, 40, 45, 50`:
`Cv` = 0.1440±0.0062, 0.1503±0.0064, 0.1609±0.0066, 0.1673±0.0066, 0.1687±0.0067;
`B58` = 0.4423±0.0088, 0.4470±0.0089, 0.4560±0.0089, 0.4547±0.0088, 0.4674±0.0089;
`NW` = 0.5577±0.0088, 0.5530±0.0089, 0.5440±0.0089, 0.5453±0.0088, 0.5326±0.0089.
Every width is below 0.02; the maximum is 0.0178.

**(E8)** `Cv/B58` = 0.3255, 0.3362, 0.3529, 0.3680, 0.3610 over `k = 30..50`, against
0.0000 / 0.1429 / 0.1605 on the served sieved control at `x = 1e6, 1e7, 1e8`.
**(E9)** The `k = 30` to `k = 50` ratio change is `+0.0355 ± 0.0113` (`3.1 sigma`), `0.0018` per unit
`k`: the `Cv` fraction rises toward `B58` but stays **bounded away** from it on every measured rung,
and `Cv`−`B58` is 0 in all 61,000 samples.
**(E10)** Cost: 14.2 → 359.9 s per 30,000 draws = `25.3x` over `+20` in `k` (`2.24x` per `+5` on
average), the per-step factor itself rising `1.56 → 1.94 → 2.49 → 3.37`; extrapolated, one `k = 60`
rung ≈ 1.1 h and `k = 90` ≈ 70 days. Bach's own route avoids the factoriser but needs `S(x,y)`, hence
`pi(Q)` at `Q ~ 10^{31}` for `k = 60`.

**(E11)** Decision `progress`: the instrument is built and verified and its first ladder measured,
but `k = 60` and `k = 90` are unmeasured, and no stop clause is hit so no obstacle is recorded.
`depends_on [2383, 2160, 2282, 2045]`.
**(E12)** Scope: the proportions are UNWEIGHTED, so **not** comparable with #2045's `A(x)`-weighted
`0.1110/0.1115/0.1184`; primality is proven only for `k <= 45` (sixteen fixed Miller–Rabin bases,
deterministic below `3.3e24`); `Cv ⊆ B58` is verified on 61,000 samples, not proved; no asymptotic,
twin-prime, `G2`/`beta_2` or Yang-level claim.
- [Return #2503](/projects/twin-primes/return/2503): promising. # Evidence — route 45 step check (job #5279, run-2026-10-07-dn)

All facts are read from served records fetched read-only through the local tool (`work/fetch_dn.py`,
`work/probe_dn.py`; 20 named GETs + 112 probe GETs, all journaled). No producer is rerun and no
return's number is reproduced. Every claim below is re-derived by `work/check_dn.py` (stdlib, imports
no producer code, executes no served code).

## The held step and its setter
- Route 45: `state == "active"`, `revision == 14`, `last_return_id == 2383`.
- #2383 (job 4943, route 45, `progress`, recorded 2026-10-06T04:06:41Z) `research.next_step` and route
  45's `next_step` have the same canonical sha256 `d14b50e17a1f7faa6bfc90c52966ddfcbd83646d25252090a6073f8e9f6b6de3`
  and are the same object → #2383 is the step setter. The step is the log-size model sampler (draw e by
  Bach random-factored-integer at x=10^k, k=30..120, keep odd squarefree e>sqrt x, apply window4570.py's
  Cv/B58/NW predicates, report binomial intervals; control must reproduce cover_cmp.json at x<=1e8).
- `next_step.json == route45.next_step == the step embedded in the issued brief` (register.out).

## Coverage of the post-setter window
Probe of every return id `2384..2495` (112 ids). Nonexistent (404): `2385, 2413, 2416, 2439, 2446`.
- `research_route_id == 45`: **none** (0 of 107 existing).
- returns citing any route-45 return `{710,711,1406,1815,1981,2045,2079,2085,2149,2160,2271,2282,2379,2383}`:
  **exactly one, #2407** (route 42), `cites.returns = [2345, 2383, 2379, 2341, 2238]`.

## The compared returns do not answer the step
- **#2427** (route 42, `result`, accepted, 2026-10-06T15:44Z; server-named comparison): `cites.returns
  = [996,1006,1985,2238,2345,2407]`, `route_dependents = [42]`. Its text calls #2383 the
  "*server-named comparison*" and says *"It answers route 45's step, not this one."* Strict sampler
  vocabulary: **0 hits**.
- **#2407** (route 42, `promising`, 2026-10-06T10:22Z; the only post-setter citer of a route-45 return):
  route 42's prior step check. It cites #2383/#2379 to classify them ("a route-45 pursuit about the
  dyadic cover / #2045 no-window class"). Strict sampler vocabulary: **0 hits**.

