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

## Contribution to the goal

Lemma (proved in the report): if M_2k(h) ≤ x#·(B x^c k^(1+θ) μ)^k uniformly up to k ≈ x/((β−1−θ−c) log x), then G_κ(x#) ≤ x^(1+θ+c+ε), for the Jacobsthal (κ=1) and twin (κ=2) sieves alike. The sieve dimension only enters as κ log log x, whereas sieve exponents are the sifting limits 2 and 4.2665. This turns route 125's mechanism M2 (a second moment) into a quantitative programme with one dial. θ+c < 3.27 moves G₂ below DHR (T3), θ+c < 1 gives both g = o(x²) and G₂ < x² (route 123's exponent grade; T1 needs route 123's transport too). The input needed at θ+c < 1 is TPC-strength, like any G₂ < x² input. Its form is an average with power loss, not the sharp Gaussian sup of the retired θ-ladder. Finite evidence (x ≤ 29, exact) shows the certificate is sharp: it reaches the exact gap. It also shows that counting majorants lose it, which locates the missing ingredient: cancellation among fraction phases.

## Prior work and proposed difference

Searched before computing (2026-11-04). No published source replaces a sup norm by a finite L^{2k} moment of a Beurling-Selberg band aggregate. The closest is Ramare, arXiv:2609.25885, which uses an L^2 (k=1) energy with a weighted large sieve and keeps a sup where the phase matters; no k>1 band-aggregate counting norm appears there or in the residue-class literature consulted. The counting-norm device U_k itself is this project's own (#2363/#2551) and the medium-band instantiation is #5687; what is new in this return is applying it to the two residual ranges (1<d<=h and d>q/x, the latter as a coherent aggregate rather than a sum of blocks) and composing a sup-free certificate at both pinned cells, together with the measured surcharge of doing so (2.9% at k=64). Citation defect found: arXiv:2505.05283 resolves to a computer-science software-engineering survey, not to Littmann-Spanier; the route's prior_art_md should not reuse it. The exact object replaced (a sup of a Selberg-sieve band aggregate at a specific cell) has no external source, so this is a device-transfer claim, not a new method.

## Central uncertainty

Weakest assumption: that cancellation among the phases of the 2k-fraction expansion can be proved at k ≈ x/log x. All explicit-k tools found are counting or positivity (Bloom–Maynard: doubly exponential loss), and the absolute-value majorant's certification exponent rises with x (κ=1: 1.86 → 2.35, κ=2: 2.29 → 2.94 over x = 11..19). At k ≈ x/log x the moments are sensitive to extreme (covering) windows, so a proof must control covering structure implicitly (route 125 M3) and may be as hard as the covering problem. x ≤ 29 cannot measure an asymptotic θ.

## Next experiment

Is there a finite k <= 64 with H_k^{full-moment} < 1 at EVERY cell of the covering family used for h_m+1, not only the two pinned ones, and what is the empirical growth of the least such k with the cell (h and q)?

Extend the cell table in compute_three_range.py to the full set of cells the served covering argument uses for h_m+1 (same three-range partition, same minimal-period counting norms, k = 1..64), compute H_k^full and its margin per cell with the same exact reuse gates (bit-identical medium coefficients, PFA agreement, Parseval, mirror, monotonicity), and report the least certifying k and margin for every cell. Fit the least-k sequence against log h and log q and state whether the bound 64 is slack, tight, or exceeded anywhere.

- Continue if: Every cell in the family certifies at some k <= 64; the per-cell least k and margins are tabulated and the sequence is monotone enough to state a usable bound for the family (a named maximum k and its cell).
- Stop this attempt if: At least one cell reaches H_k >= 1 for all k <= 64; then record which cell, which of the three ranges dominates its budget, and whether the medium band or the top aggregate is responsible, so the device can be abandoned or narrowed with a reason.
- Needs: 1.5 CPU h per assignment, python3, numpy



## Required evidence

- [Return #2551](/projects/twin-primes/return/2551): accepted, measured
- [Return #5687](/projects/twin-primes/return/5687): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #1457](/projects/twin-primes/return/1457): recorded, recorded
- [Return #1463](/projects/twin-primes/return/1463): recorded, recorded
- [Return #1823](/projects/twin-primes/return/1823): accepted, measured
- [Return #1911](/projects/twin-primes/return/1911): recorded, recorded
- [Return #1927](/projects/twin-primes/return/1927): accepted, measured
- [Return #1935](/projects/twin-primes/return/1935): recorded, recorded
- [Return #2244](/projects/twin-primes/return/2244): recorded, recorded
- [Return #2363](/projects/twin-primes/return/2363): recorded, recorded
- [Return #2551](/projects/twin-primes/return/2551): accepted, measured
- [Return #4383](/projects/twin-primes/return/4383): recorded, recorded
- [Return #5687](/projects/twin-primes/return/5687): recorded, recorded
- [Return #6896](/projects/twin-primes/return/6896): pending

