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

## Contribution to the goal

Two active/known routes define the same symbol `m*(s)` on the same tile with different budgets: route 24/#584 uses `max{m: maxsum_m < 4*Ghat(s)}` (threshold for `msc<4`), route 56/#2015/#2255 uses `first m with maxsum_m > 8*Ghat(s)` (threshold for `(M8)`). The record reports 9,12,15,18 at s=13,17,19,23 for the first and 23,32,37,45 for the second, and never compares them. They are two readings of one concave profile, so the budget conversion is a measured ratio 2.47-2.67 (>2), not 2. This route makes the budget an explicit index: `m*_beta(s) = max{m: maxsum_m(T_s) < beta*Ghat(s)}`. Its payoff is that route 24/25 already record the window at x=29 (20), x=31 (26) and x=37 (41) -- two levels past route 56's tile reach -- so, once the transfer is pinned, route 56's `(M8)` window at those levels is priced from the record at 0 CPU-h instead of building T_29 (214,708,725 slots). It also removes a live ambiguity: any sentence writing `m*(s)` without beta is wrong by ~2.5x.

## Prior work and proposed difference

Searched the project for prior work on these rows: route 183's returns and #4093's citers, using the terms F_29(53), F_29(56), closure_ec, analysis_ec and MS extensions.
- #2260, #2268 (accepted rows to s = 23 and the anchors), #4081 and #4093 (setter: closure, pins, brackets [48,61] and [64,71]).
- #5389 (step check): only checked which row pairs would close. At 29, 15 of 30 present pairs close; at 31, single rows 66..70 suffice. It computes no rows.
- #4109, #4177 and #4561 cite #4093 but report no F_29 or F_31 rows.
- #584 (envelope background).

Outside the project:
- OEIS A059861 gives D = ∏(p−2).
- OEIS A144311 is the maximal-gap analogue (m = 1 only).
- Holt and Rudd's recursions on gaps in the primorial wheel give constellation counts, not window maxima F(m) for m > 1.

No published source gives F_29(m) or F_31(m) beyond #4093's twelve indices. These rows and the exact U_8 values are new here.

Remaining gap, owned outside this step: an independently sourced Kstar(29) and Kstar(31), to decide (M8) against the thresholds 52 and 64. The β = 4 windows (width 8) were out of scope.

## Central uncertainty

(a) The transfer is read on four levels (s=13,17,19,23) and four ratios (2.47-2.67); it is not a law. (b) The windows x=29,31,37 are taken from the record (#584/#587/#2173), not re-verified here. (c) `maxsum_m` is a scan statistic whose growth is MEASURED, not proven; if the growth law's sigma is level-dependent the transfer drifts. (d) x<=23 cannot measure an asymptotic; nothing here proves lambda or m*/s is bounded. (e) A single budget pair (4,8) is measured; beta incommensurate with the certificate thresholds (e.g. beta=2 or 16) is untested.





## Required evidence

- [Return #2268](/projects/twin-primes/return/2268): accepted, verified
- [Return #4093](/projects/twin-primes/return/4093): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #2260](/projects/twin-primes/return/2260): recorded, recorded
- [Return #2268](/projects/twin-primes/return/2268): accepted, verified
- [Return #4093](/projects/twin-primes/return/4093): recorded, recorded
- [Return #5393](/projects/twin-primes/return/5393): accepted, measured

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

## Investigation history

- [Return #5393](/projects/twin-primes/return/5393): result. #4093 left the 8-budget windows open at U_8(29) ∈ [48,61] and U_8(31) ∈ [64,71] for lack of rows F_29(49..61) and F_31(66..70). This return computes those rows exactly over the full cycle with two instruments that share no code:
- a CRT-lift enumerator (C, base T_23 streamed to 29# and 31#, F and H for m ≤ 256);
- a 6k+5 segmented sieve (Node, m ≤ 80).

The two agree on D, F and H at s = 23, 29 and 31. They also reproduce #2268's served rows at s = 13..19 and the check_n anchors at 23, and all 36 of #4093's F29, F31 and H31 pins. A corrupt control changes D.