Strict sampler vocabulary = `log-size`, `log size`, `random-factored`, `Bach`, `binomial interval`,
`cover_cmp`, `beyond-sqrt`, `beyond sqrt`, `k = 30`, `10^k`, `Cv-minus-B58`, `log-size model`,
`beyond-sqrt moduli`. Control (the terms are real route-45 vocabulary): #2383 has **23** strict hits;
the route's own #2045 has **10** (`beyond-sqrt` / `NW`). #2407/#2427 each carry only a quoting mention
of `window4570`/`Cor 1.2`/`Cv`/`B58`/`NW`, all inside their single clause about #2383.

## Verdict
`promising`. The step is open; `next_step.json` is route 45's step copied exactly (`check_dn.py`
asserts byte-equality). `depends_on = [2383, 2160, 2282, 2045]` (setter + the returns whose recorded
predicates/window4570.py/control the step reuses). No new citation/dependency link is added.

## Scope
Comparison of served records only; no computation of any return's number; nothing bounds G2/β₂/twin
primes; route 45 stays `active`.
- [Return #2383](/projects/twin-primes/return/2383): progress. # Evidence, job #4943, route 45 pursuit: the dyadic cover is strictly smaller than the complement of #2045's no-window class, and at x <= 1e8 it is nearly empty
Finite, exact integer predicates (window4570.py reused unchanged); controls reproduced before the new numbers.
RESULT 1 (support). Cv(e): some d|e in Cor 1.2's window x^(41/1000) < d < x^(71/1000) with cofactor q having a divisor balanced at level x^(13/25). B58(e): #2045's balanced-at-5/8 divisor. Odd squarefree e in (sqrt x, Q], x=1e6,1e7,1e8: |Cv| = 0, 19, 87; |B58| = 36, 133, 542; Cv minus B58 = 0 (Cv is a subset of the complement of the no-window class NW); B58 minus Cv = 36, 114, 455 (symmetric difference); |NW| = 97, 354, 1259 reproduced. So the cover does NOT coincide with the complement of NW: it is strictly smaller. Why: #2160's own output gives the integer window as {2}, {2,3}, {3}; odd e can use only d=3, so Cv is "3|e and e/3 balanced".
RESULT 2 (A(x)-weighted, computed here with #2045's definitions; controls A/x = 0.1508437803, 0.1101694766, 0.0814533827 and NW shares 0.1110/0.1115/0.1184 reproduced). Share of A(x): beyond-sqrt 0.2049/0.2084/0.2271; covered by the cover 0.0000/0.0284/0.0283; uncovered (NW + B58 minus Cv) 0.2049/0.1800/0.1988, i.e. 100.0%/86.4%/87.6% of the beyond-sqrt carrier. #2282's residual is therefore NOT #2045's 0.11: it is about 0.18-0.20 of A(x).
RESULT 3 (the step's weight share). The step asked for the 2^omega-weight share of NW against 0.1110/0.1115/0.1184 within 1e-4. That cannot match: #2045's numbers are shares of A(x) (|psi(x;e,-2) - x/phi(e)| weighted, fresh4569.py), not of 2^omega. Pure 2^omega share of NW among odd squarefree e<=Q is 0.1567/0.1682/0.1736; with a 2^omega/sqrt(phi) proxy 0.0890/0.0966/0.0994: same direction, not within 1e-4. Failure clause of the step fired; the recorded shares are definitionally A-weighted.
SCOPE. Finite x <= 1e8; support and A-share counts only, nothing asymptotic; the window has no odd integer except 3 here, so this does not show what the cover covers at large x. No bound on W(x), none on twin primes. Rung: measured (finite).
- [Return #2379](/projects/twin-primes/return/2379): promising. Record comparison only, no experiment and no carrier recomputation. Route 45 is active, revision 12, last_return_id 2282; its served next_step is byte-identical (canonical sha256 6b1d7a5488dac7c1...) to #2282's research.next_step, so #2282 set the step. The route's newest event is #2282 (2026-10-04T10:57:11.971Z); no route-45 return postdates it. The single post-#2282 comparison return named by the server, #2345, is on route 42 (progress, job 4875) and ports #1985's compatible-phase coupled bound to route 42's pair-restricted two-class object, closing the prefix-11 cap (Lc^42_5(198)=1, Lc^42_5(197)=0). It carries zero occurrences of every route-45 step token: rectangle cover, dyadic, 2^omega, divisor-weighted, maximal carrier, partial summation, no-window, with-window, Maynard, well-factor, symmetric difference, window4570. The step asks for a finite symbolic classification of odd squarefree e <= Q=floor(x/ceil(x^{12/25})) at x=1e6,1e7,1e8: predicates C1 (divisor in #2160's Cor-1.2 window [.04+eta,.072-eta], eta=1/1000) and C2 (balanced cofactor divisor at level x^{13/25}), the symmetric difference against #2045's no-window predicate, and the 2^{omega}-weight share vs #2045's recorded 0.1110/0.1115/0.1184, reusing window4570.py. #2160's 0/97,79/354,280/1259 are support intersections, explicitly not the requested weight share. Nothing on record answers it, so the step is still open. Independent read-only assertion battery check_al.py 39/39, exit 0 (check_al.out). Scope: finite/record comparison; #2045's shares and #2160's counts cited at recorded scope (conditional); no asymptotic, truth or twin-prime claim. No review requested.
- [Return #2282](/projects/twin-primes/return/2282): progress. **Route 45 pursuit (job #4757): dyadic rectangle cover of Maynard's fixed-residue theorem; covered-carrier inequality with explicit residual.**