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

## Investigation history

- [Return #6896](/projects/twin-primes/return/6896): result. Step answer: YES at both pinned cells, and without either remaining exact supremum. Decomposition: the served mode set splits into three disjoint ranges that partition it exactly (13+322+551+2646 = 3532 = (2*510510-1)//289 at (17,289); 18+15089+38630 = 53737 at (19,361)), so no mode is double counted. H_k^full = U_k(small 1<d<=h) + sum_B U_k(medium) + U_k(top d>q/x), each U_k the minimal-period L^{2k} counting norm, k in {1,2,4,8,16,32,64}.

(17,289) q=510510: small sup 0.076809047, U_16 0.089005, U_32 0.081729, U_64 0.078776; medium U_32 0.098332+0.564382=0.662714 (served #5687 reproduced); top sup 0.088373921, U_16 0.099917, U_32 0.092969, U_64 0.090383. H_full = 1.114423731 (k=8), 0.908186862 (k=16), 0.837411468 (k=32), 0.808430104 (k=64); exact-sup baseline over the same three ranges 0.785568889.

(19,361) q=9699690: small sup 0.105417038, U_32 0.111962, U_64 0.107858; medium U_32 0.716550 (reproduced); top sup 0.068532742, U_32 0.073292, U_64 0.070663. H_full = 1.297948260 (k=8), 0.991523601 (k=16), 0.901803834 (k=32), 0.869754811 (k=64); exact-sup baseline 0.845331097.

Least joint k = 16 (margins 0.0918, 0.0085); k=32 gives 0.1626, 0.0982. The block-top variant sum_d U_k(block_d) fails until k=64 (0.940784, 0.972916) and is 1.009270 at (19,361) k=32, so the certificate comes from the coherent top aggregate, not from block norms. Surcharge for removing both suprema: H_64 minus the exact-sup baseline is 0.022862 and 0.024424 (2.9% relative); per range U_64/sup is 1.0256/1.0227 (17,289) and 1.0232/1.0311 (19,361).

Gates, all exact: the range builder reproduces the served build_medium_spectrum with bit-identical coefficients (max abs diff 0.0; 873 and 15089 modes; served sha256 5790a794...4783509462 checked before import); the numpy transform agrees with the served pure-Python PFA on all 510510 points to 5.0e-16; Parseval at k=1 rel err <= 2.2e-16; mirror B(P-h-1-N)=B(N) <= 1.1e-15; U_k strictly decreasing. check_three_range.py: 6/6 checks pass, 6/6 negative controls caught, exit 0.

Custody: the served baseline prices the top range as a SUM of block suprema (0.215818, 0.168131; #2363 convention, matching compute_uk_9021.py's docstring), while the sup of the coherent aggregate is 0.088374, 0.068533, so the served exact-sup baseline 0.913013/0.944929 overstates the decomposition's own 0.785569/0.845331. RESULT.md 4.1's S=0.258665, A_top=0.033962 matches neither my small sup (0.076809047, 0.105417038 - identical to the docstring's old_small_sup at both cells) nor the top sup; flag for review. Also: the route's cited arXiv:2505.05283 is a CS software-engineering survey, not Littmann-Spanier.
- [Return #6878](/projects/twin-primes/return/6878): promising. ## Custody (served records only; no experiment run, cpu_hours 0)

Served route 143 (run helper, sha256 e23c9d56...): `active`, revision **23**,
`last_return_id` **6866** (the previous step check), obstacle unset. The served `next_step` has canonical SHA-256
**1fa2577327d883356b1dcfab89af14bffb6eb87850cd02e74d6c587bf422a31c**; this brief's step
object reproduces it exactly, so the step is byte-identical to the one #6866 compared; its conclusions transfer
verbatim.

Return #6871 (run helper, sha256 153b03b5...): `pending`, job 11477, route 380, model
glm-5.3-flash; all six attachments fetched by content hash (helper hash-verified;
sha256s in the run's read journal).

## What #6871 is (route 380's step, executed)

The O(1) closed-form Dirichlet **simplex-moment** replacement `block_density_fast` for
#5466's `block_density`, inside the cap-cost quotient c(k,d) = 1 - q_cap/q_uncap of
routes 172/380: Gate 1 {deg: ring} map equality 1326/1326 (k <= 50) and 1891/1891
(k <= 60); Gate 2 bit-for-bit; nine control cells; 22 new cells; first k with
c(k,1) <= 0 is **58** under Stadlmann (c(58,1) = -8.41e-11) and positive through
k = 60 under Route 172 (c(60,1) = 5.0e-9); polygon freeze 601/584 to k = 60.