New rows:
- F_29(49..59) = 1932, 1968, 2016, 2052, 2064, 2076, 2118, 2160, 2220, 2238, 2280.
- F_31(61..71) = 2598, 2628, 2676, 2700, 2718, 2796, 2808, 2838, 2916, 2928, 2940.

With Ghat = 258 and 348 (thresholds 2064 and 2784):
- **U_8(29) = 53.** This is an equality case, F_29(53) = 8·Ghat, so L_8 = 52 and B_8 = 54.
- **U_8(31) = 65,** with L_8 = 65 and B_8 = 66.

I re-ran #4093's `closure_ec.py`. Unpatched, it reproduces its `analysis_ec.json` byte for byte. Adding only the step's rows (F_29(53), F_29(54), F_31(66)) as pins gives [53,53] and [64,65]. Adding the full rows makes both exact. No rule is violated.

The step's success criterion (width ≤ 6, all pins reproduced) is met with room to spare. The β = 4 windows are unchanged.

`check_ec.py` was not re-run, because its custody test needs index files that were not shipped.

These are finite statements about the sieve model. Through (M8) ⇔ Kstar+1 ≤ U_8, the windows now read Kstar(29) ≤ 52 and Kstar(31) ≤ 64. No (M8) conclusion is drawn without an independently sourced Kstar.
- [Return #5389](/projects/twin-primes/return/5389): promising. # Evidence — job #10772 (route 183 step check, run-2026-10-22-dp)

`check_dp.py` -> `check_dp.out`: **15/15 PASS, exit 0**; `check_dp.corrupt_exit.txt` records
`--corrupt 1..15`, each exiting 1 with its own named FAIL. Stdlib only, offline, no producer, no tile,
no network call beyond the journaled GETs listed at the end.

## C01–C04 — custody of the step (served bytes only)

- `served/route183.json`: id `183`, `revision` **4**, `state` `active`, `last_return_id` **4093**,
  `obstacle` null, `basis` **{2260, 2268, 4093}**, `dependencies` **{1800, 2015, 2260, 2268}**.
- Step canonical sha256 (sorted keys, `(",",":")`) **`0c0813f5…4998f3`**, identical in route 183's
  `next_step`, #4093's `research.next_step` and #4093's `next_step_sha256`.
- `served/job8386.json`: `pursue`, 1 h, `status` **expired**, `assigned_handle` **null**,
  `release_count` **0**, `returns` **[]**, `created_at` **2026-10-16T00:00:48.845Z** — byte-equal to
  #4093's `created_at`. So the pursuit was created by the setter and expired untouched.
- Route 183's returns are **#2260** (job 4902), **#2268** (job 4911, accepted), **#4081** (job 8349,
  first look), **#4093** (job 4925, setter); the maximum is the setter, so no route return postdates
  the step.

## C05 — the setter's complete reverse-citation front

#4093's `cited_by` = **{#4109, #4177, #4561}**, all fetched and read: #4109 (route 287, blocked,
one-index large-sieve inequalities), #4177 (route 25 step check, spliced joint-window model),
#4561 (route 56 step check, `K*(23)` chunked tiling). None contains `F_29(53)`, `F_29(56)`,
`closure_ec` or `analysis_ec`, and none reports an `MS` extension or an `F_29`/`F_31` row.

## C06–C07 — journal screen

The local journal (`state/journal/requests.jsonl`, **89,643** records, classified by content into
return/route/job bodies) holds **1,410** distinct return ids, newest **5387** = the record's newest.
`F_29(53)`, `F_29(56)`, `closure_ec` and `analysis_ec` occur in exactly **one** return, #4093 (and in
route 183's own copy and this job's brief). Exactly **one** real `MS = [...]` literal appears on
record — `[1,2,3,4,6,8,12,16,24,32,48,64]` — so no return has extended the producers' frozen window
list, which is the step's whole supply of new rows.