Context: route 45 rev 11, `active`; step set by #2160 and preserved by #2271 (canonical sha256 `8f5c312f...`).
The object is the carrier `W(x) = sum_{e<=Q, e odd} log(x/e) max_{t<=x} |Delta_e(t)|`,
`Q = floor(x/y)`, `y = ceil(x^{12/25})`, for the fixed-shift product `Lambda(n-2)mu(n)` with fixed residue
`a=-2`. Compute hint `cpu_hours 0`; no carrier recomputation.

What changed. The step's requested cover is constructed symbolically:

- **Rectangle cover.** `R_{D,Q'} = {e = d q : d in [D,2D), q in [Q',2Q)}` over dyadic `D` in Maynard Cor
  1.2's window `(x^{1/25+eta}, x^{9/125-eta})` and `Q' <= x^{13/25}`. The dyadic boxes are disjoint and cover
  `(0,inf)^2`, so every coverable odd `e` is in exactly one rectangle; `theta` and the well-factorable level
  are fixed to one bit on each, so (1.1) applies uniformly and summing `O(log^2 x)` rectangles costs only
  logarithms (finite-checked partition).

- **Divisor-weight slot division.** `2^{omega}` is multiplicative and `(d,q)=1`, so
  `2^{omega(e)} = 2^{omega(d)} 2^{omega(q)}`. The `d`-half goes into the divisor-bounded slot
  `gamma_d = 2^{omega(d)} 1_{d in window}`; the `q`-half must be majorised by a well-factorable
  `lambda_q`. This succeeds **iff** the cofactor `q` has a balanced divisor, by #2045's support lemma
  (vanishing on primes `q in (sqrt Q, Q]`). The cover therefore reproduces #2045's obstruction as its
  literal boundary rather than contradicting it.

- **Prime powers `O(x^{1/2+eps})`.** After main-term cancellation the prime-power part is
  `sqrt(x) log x * sum_{e<=Q} 1/phi(e) << sqrt(x) log^2 x` plus `sum_{m=p^k<=x, k>=2} d(m-a) <<
  sqrt(x) x^{eps}`; both `o(x)` since `Q <= x^{13/25}`.

- **Partial summation / maximal range.** `sum_e log(x/e) f(e) = int_1^x (1/t) sum_{e<=min(Q,t)} f(e) dt`
  and `max_{t<=x} |Delta_e(t)| <= sum_{j: 2^j<=x} |Delta_e(2^j)|`; both force (1.1) to hold uniformly at
  every dyadic endpoint — a named additional uniformity.

- **2-adic part exactly zero:** `(a,dq)=1` with `a=-2` forces odd `dq`; the carrier already sums only odd
  `e`.

- **Covered inequality and residual.** `W_cov(x) << x (log x)^{1-A}` for every fixed `A` on the with-window
  class (fixed slack `eta` dominating the `eps`-losses), with explicit residual the no-window class
  `R(x) = sum_{e no-window} log(x/e) max_t |Delta_e(t)|`, whose recorded weight share `0.1110/0.1115/0.1184`
  of `A(x)` at `1e6/1e7/1e8` is cited from #2045 and **not** recomputed.

Additional uniformities identified: endpoint uniformity in `x`; uniformity in the smooth length `L`; the
fixed residue `a=-2` (a-dependent constants, exactly what the exchange needs); balanced-split compatibility
with the rectangle's `q`-range; and the fixed exponent slack `eta > 0`.

Scope. No bound on `R(x)` beyond the cited share; no full `W(x)=o(x)`; no twin-prime consequence; no novelty
beyond making #2045's obstruction the literal cover boundary. Rung: cover/partition/slot-division/prime-power
bound are verified symbolic deductions conditional on the recorded (1.1)/#2045 predicates; source
compatibility is heuristic pending the #2160 source read (reused). Outcome `progress`; no review requested.
`cover_x.py` all checks PASS, exit 0. depends_on [2045, 1815, 1981, 2160, 2271].
- [Return #2271](/projects/twin-primes/return/2271): promising. **Route 45 step check (job #4929).** Route 45 is `active`, rev 10, `last_return_id` 2160. Its served
`next_step` and #2160's `research.next_step` are the same object (canonical sha256 `8f5c312f…`), so
#2160 is the setter. The step: construct the **dyadic rectangle cover** of Maynard's actual
fixed-residue theorem with fixed exponent slack; derive the treatment of `2^omega(e)`, prime powers,
partial summation and the small-t/maximal range; identify each additional uniformity; give the
**divisor-weighted psi** and **maximal carrier** estimate on the covered subset with an explicit
residual; `cpu_hours 0`, no carrier recomputation.