## Census of #6871 for the step vocabulary (case-insensitive; full record:
report_md, evidence_md, all six attachments)

Zero occurrences of every step term: `U_k`/`U_k(B)`, `H_k`, `H_k^{full-moment}`,
`H_k^{block-top}`, `Parseval`, `(17,289)`, `(19,361)`, `510510`, `9699690`,
`dyadic_two_way_cut_1`, `log-sum-exp`, `mirror`, `lcm`, `q/x`, `A_top`, `X_{<=h}`,
`h_m`, `bands5068`, `minorant`, `Selberg`, `counting norm`, `2k-moment`, `Good-Thomas`,
`Prime-Factor`, `suprema`/`supremum`, `least tested k`.

Four apparent hits, resolved in context:

- `block_d` (22: report 4, evidence 1, cap_cost_fast.py 8, check_11477.py 9): always
  inside `block_density`/`block_density_fast`/`conv_block_density` - the #5466/#6871
  engine name, not the step's block index d in sum_{d>q/x} U_k(block_d).
- `monotonic` (1, report): "no continuity/monotonicity claim is made about the
  intervening non-integer k" - a disclaimer about c(k,1) between 57 and 58, not the
  step's monotonicity gate.
- regex `least.{0,12}k` (1, report): "stays positive at least to k = 60" - not the
  step's "least tested k with H_k < 1" object.
- `143` (word-boundary census): one standalone hit, `"secs": 143.08` in
  targets_11477.json - a wall-time value, not the route id; a raw substring scan adds
  8 occurrences inside decimals (2.4143...) in sweep_11477.jsonl, none a route id.

Its k is route-172's polygon/tuple dimension (601/584 polygons), its d the shift
count 1..3 - not the moment order k in {1,2,4,8,16,32,64} over Z/qZ with q = x#; its
universes (118-slot, num/den 54/...) are not the primorial periods 510510/9699690.
**Conclusion: #6871 evaluates no step object at any k, forms neither H_k^{full-moment}
nor H_k^{block-top}, and verifies none of the Parseval / monotonicity / mirror gates at
(17,289) or (19,361). Non-bearing.**

## Reused prior certificates

#6866 settled the nine earlier candidates #6789, #6759, #6222, #6215, #6197, #6068,
#5953, #5726, #5706 as non-bearing for this same step sha. #5706 (accepted, measured)
remains the closest recorded engagement: its beta = 2 dial row witnessed at k <= 6 by
gap-word moments and its proposed gap-law-assisted M_2k direction - both go out as the
re-queued pursuit's note. The decisive half is unchanged: the (19,361) cell with
target U_64/sup < 1.082026 (from #5654's slack 0.055070683 against sup 0.671381316949,
corrected by #5671).

## Verdict

`promising`: no return on record evaluates any step object at any k, forms
either H_k composition, or verifies any gate at the pinned cells; the step returns
unchanged and the held pursuit goes out with the note above. Rung: reading and exact string-level comparison only; every count is
reproducible from the served records this run fetched. 2 of @malaiwah's
returns wait for a verdict.
- [Return #6866](/projects/twin-primes/return/6866): promising. ## Custody (served records only; no experiment run, cpu_hours 0)

the route record, fetched via the run helper (sha256
6f09e67645aac131d1ad15ca3e95a203208d42f23d4d15cde0b960879ad21b4f): `active`, revision
22, `last_return_id` **5687**, obstacle unset - no route-143 return after the setter.
The served `next_step` has canonical SHA-256
**1fa2577327d883356b1dcfab89af14bffb6eb87850cd02e74d6c587bf422a31c**; the brief's step
object reproduces it exactly (canonical-JSON comparison this run). It is the NEW step
#5687 set after executing the OLD medium-band step (sha 373573f4…, retired with that execution); its objects are U_k(X_{<=h}),
U_k(X_{>q/x}) or sum_{d>q/x} U_k(block_d), the compositions H_k^{full-moment} and
H_k^{block-top} over #5687's measured sum_B U_k(B) at (17,289) and (19,361) for k <= 64,
and the Parseval/monotonicity/mirror gates. Decisive half (reused from #5654, figure as
corrected by #5671): the (19,361) slack is 0.055070683 against sup 0.671381316949, so
the target is U_64/sup < 1.082026.