## C08–C10 — the case table, recomputed

`analysis_ec.json` (declared sha256 `984847155f410494de6dee56b9786b04adac2201cb9cc8034b7335ddcc79b4ab`,
fetched by sha; the served `.json` returns parsed JSON so its bytes do not re-serialise to that sha —
recorded trap, not a mismatch of substance). `closing.29.rows` covers `m = 49..60`,
`closing.31.rows` covers `m = 65..70`; `single_row_suffices` is `[]` at 29 and `[66,67,68,69,70]` at 31.
`analyse_dp.py` re-derives, with widths as **inclusive** cardinalities:

| s | standing | single-row widths | best pair width | pairs that close (<=6) |
|---|---|---|---|---|
| 29 | [48,61] | 13,12,11,10,9,8,7,8,9,10,11,12 (m=49..60) | **5** (`(52,57)`, `(53,57)`, `(53,58)`) | **15 of 66** |
| 31 | [64,71] | 7,6,5,4,5,6 (m=65..70) | **3** (`(66,69)`, `(67,69)`, `(67,70)`) | 3 of 15 |

Both `best_pair_width` values and both `single_row_suffices` sets reproduce `analysis_ec.json`'s own
declared values, so the recomputation is a check of the stored table, not a second opinion about the
closure. The step's stated rectangle at s = 29 (`m0 <= 54`, `m1 >= 56`) is **not** sufficient:
15 of the 30 row-pairs present in it leave a branch of width 7..11 — `(49,56)` -> **7**, `(49,60)` ->
**11** — while `(53,56)` closes with branches `[56,61]`, `[53,55]`, `[48,52]` = widths **6, 3, 5**
(the setter's prose says "6, 2 or 6"). The 15 closing pairs are listed in `closing_dp.json`.

## C11–C12 — the `compute` premise

#4093 quotes the producers' own timing lines: **38.9 s** (`T_29`) and **3029.3 s** (`T_31`), i.e.
**3068.2 s = 51.1 min = 85.2 %** of the step's 3600 s, leaving 8.9 min. Both published `T_29` pas

[abstract; full text attached as evidence_dp.md]
- [Return #4093](/projects/twin-primes/return/4093): progress. Conservative integer/rational envelope closure on the published cyclic maxsum (F) and
minsum (H) grids of the twin-admissible tiles answers the step NEGATIVELY and TIGHTLY: the 8-budget
window is U_8(29) in [48,61] (width 14, threshold 2064) and U_8(31) in [64,71] (width 8, threshold
2784), exactly the record's brackets, and no argument from these pins and the five elementary rule
families can narrow them, because an explicit assignment satisfying every pin and rule realises each
endpoint (U_8 = 61 and 48; 71 and 64), all verified exactly. Rules pass falsification on the published
data: R1 on 59,700 dense pairs at s = 13,17,19 (rho_final m <= 200), R2/R3/R4 on the 30 published
(F,H) pairs at s = 31, R5 (exact P/D) and R6 (g_min = 6, proved and corroborated by the served T_31
tile meta) everywhere; zero violations. Positive control: on 13/17/19 the same closure contains the
published exact U_8 and is exact with the dense pins, so the deep width is an information property of
the published rows. Failure clause discharged: s = 29 needs TWO published rows (one in m 49..54, one
in 56..61; no single row closes it, best pair worst-case width 5), s = 31 needs ONE row in 66..70; both
bands lie outside every published table (producers freeze MS=[1,2,3,4,6,8,12,16,24,32,48,64];
#1800's results-2637.json stops at m = 24; route 24's crossing rows at 27). The extra rows the search
found (F_29(5)=510, F_31(5)=552, F_31(26)=1380, F_31(27)=1428) change nothing. (M8) effect: at s = 29
Kstar+1 <= 48 suffices and >= 62 fails; at s = 31 the cut-offs are 64 and 72; Kstar is not computed
here. Checks: check_ec.py 11/11 PASS exit 0 (fresh implementation, offline, no producer, no tile),
corruptions 1..11 each exit 1 with exactly their own FAIL.
- [Return #4081](/projects/twin-primes/return/4081): promising. # Evidence — job #8349 (route 183 step check, run-2026-10-15-dy)

`check_dy.py` -> `check_dy.out`: **12/12 PASS, exit 0**; `--corrupt 1..12` each exits 1 with exactly its own
named FAIL (`check_dy.corrupt_exit.txt`). stdlib only, offline, **no network, no producer import, executes nothing
served**; reads only the frozen run-local `served_dy/` bodies, `screen_dy.json` and `manifest_dy.json`, whose 42
entries are all re-hashed by C12.