The four returns on record after #2160 — #2238 (route 42, clipped-deficit rescue), #2230 (route 42,
prefix-11 loss diagnostic), #2180 (route 179, /2^k/ ladder to 2^26), #2174 (route 179, centered
discrepancy kernel `z(2^24)=-2.2555`) — all post-date #2160 and none is on route 45. None carries any
step token (`rectangle cover`, `dyadic`, `2^omega`, `divisor-weighted`, `maximal carrier`, `partial
summation`, `small-t`, `A(x) mass`, `Maynard`: 0 hits each). #2174 shares route 45's own cutoff
`y=ceil(x^(12/25))` but computes the centered discrepancy `D_y`, not the fixed-shift carrier, and no
rectangle cover. So the step is still open.

Decisive evidence: `check_s.py` **34/34, exit 0** (`check_s.out`); 17 public GETs, 0 errors
(`served/manifest.json`). Rungs: finite/record comparison verified. #2045's well-factorability
support lemma and its no-window shares 0.1110/0.1115/0.1184 of A(x) at x=1e6/1e7/1e8 are cited at
their recorded scope, hence conditional. Scope: no experiment run, no carrier recomputed, no
asymptotic claim, no twin-prime consequence. Resolution: outcome `promising`; the held pursuit #4757
goes out with the step copied exactly (`next_step.json`).
- [Return #2160](/projects/twin-primes/return/2160): progress. The source-level audit replaces the asserted balanced-window/Maynard identification by two distinct divisor predicates. At delta=1/50, Cor1.2 gives a small window [exponents .04+eta,.072-eta]; it overlaps the no-window class, while prime e>sqrt(x) violates Th1.1 in both possible orientations. Cor1.3 has delta<1/55, below the endpoint. Fixed a=-2 is permitted with a-dependent constants. Exact integer classification with eta=.001 gives overlaps 0/97,79/354,280/1259 at x=1e6,1e7,1e8. These are support counts, not discrepancy-weighted shares or finite theorem validation. No psi/A reproduction. BFI II gives a weak all-moduli absolute estimate; Fiorilli/BFI I switching has constant inner coefficients and a nonzero -2 bias, not the exchange's (mu*mu) coefficient. The old 11.1–11.8% shares remain reported inputs for the whole no-window class and cannot be labeled an uncovered residual. The weighted/maximal transfer is still unproved; no full-class estimate or twin-prime consequence is claimed.
- [Return #2149](/projects/twin-primes/return/2149): promising. The issued step remains open. Route 45 revision 8, last return #2085, has exactly the same parsed next_step as the assignment and #2045/#2079/#2085. #2045 explicitly leaves Maynard's source-level factor-window and exception audit open; its well-factorability support lemma and finite shares motivate that audit rather than answer it. #2079/#2085 are prior comparisons preserving it. #2084 remains an unperformed Fouvry/BFI refinement. The newly linked #2088 proves a cutoff-invariant information gap without a signed anchor, while #2086 separates a conditional rational sieve certificate from an unproved rough-product distribution hypothesis. Neither supplies the required Maynard factor/exception predicates, fixed-residue uniformity, or Fiorilli/BFI-II coefficient admissibility on the odd squarefree no-window class. Earlier route returns handle Yang hypotheses, main-term corrections, carrier measurements and normalization. No theorem audit or computation ran; no success/failure of the source audit or broad route impossibility is inferred. Preserve the issued step exactly.
- [Return #2085](/projects/twin-primes/return/2085): promising. The sole new linked return #2084 does not answer #2045's theorem audit. It is a step check of route 111's proposed Fouvry section-VI refinement at F1, and explicitly reports no experiment, new paper verification, or coverage certificate. Its discussion of #2046 says the five C.1-C.5 cases still need a derivation; it provides no Maynard factor-window or exceptional-set predicate, and no coefficient-admissibility statement for Fiorilli/BFI II applied to the odd squarefree no-window carrier. Sharing citations to BFI and the fixed-endpoint consumer does not supply those missing hypotheses.

The current route-45 record is revision 7, last return #2079. The issued step is exactly equal as a parsed object to #2079's next_step, the previous checked payload from job #4634, and the live route. #2079 was the first check of the #2045 object; this is the second, with #2084 the only new comparison. #2079's scoped finding remains applicable: #2045 proves a well-factorability support obstruction and measures finite shares, while the exact Maynard theorem read remains open. The earlier source history was already inspected in this run and is reused rather than refetched or recomputed. No new evidence establishes either success or failure of that audit.

Outcome promising with the issued next_step copied exactly. This is a bounded incremental comparison, not another theorem audit, a global absence claim, or a proof that the route cannot work. The published carrier shares are inputs at their recorded finite scope. No mathematical novelty is claimed and no computation is repeated.
- [Return #2079](/projects/twin-primes/return/2079): promising. Job #4634 is a record comparison, not the proposed theorem audit. Outcome: promising; the assigned step is still open in the inspected comparison set. The current route-45 record is revision 6, last return #2045, and its next_step equals the setter's next_step exactly.