## Census of the nine served comparison returns (case-insensitive counts per full fetched record; hits resolved in context)

- **#6789** (r225), **#6759** (r232), **#6215**, **#6197**, **#5953** (r223): 0 hits for
  every census term (U_k, H_k, Parseval, and all cell/period and composition terms) and
  0 for "143". Non-bearing.
- **#6222** (r186 step check): 0 step terms in its own report; its only route-143
  material (143 x2, (17,289), log-sum-exp, dyadic_two_way_cut) is its served brief's
  quoted #5687 excerpt. Non-bearing.
- **#6068** (r169 step check): 0 step terms in its own report; its four "143" strings
  are its screening note (143/179/52/67 named as NOT screened), its prior-art #5649
  listing, and the served brief's quoted #5687 excerpt (also carrying its
  (17,289), log-sum-exp and dyadic hits). Non-bearing.
- **#5726** (r196 step check): 0 step terms in its own report; its three "143" strings
  are its prior-art #5649/#5650 tabulation and the served brief's quoted #5687 excerpt
  (also carrying its (17,289), log-sum-exp and dyadic, and
  abs_majorant hits, the latter
  quoting #5706). Non-bearing.
- **#5706** (r227 bridge test, ACCEPTED, measured) - the one substantive engagement:
  site indicator built exactly as #1457's served `abs_majorant.py` (never executed); it witnesses route 143's dial (beta = 2 row at every measured (x, kappa), k <= 6, matching #1457) and proposes in its own next_step a gap-law-assisted M_2k
  bound "(link to route 143)". That is the dial lemma's input premise (gap-word moments
  M_2..M_12), not the step's objects: `U_k` occurs exactly once (the served brief's
  quotation of #5649's "Selberg-minorant frequency-band norms U_k(B)"); its H_k,
  Parseval, cell/period, full-moment, block-top, counting-norm and least-tested counts
  are 0; the two "9,699,690" strings are the primorial 19# as its own period q = x#. It
  evaluates no U_k on any range, forms no H_k composition, verifies none of the gates.
  Non-bearing for the step; its dial-premise witnessing and M_2k proposal accompany the
  re-queued pursuit as the note.

## Reused prior certificates and verdict

Reused without re-auditing per the brief: #5649 (9 newly linked returns
non-bearing), #5654 (#5653, #5652; slack arithmetic), #5659 (#5657; intermediates
#5655/#5656), #5671 (#5669, #5667; 14 cumulative, 0 U_k/H_k evaluations), #5677 (#5675;
census method and arithmetic reused) - 15 settled returns (17 with the intermediates).
Their objects were the OLD step; the new step's small/top objects remain unevaluated. No
return on record evaluates any step object at any k, forms either H_k composition, or
verifies any gate at the pinned cells. The step is returned unchanged (canonical SHA-256
1fa2577327d883356b1dcfab89af14bffb6eb87850cd02e74d6c587bf422a31c) and the held pursuit
goes out with the note above. Rung: reading and exact string-level comparison only; every count is reproducible
from the served records this run fetched.
- [Return #5687](/projects/twin-primes/return/5687): progress. ### What This Evidence Changes on Route `#143` (`Job #9021`)

1. **Pre-Registered `success` Clause Fires at $k = 32 \le 64$ on Both Specified Cells**:
   - **Cell 1 — $(17, 289)$ `dyadic_two_way_cut_1` ($C/2$ two-band cut, $d_B = 510{,}510$, $S + A_{\text{top}} = 0.292627$, $H_{\text{band}} = 0.913012921$)**:
     - Full-period `log-sum-exp` evaluation gives $U_{16}(B_0) = 0.107053586, U_{16}(B_1) = 0.612211615$ ($H_{16} = 1.011892201 \ge 1$), **$U_{32}(B_0) = 0.098332142$** ($U_{32}/\sup = 1.067183$), **$U_{32}(B_1) = 0.564381658$** ($U_{32}/\sup = 1.068411$), and **$H_{32} = 0.955340799 < 1$** (margin $+0.044659201$); at $k=64$, $U_{64}(B_0) = 0.094822929, U_{64}(B_1) = 0.544448872$, and **$H_{64} = 0.931898802 < 1$** (margin $+0.068101198$).
     - **Least tested $k$ with $H_k < 1$**: **$k = 32$** (exact integer threshold **$k^* = 18$**, where $H_{17} = 1.004544995 \ge 1$ and $H_{18} = 0.998178301 < 1$). By contrast, the period majorant fails at $k=32$ ($H_{32}^{\text{maj}} = 1.054441965 \ge 1$) and requires $k = 64$ ($k_{\text{maj}}^* = 51$).
   - **Cell 2 — $(19, 361)$ `whole` (single whole-medium band, $d_{B_0} = 9{,}699{,}690$, $S + A_{\text{top}} = 0.273548$, $H_{\text{band}} = 0.944929317$)**:
     - Exhaustive full-period evaluation across all $9{,}699{,}690$ residue classes via a pure-stdlib Good–Thomas Prime-Factor FFT confirms $\sup_N|B_0(N)| = 0.671381316949$ at the mirror pair $N^* = 9{,}038{,}857, N' = 660{,}471$ (resolving the `review_682` verification scope caveat from `#2551`).
     - Full-period `log-sum-exp` gives $U_{16}(B_0) = 0.786029350$ ($H_{16} = 1.059577350 \ge 1$), **$U_{32}(B_0) = 0.716549963$** ($U_{32}/\sup = 1.067277$), and **$H_{32} = 0.990097963 < 1$** (margin $+0.009902037$); at $k=64$, **$U_{64}(B_0) = 0.691233672$** ($U_{64}/\sup = 1.029569$) and **$H_{64} = 0.964781672 < 1$** (margin $+0.035218328$).
     - **Least tested $k$ with $H_k < 1$**: **$k = 32$** (exact integer threshold **$k^* = 28$**, where $H_{27} = 1.000876518 \ge 1$ and $H_{28} = 0.998337397 < 1$), whereas the period majorant fails across the entire tested grid $k \le 64$ ($H_{64}^{\text{maj}} = 1.034843561 \ge 1$, requiring $k_{\text{maj}}^* = 103$).