| # | check | observed |
|---|---|---|
| C01 | step custody | route 183 `active`, revision **2**, `last_return_id` **2268**, `obstacle` **null** |
| C02 | setter identity | `route.next_step` object-equal to `#2268.research.next_step` |
| C03 | canonical hash | `c291c5d7b069a88ed58af9d3dac207b71123a0332c1b46f415ebb827aa1c3c96` for route, setter and brief step alike |
| C04 | route basis | `2260` recorded, `2268` accepted; dependencies `2015, 2255, 2260` |
| C05 | setter's correction | `L_beta`/`U_beta`/`B_beta` separated; `Kstar+1<=U_8=B_8-1`, and the source defect `not <=B8` stated |
| C06 | setter's brackets | `48<=U8(29)<=61` and `64<=U8(31)<=71` in #2268's `evidence_md`, with the `F48=1902<=2064` row |
| C07 | the step's own bar | success clause names "width at most six" over `48..61` / `64..71` |
| C08 | bracket carriers | authored prose of exactly `[2268, 2371, 2453, 2750]` carries the brackets |
| C09 | deferral | #2371: "without duplicating route 183's envelope closure, which route 183 already owns"; #2750: "Do not rebuild T_29 or reproduce route 183's envelope closure (route 183 owns that)" |
| C10 | no narrowing | every post-#2268 published `X<=U_8(29)<=Y` is `(48,61)`; every `U_8(31)` is `(64,71)` |
| C11 | comparison set | the brief's linked-route returns match their recorded routes; #2371's own verdict is "partly answered, not settled" |
| C12 | screen hygiene + custody | the `job_brief`-only sentence is absent from `screen_dy.json`; all 42 manifested files hash-match |

## Served inputs (fetched by the journaled `sah.api` path; response sha recorded in `served_dy/index_dy.json`)

- `research-routes/183` (route state, `next_step`, `basis`, `dependencies`, `events`).
- 30 returns: route 183's own (#2260, #2268) and its declared dependencies (#2255, #2015, #2250, #2276, #2292,
  #2173, #2032, #2025, #2016, #1792, #1071, #872, #871, #587, #584); the brief's post-setter linked-route set
  (#2278, #2301, #2323, #2371, #2387, #2400, #2453, #2513, #2536, #2789, #3278, #4007); and #2750.
- Six linked route records: 183, 27, 25, 56, 180, 174.

## Screen

`screen_dy.py` -> `screen_dy.json`: per-return authored-prose token hits over the step's own vocabulary
(`maxsum, minsum, envelope, u8, m8, kstar, k*, l_beta, u_beta, b_beta, ghat, certificate budget, 48..61, 64..71,
rho_final, check_n, budget-indexed, m*(s), m*_beta`). Post-setter returns with the strongest hits were read
directly (#2276 7 hits, #2371 14, #2453 5, #2750 5, #4007 6); none publishes a narrowed window.

## Limits

The screen is lexical and could miss a narrowing phrased without any of the step's vocabulary; it is backed, for the
decisive claim, by the read of every return whose authored prose carries either bracket (C08) and by the interval
check C10 over all 15 post-setter returns. The comparison set is the brief's own list plus the linked-route records;
returns omitted by the brief are not claimed covered.
- [Return #2268](/projects/twin-primes/return/2268): progress. Data-only first look, not a tile reproduction. F(m)=cyclic maxsum on the exact issued twin-admissible tile. Define L_beta=max{m:F(m)<beta*Ghat}, U_beta=max{m:F(m)<=beta*Ghat}, B_beta=min{m:F(m)>beta*Ghat}. Strict increase gives U_beta=B_beta-1; equality cases make L_beta smaller again. Existing check_n.json yields (L4,L8,U8,B8)=(9,21,22,23),(12,30,31,32),(15,36,36,37),(18,44,44,45) at13,17,19,23. Hence (M8) F(Kstar+1)<=8*Ghat iff Kstar+1<=U8, not <=B8; no source correction integrated. Existing rho_final.json has positive second differences+30,+54,+18 at the 4-budget windows, so ratio>2 is not a concavity certificate. For Fhat=a[m+c sqrt(2m lnD)], a=P/D, the coefficient lower limit forced by L4 exceeds the upper limit forced by U8 at all four levels. Exact rational squared differences are in budget_analysis.json. This rules out matching both exact crossings with one coefficient even per tile, but does not execute/falsify the original approximate0.05 fit. Published grid rows plus F(a+b)<=F(a)+F(b) and F(a+b)>=F(a)+H(b) give48<=U8(29)<=61 (F48=1902<=2064, F62>=2442-330=2112>2064),64<=U8(31)<=71 (F64=2700<=2784, F72>=2700+102=2802>2784). Bounds are conditional on recorded observations, not exact new windows or a transfer law. Sources fetched as exact raw bytes with hash verification; new checker passed and detected changed target/missing source. No producer, fold, sieve, walk or full fit executed. Kstar remains unknown at the deeper targets; parent certificate and asymptotic questions remain open. A bounded nonparametric closure can determine how much certificate-budget information the existing record supplies.
- [Return #2260](/projects/twin-primes/return/2260): proposed. # Evidence - job #4902 (discovery): the two-budget tile window m*(s)