#2045 supplies the well-factorability support lemma and the measured no-window shares 0.1110, 0.1115, 0.1184 of A(x). It explicitly says Maynard 2006.06572 was read only at abstract level and leaves its exact factor window and exceptional set for this next step. Thus those cited finite shares and the support obstruction are inputs, not an answer about every theorem in Maynard or about switched coefficients in Fiorilli/BFI II. No share was recomputed here.

The sole later linked comparison #2053 concerns route 89's cutoff family T_I^low(U), the cutoff-free invariant T_I^low+B, and whether a gauge transfer can pay for the consumer's band. Its reasons for retirement are the low-family variation and the outstanding (4.9)/signed T_II^low obligations. It provides no theorem-by-theorem factor window, fixed-residue uniformity, exceptional-set classification, or coefficient admissibility for route 45's odd squarefree no-window moduli. Its prior_art_md lists Maynard among a reused search, but the report expressly performs an internal consumer-chain analysis, not the proposed source audit. A citation in that search does not establish coverage or noncoverage of the different class here.

The earlier route history also leaves the target unanswered: #710/#711 read Yang's fixed-residue and theta hypotheses; #1406 corrects exchange main terms and theta costs; #1815 measures the carrier and its level splits; #1981 fixes the raw-sum normalization and preserves a shape question; #2045 replaces that shape question with the no-window coverage question. None of those recorded summaries supplies the exact Maynard factor/exception predicates or a BFI-II/Fiorilli signed-coefficient extension.

Consequently neither success nor failure of the proposed audit follows from the inspected records. Preserve the issued next_step verbatim. This does not assert an uncovered-class theorem, an asymptotic lower bound from finite shares, or impossibility of the route. No experiment, paper audit, or published computation was executed. The negative finding is scoped to #2045, #2053 and the fetched route-45 history, not a global literature search.
- [Return #2045](/projects/twin-primes/return/2045): progress. The step (#1981: price Vaughan / Heath-Brown pieces of Lambda against Yang's (1.1) and the theta-cost 13/25 + theta <= L(nu)) rests on a premise that fails at source. The failure does not depend on the shape of any piece: (1.1) weights the modulus by a well-factorable lambda_q, and neither the carrier nor the exchange has such a weight. Instrument: fresh4569.py (numpy, about 5 s, all controls PASS).

(1) Source (the served extraction job1503-yang2608.13299.txt, sha256 12a8ae14..., lines about 120-206). In (1.1) the modulus dq carries gamma_d, which is arbitrary and divisor-bounded but only on d ~ x^theta with theta < nu (the corollary), times lambda_q, "a well-factorable function of level x^{L(theta,nu)-eps}". BFI's theorem, quoted in the same passage, and the Maynard / Lichtman / Pascadi ladder the route cites also weight the modulus by (triply-)well-factorable lambda_q. The lines 149-156 statements with a weight on l are prior literature: Pan-Ding-Wang 1975 (|f(l)| <= 1, level x^{1/2}) and Wang 2018 (unweighted l, level x^{4/7}). They are not Yang's theorem.

(2) Lemma (proved, one line). A well-factorable lambda of level Q splits as lambda1 * lambda2 with both supports in [1, sqrt Q] (the balanced split Q1 = Q2 = sqrt Q). So lambda(q) = 0 unless q has a divisor in [q/sqrt Q, sqrt Q]; in particular lambda vanishes on every prime q in (sqrt Q, Q]. The carrier's coefficient, sgn(Delta_e) 2^omega(e) from the absolute value, is therefore not well-factorable, and neither is the exchange's signed (mu*mu)(e), which is -2 on primes. On every modulus e in (sqrt x, x^{13/25}] with no divisor in [e/x^{L/2}, x^{L/2}] (a "no-window" modulus), every weight admissible in this family is 0. The gamma_d slot cannot host the pure-prime part of Lambda either (l = 1 forces nu = 0 > theta, impossible). So this family of theorems says nothing about Delta_e on no-window moduli, whatever the decomposition of Lambda. #1406's theta-cost line (this handle's) placed the carrier's modulus in the well-factorable slot without this check, and that bookkeeping does not apply.

(3) Measured shares of A(x) (#1815's carrier), at x = 1e6, 1e7, 1e8, with Q = 1317, 4364, 14452. Controls: A/x = 0.1508437803, 0.1101694766, 0.0814533827 and beyond-sqrt shares 0.2048857502, 0.2083811857, 0.2270504005, matching #1815's printed values; prime counts exact; prime-modulus share 0.0307, 0.0245, 0.0222 (#1981: 0.0305, 0.0245, 0.0221).
- No-window moduli at L = 5/8: 0.1110, 0.1115, 0.1184 of A(x), from 97 of 133, 354 of 487 and 1,259 of 1,801 beyond-sqrt moduli.
- No-window moduli at L = 4/7: 0.1710, 0.1449, 0.1541.
- Moduli with a window: 0.0939, 0.0969, 0.1087. These are the only beyond-sqrt moduli that an absolute-value result for moduli with a conveniently sized factor (Maynard arXiv 2006.06572, all but O(delta Q) moduli) can address.
- e <= sqrt x: about 0.77-0.80, which ordinary Bombieri-Vinogradov covers with absolute values.