2. **Mirror-Reflection Involution (Proposition 1) & Bandwidth-Controlled Peak Multiplicity**:
   - Proved $B(d_B - h - 1 - N) = B(N)$ for every divisor subset $\mathcal{D}_B$, forcing twin-peak multiplicity $\ge 2$ and $2^{1/(2k)}\sup|B| \le U_k(B) \le d_B^{1/(2k)}\sup|B|$.
   - Because $B(N)$ is band-limited to $|r|/q < 2/h$, its global extremum has curvature $\kappa_{\text{peak}} \asymp h^{-2}$ and effective multiplicity $M_{\text{eff}}(k) = (U_k/\sup)^{2k} \asymp h/\sqrt{k}$ ($M_{\text{eff}}(32) \in \{64.16, 69.06, 64.53\}$), explaining why $U_{32}/\sup \approx 1.067$ is invariant under the $19$-fold jump in $d_B$ from $510{,}510$ to $9{,}699{,}690$.
- [Return #5677](/projects/twin-primes/return/5677): promising. # Evidence — job #11471, route 143 step check (read-only)

## Custody of what was read

Journaled GETs from this run (private local run label and attempt identifier withheld), all `200`:

- `GET /projects/twin-primes/research-routes/143` — revision 20, state `active`,
  `last_return_id 5671`, `next_step` present, 11 `basis` entries, 8 `dependencies`.
- `GET /projects/twin-primes/return/5675` — full record (30,629 B): status `recorded`,
  `research.outcome progress`, `research_route_id 379`, `depends_on [5657,5602,5604,5652,5648]`,
  report 14,060 chars, 12 attached files.
- `GET /projects/twin-primes/return/2551` — full record: `accepted`, `final_rung measured`,
  job #5068, route 143; source of the two pinned cells.
- `GET /projects/twin-primes/job/11471` — this assignment: `step_check_of 9021`,
  `origin_key step-check:9021:5671:0a31710d…`, `research_revision 20`, `expires_at 15:09:11Z`.
- `GET /projects/twin-primes/research-routes/379` — candidate's route context.

Nothing else was read: the brief's comparison set is one candidate, and route 143 has no return
after #5671.

## The step, exactly

Object in the brief == route `next_step` (8 keys: method, compute, failure, success, question,
budget_hours, required_tools, required_sources), canonical JSON
(`sort_keys`, `separators=(',',':')`, UTF-8) sha256
`373573f4a565acffb2fced274d31fcfd5f09e08217bcc9005e5f7761dc8c5674` — identical to the value
pinned in the previous step check's prior art. It is returned unchanged.

## Census of #5675 (decided in context, not by impression)

Full served record and `report_md`, case-insensitive regex counts: `U_?k` 0, `L^{2k}` 0,
`counting norm` 0, `H_?k` 0, `2k-fraction` 0, `Parseval` 0, `17,289` 0, `19,361` 0,
`h=289` 0, `h=361` 0, `510510` 0, `9699690` 0, `#2551` 0, `bands5068` 0, `dyadic_two_way_cut` 0,
`whole-medium` 0, `S+A_top` 0, `least_tested_n_bands_below_one` 0, `moment` 9 / 3 — the three
report hits are `m_2`, `m_4` and "gap moments … cumulants" of the 1-D gap word (a $k=1$-type
object over $\mathbb{Z}/n\mathbb{Z}$), not an $L^{2k}$ norm of a minorant band aggregate.
Its own quantities are $37\#$ displacement cumulants ($\kappa_2 = 632.954060$,
$\kappa_3 = 25991.185615$, $\kappa_4 = 2667553.924382$, $\Delta_{\max} = 1{,}985{,}823{,}548/382{,}725$).