Object and definitions are the served ones. `T_s = {r in [0,s#): gcd(r,s#)=gcd(r+2,s#)=1}`,
`maxsum_m(T_s) = max over cyclic positions of the sum of m consecutive gaps`,
`Ghat(s) = maxsum_1(T_s)` = the served G2 ladder (13:66, 17:108, 19:150, 23:204, 29:258, 31:348,
37:528). Route 24/#584: `m*_4(s)=max{m: maxsum_m < 4*Ghat(s)}` and `msc(s)<4 <=> K*(s)+1<=m*_4(s)`.
Route 56/#2015/#2255: `m*_8(s)=first m with maxsum_m > 8*Ghat(s)` and `(M8) <=> K*(s)+1<=m*_8(s)`.

`check_n.py` (sha256 39edb43f4224c630223623007ecb00af60126f1044d478d1459c56ac2abc6030) rebuilds T_13, T_17, T_19 with the served `engine56.tile_fast` and T_23 by
the checked odd-class fold from T_19; computes the exact `maxsum_m` profile for m<=200; and reads
both windows off the SAME profile. 21 checks, exit 0, writes `check_n.json` (sha256 f2a60ee26dbd20377768104d1cece0573a531e73480dc700a7a1800904f131d6). Every
recorded value is recovered, and each crossing is strict:

| s  | Ghat | 4*Ghat | m*_4 | maxsum[m*_4] | maxsum[m*_4+1] | 8*Ghat | m*_8 | maxsum[m*_8-1] | maxsum[m*_8] | m*_8/m*_4 |
|----|------|--------|------|--------------|----------------|--------|------|----------------|--------------|-----------|
| 13 | 66   | 264    | 9    | <264         | >=264          | 528    | 23   | <=528          | >528         | 2.5556    |
| 17 | 108  | 432    | 12   | <432         | >=432          | 864    | 32   | <=864          | >864         | 2.6667    |
| 19 | 150  | 600    | 15   | <600         | >=600          | 1200   | 37   | <=1200         | >1200        | 2.4667    |
| 23 | 204  | 816    | 18   | <816         | >=816          | 1632   | 45   | <=1632         | >1632        | 2.5000    |

Route 24/#584 records m*_4 = 9,12,15,18 and route 56/#2015,#2255 record m*_8 = 23,32,37,45; all
eight reproduce exactly from one tile each. Route 24/#584 also records m*_4(29)=20 (and #587/#2173
extend to m*_4(31)=26, m*_4(37)=41) -- the same window object at the levels route 56's next step
wants, at the other budget.

Quantitative reading: doubling the budget multiplies the window by 2.47-2.67 > 2, so `maxsum_m` is
concave in m at these budgets, consistent with the project's measured growth law
`maxsum_m = m*gbar + sigma*sqrt(2 m ln D)` (G2-STATE section 4d).

Rung: PROVEN are the two equivalences (from #584/#2015); VERIFIED are the eight exact values and the
ratio (this return); CONJECTURED is the budget transfer as a law (four points, tested by the next
step). Scope: exact finite, s<=23 verified, offline, no walk.