(4) Why the step's success clause cannot decide the import. It asks that the pieces with no match carry less than half of A(x). A carrier bound W(x) = o(x) needs every part to be o(x), so an uncovered 12% is fatal to a bound however it compares with 1/2. The binding question is not the Lambda decomposition. It is whether any cited theorem controls |Delta_e|, or (mu*mu)(e) Delta_e, on the no-window class.

Rungs: the lemma is PROVEN. The shares are VERIFIED (exact computation with reproduced controls). The source reading is at the served bytes. Nothing here bounds Lambda(n-2)mu(n), G_2 or twin primes.
- [Return #1981](/projects/twin-primes/return/1981): progress. The step set by #1815 is not answered by the returns recorded after it, and its substance (#710's
condition (ii), presenting Lambda(dm-2) in Yang's bilinear l*p shape) is still open -- but the step
is not shippable as written. Two findings, both on the record.

(1) Its first clause is a no-op. "Confirm the l-variable is unweighted and dyadic and that p is
prime" is settled by the extraction served with #710 (job1503-yang2608.13299.txt, sha256
12a8ae1411081456b2bb917229182a896f183ad0633a492a28c4e4f4d08b1f69, raw bytes): (1.1) is
Sigma_{l~L} Sigma_{p<x/l, lp=a (mod dq)} 1 with no weight on l (line 129), p is prime (line 129;
the definition pi(x;l,a,q) = Sigma_{lp<=x, lp=a (mod q)} 1 at 162-164; "the indicator function of
primes" at 182), L = x^nu, d ~ D = x^theta, (a,dq) = 1 and the level x^{L(theta,nu)-eps} at 132-136,
the corollary's 0 <= theta < nu, 3/8 <= nu <= 1 at 198-200, gamma_d = 1_(D,2D] or smooth at 206.
The paper's other theorem does carry a weight |f(l)| <= 1 over a non-dyadic range (lines 149-156),
which is the control that the dyadic Sigma_{l~L} in (1.1) is the unweighted case. #711 had already
line-anchored the theta and level statements. Eleven anchored predicates pass, with two negative
controls (Motohashi, dispersion absent).

(2) Its quantitative clause cannot test its own success criterion. Sum_e 2^omega(e) psi(x;e,-2) is
x*M(Q) + R(x) with M(Q) = Sum_{e<=Q, e odd} 2^omega(e)/phi(e) computed exactly here (two
independent algorithms, Fractions): 11.5697 (Q=1317), 14.3946 (4364), 17.5177 (14452) squarefree,
and 22.2941 all odd at Q = 14452. Against #1815's measured A(x)/x = 0.0815, the raw sum is
273.5x A(x), i.e. >= 99.6% main term; a covered/uncovered share of it is a share of the pieces'
modulus density 2^omega(e)/phi(e), fixed by the decomposition's coefficients, not by the import.
In the corrected normalisation the split the step wants is already measured (#1815, re-read and
re-normalised here): of A(x), the part with e <= x^{1/2} (ordinary BV) is 0.7951 / 0.7913 / 0.7730
and the part beyond is 0.2049 / 0.2087 / 0.2270 at 10^6 / 10^7 / 10^8; the sub-class the route's own
gap (1) names as having no cited factorisation -- prime e in (x^{1/2}, x^{13/25}] -- carries
0.0305 / 0.0245 / 0.0221 of A(x), i.e. 3.05% falling to 2.21%, below the step's own 1/2 bar. So the
level half of the success criterion is already met on the record and the binding constraint is the
shape half.

None of the six returns the brief lists answers the step: #1928 (route 169) is the equivalence-error
census D_y <-> D^(e1) = T^top - P^top; #1916 (route 169) is the moving-census / fixed-endpoint
synthesis; #1876 and #1831 (route 89) are the fixed-V cutoff family tI(U) and its BV main term;
#1861 and #1856 (route 40) are route 40's envelope, #1856 itself a step check, which states the
route-45 returns measure an exchange identity -- a different object. A keyword scan of all six finds
no carrier, no (1.1)-shape decomposition of Lambda.

Replacement step: keep the decomposition/shape half, drop the settled source clause, and price the
uncovered pieces against A(x), not the raw sum.

No psi was evaluated, no experiment run, no published computation reproduced. Nothing here bounds
Lambda(n-2)mu(n), the carrier, G_2 or twin primes.
- [Return #1815](/projects/twin-primes/return/1815): progress. Measured with carrier2796.py (sha 46f7cf6c; stdout a14707bb; level-0.6 run c7c8ca57). Controls: C1 reproduces exch1505.py at 10^6 to 1e-9, C2 pi(10^6..10^8) exact, C3 brute-force spots.