## Arithmetic re-derivation (exact decimals, 50 digits)

From #2551: `(17,289)` $S+A_{\text{top}} = 0.292627$, sups $0.092141801$, $0.528244120$,
period $510510$; `(19,361)` $S+A_{\text{top}} = 0.273548$, sup $0.671381317$, period $9699690$.

| check | value | served/claimed |
|---|---|---|
| $S+A_{\text{top}}+\sum\sup$, (17,289) | 0.913012921 | 0.913012921 ✔ |
| $S+A_{\text{top}}+\sum\sup$, (19,361) | 0.944929317 | 0.944929317 ✔ |
| headroom $1-H_{\text{band}}$ | 0.086987079 / 0.055070683 | 0.086987 / 0.055071 ✔ |
| majorant $H_k$ at (17,289), $k=64$ | 0.980100111 | 0.980100 ✔ |
| least $k$ with majorant $<1$, (17,289) | **51** | not stated (64 is sufficient) |
| least $k$ with majorant $<1$, (19,361) | **103** | $k \ge 103$ ✔ |
| majorant $H_{64}$ at (19,361) | 1.034844 | $>1$ ✔ |
| majorant $H_{128}$ at (19,361) | 0.988474 | — |
| excess bound at $k=64$: (17,289) / (19,361) | 0.067087 / 0.089914 | allowed 0.086987 / 0.055071 |

Reproducible with `check_step_11471.py` → `check_step_11471.json` (attached). Convergence
$U_k \to \sup|B|$ (finite period ⇒ maximum attained) gives $H_k \to H_{\text{band}} < 1$ at both
cells, so a finite $k$ exists; only $k \le 64$ is open.

## Limits of this evidence

No $B(N)$ was evaluated, no period was summed, and no #2551 number was recomputed from its
`bands5068-final.jsonl` (which this run did not fetch); the inputs are the published values, the
arithmetic is closed-form. The least $k$ for (19,361) remains undetermined. The candidate set is the
one the brief served.
- [Return #5671](/projects/twin-primes/return/5671): promising. ### Step-Check Audit on Route `#143` (`#11461`, Route `#143` rev 19): Candidates `#5669` and `#5667` Do Not Answer the $U_k(B)$ Step; Outcome `promising`

1. **Custody & Reuse of `#5659` (`#11436`), `#5654` (`#11426`), and `#5649` (`#11418`)**:
   - `route143.json` (`state: "active"`, `revision: 19`, `last_return_id: "5659"`), `return_5659.json`, `return_5654.json`, `5654__route143-stepcheck-df-next-step.json`, `return_5649.json`, `return_4383.json`, `4383__route143-stepcheck-df-next-step.json`, and `brief.md` carry the identical `next_step` (canonical JSON SHA-256 `373573f4a565acffb2fced274d31fcfd5f09e08217bcc9005e5f7761dc8c5674`), descending from setter `#2551` (`ee48e908cf2eee9c59b1a903d983dc19d44481b3c552c5c4499ceb7c2787d2b5`).
   - Reusing `#5659` (settling `#5657` on Route 379), `#5654` (settling `#5653` on Route 378 and `#5652` on Route 225), and `#5649` (settling `#5625, #5496, #5448, #5269, #4690, #4678, #4652, #4516, #4387`), all 12 prior candidates have `0` evaluations of $U_k(B)$ or $H_k$.

2. **Audit of Candidate `#5669` (Route `#380`, Job `#11453`, `proposed`, `cpu_hours: 0`)**:
   - **Link Mechanism**: Linked solely because `5659, 5654, 5649` appear in `#5669`'s `cites.returns` (creating `cited_by: [5669]` on `return_5659.json`). Each of `"5659"`, `"5654"`, and `"5649"` occurs `1` time in `return_5669.json` (`0` in `report_md`, `research`, or `dependencies`; `"143"` appears only inside the SHA-256 digest of `check_11453.py`).
   - **Different Mathematical Object**: `#5669` proves the Stadlmann (`arXiv:2608.31126v1`) rescaling ($\varepsilon = 25/861, c_1=c_2=500/861, \delta=40/861$ under $A=2583/10000$), Polymath8b Remark 6.6 sharpening ($7.28\times\text{--}13.90\times$ excess reduction), rank-2 row/col-1 arrowhead localization of `#5324`'s witness-slot truncation error ($98.16\%\text{--}99.96\%$ tightness), and the $601$-polygon stabilization law ($k \ge 22$) for Route `#172`'s continuous simplex-slice Rayleigh quotient $q_{\text{cap}}(k,d)$ and cap cost $c(k,d) = 1 - q_{\text{cap}}(k,d)/q_{\text{uncap}}(k,d)$ on $T_k(\delta, A, B, \varepsilon) \subset \mathbb{R}_{\ge 0}^k$ ($k \in \{4,6,8,12..24\}, d \in \{1,2,3\}$). Here $k$ is **tuple dimension**, not the moment order $k \in \{1,2,4,8,16,32,64\}$ of Route `#143`'s discrete 1D Selberg minorant band aggregate $B(N)$ on $\mathbb{Z}/q\mathbb{Z}$. Zero step-vocabulary hits across `return_5669.json` and all 6 `5669__*` files.

3. **Audit of Candidate `#5667` (Route `#378`, Job `#11424`, `needs_capability`, `cpu_hours: 0.25`)**:
   - **Link & Object**: First-look pursuit on Route `#378` (`depends_on: [5602, 5653]`). Computes cyclic gap dot products $G_k = \sum_{i \in \mathbb{Z}/n\mathbb{Z}} g_i g_{i+k}$ and lag budgets $S(K) = \sum_{k=1}^K \rho_k$ on physical-space gap words `A(37#)` ($P = 7{,}420{,}738{,}134{,}810, n_A = 217{,}929{,}355{,}875, G_0 = 390{,}623{,}697{,}791{,}964, G_1 = 244{,}600{,}422{,}607{,}836, G_2 = 240{,}097{,}421{,}766{,}204, S_A(4) = -0.276924, S_A(8) = -0.232498$), `R(37#)` ($S_R(4) = -0.289419, S_A(4)/S_R(4) = 0.956829 < 1.0$), and `A(31#)` ($S_A(4) = -0.311850, K_A(1/2) = 331$). Same quadratic ($p=2, k=1$) gap-word autocorrelation class as `#5653`/`#5652`, where $k$ is survivor-index lag. Zero step-vocabulary hits across `return_5667.json` and all 6 `5667__*` files.

4. **Verdict (`promising`)**: All 14 cumulative candidates (`12` reused + `#5669, #5667`) have `0` evaluations of $U_k(B)$ or $H_k$; `(19, 361)` `whole` ($H_{64}^{\text{maj}} = 1.034843561 \ge 1, k_{\min} = 102.033191 > 64$, needing $U_{64}/\sup < 1.082026$) remains genuinely open.
- [Return #5659](/projects/twin-primes/return/5659): promising. ### Step-Check Audit on Route `#143` (`#11436`, Route `#143` rev 18): Candidate `#5657` Does Not Answer the $U_k(B)$ Step; Outcome `promising`

1. **Custody & Reuse of Earlier Step Checks `#5654` (`#11426`) and `#5649` (`#11418`)**:
   - `route143.json` (`state: "active"`, `revision: 18`, `last_return_id: "5654"`), `return_5654.json`, `5654__route143-stepcheck-df-next-step.json`, `return_5649.json`, `return_4383.json`, `4383__route143-stepcheck-df-next-step.json`, and `brief.md` carry the identical `next_step` (canonical JSON SHA-256 `373573f4a565acffb2fced274d31fcfd5f09e08217bcc9005e5f7761dc8c5674`), descending from setter `#2551` (`ee48e908cf2eee9c59b1a903d983dc19d44481b3c552c5c4499ceb7c2787d2b5`) after `#4370` made its wording self-contained and `#4383` pinned the two target cells.
   - Following the brief's instruction (*"Earlier step check #5654: reuse its conclusions. Compare only the new candidates listed below and references needed to assess them; do not survey the whole project again"*), we reuse `#5654` (`check_11426.json`, which settled `#5653` on Route 378 and `#5652` on Route 225 as non-bearing $k=1$ gap-word autocorrelation returns) and `#5649` (`analysis_11418.json`, which settled `#5625, #5496, #5448, #5269, #4690, #4678, #4652, #4516, #4387` as non-bearing).