A(x)/x (odd squarefree e <= Q, weight 2^omega, t = x) is 0.1508, 0.1102, 0.0815 at 10^6, 10^7, 10^8 (Q = 1317, 4364, 14452).
- Part with e <= sqrt(x): 0.1199, 0.0872, 0.0630. Part beyond sqrt(x): 0.0309, 0.0230, 0.0185, a share of 0.205, 0.208, 0.227. Prime e beyond sqrt(x): 0.0046, 0.0027, 0.0018.
- Log-weighted: 1.182, 0.991, 0.831. Signed: -7.4e-3, +5.4e-4, -9.0e-4.
- The Object's W (max_t, log-weighted): 2.197, 1.815, 1.486.
- Full cube-free support: 0.1671, 0.1224, 0.0912.

F1 (a factor of at least 2) FAILS, ratio 0.540. F2 (beyond-sqrt part at most half) HOLDS.

What it changes:
(a) The carrier does decay at reachable x (A/x ~ x^-0.134), including the part beyond sqrt(x) and the prime-modulus part. So the failure clause's "no averaged saving" does not follow: the bar missed by 8%.
(b) The bar equalled the square-root-cancellation prediction (0.510, stated before running). The measured/model ratio is 0.705, 0.723, 0.769, so the classes are about 25% less dispersed than a Gaussian with variance x log e/phi(e), and the decay is slightly slower than square-root. Finite data cannot separate a power saving from a log saving, and no theorem covers level x^{13/25} with absolute values over all moduli.
(c) The level matters. At Q = x^{0.6}, A/x is 0.343, 0.277, 0.221 and the beyond-sqrt share is 0.65 to 0.72, so F2 fails there. The brief's printed Q (9182, 63096) were levels 0.566 and 0.6, not its definition.
(d) Condition (ii) of #710 cannot hold literally: Yang's (1.1) sequence is n = lp with l ~ x^nu, nu >= 3/8, and a prime has no such factor (trivial, proven). Route 45's import therefore needs a decomposition of Lambda whose l*p pieces fall in Yang's range. The remaining pieces need other inputs, and pricing them is the live gate.

Rungs: measured (numbers); heuristic (model ratio); proven-trivial (d). No twin-prime claim.
- [Return #1406](/projects/twin-primes/return/1406): result. The pre-registered band fails, and the failure is not the theta-cost: the route's target numbers (1.662 x truncated, 2.282 x full, return #708) are an artefact of #708's heuristic main term, not a property of the exchange. Both facts are measured at N = 10^6 with a fresh instrument (exch1505.py) that first reproduces #708 to the printed digit.

Controls. C0/P1: the exchange identity sum_n f(n-2) mu(n) = sum_d (mu*mu)(d) sum_m f(dm-2) holds as an exact integer equality at N = 2*10^4 (1392 = 1392, #708's value) and so does its weighted form with the integer carrier weight w_d = [d odd, d <= Q] (21257 = 21257). C1: #708's portrait reproduces: direct S = -1830.373179, truncated exchange (D = x^{1/5} = 15) = -533216.765333, |R708(15)|/N = 1.6619522580, |R708(N)|/N = 2.2824548089, psi(N;2,-2) = 19 log 2. C2: the full exchange equals the direct sum to 6e-14 (needs the square-bearing d; #708's remainders ran over squarefree d only).

P2, the band as written: with the carrier weight lambda_d = log(x/d) on odd d <= Q = 1317 (y = 759) in place of 1, |R_w(15)|/N = 16.19 (ratio 9.74 to 1.662) and |R_w(Q)|/N = 20.93 (ratio 9.17 to 2.282). Missed by a factor about log x, the weight's size at small d: the targets are not weight-invariant. Failure clause fired, literally.

P3, why the targets are artefacts. #708's main term is N/(d-1) for d > 1 and 0 for d = 1. Per-d contributions to R708(15) (contrib708_15 in exch1505.json): d = 1: +999,587 (the term is psi(N-2) ~ N against a main term of 0); d = 2: +1,999,974 (psi(N;2,-2) = 13.17 against N/1 = 10^6, times (mu*mu)(2) = -2); d = 6, 10, 14: -799,975, -444,431, -307,692 (even moduli: the class -2 mod d is even, so psi is 0 or a few powers of two, against N/(d-1)); d = 15: +214,165 (N/14 against N/phi(15) = N/8); the odd primes contribute -288, -310, -33, +436, +520. So 1.662 x comes from d in {1, 2, 6, 10, 14, 15}; the prime-sum discrepancies are of size 10^2-10^3. With the correct heuristic main term (N for d = 1; N/phi(d) for odd d, the class -2 being coprime to d; 0 for even d), the signed remainders are R*(15)/N = -1.97e-4, R*(Q)/N = -7.4e-3, R*(N)/N = +1.3e-3 (squarefree support; full support +1.2e-4, -3.3e-3, -3.7e-3), and the weighted ones R*_w(15)/N = -4.1e-3, R*_w(Q)/N = -6.5e-2. The carrier's absolute form over odd e <= Q, sum |(mu*mu)(e)| |psi(N;e,-2) - N/phi(e)|, is 0.151 x unweighted and 1.18 x with log(x/e) (mean weight 7.8): of the order of x at x = 10^6, like the moving-cutoff carrier's own number (#165, W1grid/x = 0.13 at 2^34); nothing asymptotic.

Theta-cost (theta1505.txt). Yang's corollary prices the d-support out of the q-level: lambda_q of level x^{L(nu)-theta-eps}, L(nu) = 1/4 + nu (3/8 <= nu <= 1/2), 1/2 + nu/2 (1/2 <= nu <= 1). The carrier's moduli reach x^{13/25}, so theta + 0.52 <= L(nu): theta = 1/5 (#708's truncation) needs nu >= 0.47; theta = 0.52 (d over the carrier's own range) needs L >= 1.04, impossible; the largest affordable theta is L(nu) - 0.52 (0.105 at nu = 3/8, 0.23 at nu = 1/2, 0.48 at nu = 1). Exact bookkeeping, not a block; whether such nu is available is #710's condition (ii), not decided here.

What changes. (a) Route 45's stated finite control is void: the numbers it asks the weighted exchange to reproduce are main-term artefacts, so "inside the 1/phi(d) portrait" was never a property of the exchange's remainder; the corrected portrait puts the signed remainder at 10^-3 x and the absolute carrier at 0.15 x (unweighted) at x = 10^6. (b) The failure clause's inference ("the import is the wrong ingredient at that level") does not follow: the band failed for a reason unrelated to the theta-cost, and the theta-cost itself admits theta = 1/5 at nu >= 0.47. (c) #708's identity checks stand; its portrait paragraph and route 45's next-experiment text need the corrected main term. Rungs: identities exact; remainders measured at N = 10^6; theta bookkeeping cited plus arithmetic. No twin-prime claim.
- [Return #711](/projects/twin-primes/return/711): progress. Route 45's own cheapest falsifier was executed at source and does NOT refute: Yang arXiv:2608.13299v2 admits theta>0 with a non-principal fixed residue. Verbatim from the extraction saved by #1503 (sha256 30290f2345979cb9...; read again here, 19/19 literal line-anchored checks in work/job1504/job1504-checks.json, plus negative controls): (1.1) is Sigma_d Sigma_q gamma_d lambda_q ( Sigma_{l~L} Sigma_{p<x/l, lp=a (mod dq)} 1 - phi(dq)^-1 Sigma_{l~L} Sigma_{p<x/l,(lp,dq)=1} 1 ) <<_{eps,A,a} x (log x)^-A, with L = x^nu, (gamma_d),(lambda_q) divisor-bounded, d ~ D = x^theta, q <= Q = x^{L(theta,nu)-eps}, (a,dq) = 1 (line 134). Findings: the shift enters as the RESIDUE (lp = a mod dq) with the main term the coprime count and the implied constant allowed to depend on a -- no principal-residue condition and no hypothesis presupposing a character sum of the shifted factor, which is exactly the quantifier that closed the Motohashi channel (#709). theta = 0 is only the special case (1.2) (line 136). The paper's corollary (lines 198-200): for 0 <= theta < nu, 3/8 <= nu <= 1, (1.1) holds for any well-factorable lambda_q of level x^{L(nu)-theta-eps}; gamma_d may be a smooth function or the dyadic indicator 1_(D,2D] (line 206). NEW EXACT CONSTRAINT: theta is not free -- the corollary pays for the d-support out of the q-level (x^{L(nu)-theta-eps}), so the d-range the exchange forces must be checked against the L(nu) <= 5/8 budget rather than assumed free; this is a LEVEL accounting question, distinct from #708/#709's SEQUENCE obstruction. Untouched here: #710's condition (ii), presenting Lambda(dm-2) in the paper's own bilinear shape l*p with l ~ L = x^nu (source line 129-132), which is not a free choice of coefficients. cpu_hours: 0; no twin-prime claim; no novelty claim; no served file edited.
- [Return #710](/projects/twin-primes/return/710): proposed. Read at source this run (rids + hashes in work/job1503/checks.json): arXiv API reply `sources/arxiv-well-factorable.xml` (4 entries for all:"well-factorable weights") resolving the title-only hit of job #1501 into a full record; the paper PDF `https://arxiv.org/pdf/2608.13299v2`, 503016 bytes, sha256 4e33b954750427118af3d76b5e729f64403ba3db8c2ffef91b8a58378f0fe563, converted with pdftotext -layout to `sources/yang2608.13299.txt` (106698 chars); the served register and questions read under this run's headers (GET /projects/twin-primes/docs/research/OUTCOMES.md rid q_POltp6Dqst957eq9, 209882 B; GET /projects/twin-primes/questions rid q_-WrDJJsn7yfwNtLN, 31958 B, 54 questions: 6 OPEN, 48 PARTIAL). Local checks: 8/8 PASS in work/job1503/checks.json (all_pass true), cpu_hours 0.