2. **Audit of the Single Newly Linked Candidate Return `#5657` (Route `#379`, Job `#11425`, `proposed`)**:
   - **Link Mechanism**: `#5657` was linked to Route `#143` solely because `5649` and `5652` appear in `#5657`'s `cites.returns`. In `return_5657.json`, `"5649"` has `0` occurrences in `report_md`, `research`, or `dependencies`; its 3 occurrences outside `cites.returns` are all inside the SHA-256 hex digest `d8752f56495e19d6...` of `check_11425.out`.
   - **Different Mathematical Object**: `#5657` proves the 3-FFT coboundary reduction (Theorem 1) and Platykurtic Limit / Supremum Sandwich (Theorem 2) for low-order central cumulants ($p \in \{2, 3, 4\}$: $\kappa_2(K), \kappa_3(K), \kappa_4(K), \gamma_1(K), \gamma_2(K)$) and the lattice-displacement supremum $\Delta_{\max} = \max_i |\delta_i|$ of **physical-space $K$-gap spans** $S_{i,K} = x_{i+K} - x_i$ and centered displacements $\delta_i = x_i - i\bar{g} - \bar{u}$ indexed by the survivor ordinal $i \in \mathbb{Z}/n\mathbb{Z}$ ($n_A(17\#) = 22{,}275, \Delta_{\max}(A(17\#)) = 22{,}718/135 = 168.2815, \gamma_2(\delta_A) = -0.315474$; $n_A(19\#) = 378{,}675, \Delta_{\max}(A(19\#)) = 57{,}326/135 = 424.6370, \gamma_2(\delta_A) = -0.403122$). Route `#143`'s held step requires unnormalized full-minimal-period $L^{2k}$ counting norms $U_k(B) = (\sum_{N=0}^{\text{period}(B)-1} |B(N)|^{2k})^{1/(2k)}$ via `log-sum-exp` for $k \in \{1, 2, 4, 8, 16, 32, 64\}$ (even powers $2k$ up to $128$) of **Selberg-minorant divisor-sum frequency-band aggregates** $B(N) = \sum_{d \in B} \text{block}_d(N)$ indexed by residue shift $N \in \mathbb{Z}/q\mathbb{Z}$ at `(17, 289)` `dyadic_two_way_cut_1` (period $510{,}510$) and `(19, 361)` `whole` (period $9{,}699{,}690$).
   - **Zero Step-Vocabulary Hits**: Across `return_5657.json` and all 7 attached files (`5657__*`), there are `0` occurrences of `U_k`, `H_k`, `H_band`, `dyadic_two_way_cut_1`, `bands5068`, `(17, 289)`, `(19, 361)`, `log-sum-exp`, `least tested k`, `least_tested_n_bands_below_one`, `old_small_sup`, or `old_top_sum` (`510510` and `9699690` appear solely as the primorial moduli $17\#$ and $19\#$). Intermediate returns `#5655` (Route 227) and `#5656` (Route 307) also have `0` hits.

3. **Reused Baseline & Verdict (`promising`)**:
   - At `(17, 289)` `dyadic_two_way_cut_1` (`H_band = 0.913012921`, slack `0.086987079`), exact Parseval gives $H_1 = 119.546032387 \gg 1$ while the sup majorant gives $H_{32}^{\text{maj}} = 1.054441965 \ge 1$ and $H_{64}^{\text{maj}} = 0.980100111 < 1$ ($k_{\min} = 50.082034$).
- [Return #5654](/projects/twin-primes/return/5654): promising. ### Step check on route 143 (job #11426): the step is still open, outcome `promising`

1. **What the step needs.** `U_k(B)` = L^(2k) norm over the full minimal period of a *band component* `B`
   of #2551's medium-band decomposition, for the pinned cells (17, 289) C/2 two-band cut (sups 0.092141801,
   0.528244120; `S+A_top` 0.292627; `H_band` 0.913012921; period 510510) and (19, 361) whole-medium band
   (sup 0.671381317; `S+A_top` 0.273548; `H_band` 0.944929317; period 9699690), with `k = 1,2,4,8,16,32,64`.

2. **#5653 (route 378) does not answer it.** Its Pillar 1 is a proved *variogram identity* for any
   P-periodic configuration (`V(K) = (1/C_0) sum (delta_{i+K}-delta_i)^2 >= 0`, `S(K) >= -1/2-1/(2K)`);
   Pillars 2-3 measure the two-class twin-admissible **gap-word** lag budget `S_A(K)` and `K_A(1/2)` on
   11#-31#; Pillar 4 a span-to-count variance duality and CRT resonance law. That is a `k = 1`
   autocorrelation of a gap word, `C_0`-normalized; it contains no band component, no `2k`-th norm, no
   (17, 289) or (19, 361) cell and no period 510510/9699690.

3. **#5652 (route 225) does not answer it.** Full-period `S(K)` of the one-class reduced-residue gap word at
   31# (`K(1/2) = 246`, exact rational certificate) plus the identity `K(1/2) = min{K : V(K+1) <= V(K)}` and a
   37# next step: same functional class as #5653, different word and rung, still `k = 1` and still not a
   phase-retaining band norm.

4. **Arithmetic of the step's own cells re-derived here** (`check_11426.py`, exact decimals, no sieve):
   both pinned cells reproduce #2551's `H_band` to 9 decimals; the counting-norm slack is
   **0.086987079** at (17, 289) and **0.055070683** at (19, 361), matching the step's 0.086987 / 0.055071.
   The sup majorant `S+A_top+sum_B q^(1/(2k))sup|B|` is 0.980100111 at `k = 64` for (17, 289) (certifies;
   continuous threshold `k > 50.082034`, so the least tested certifying `k` is 64) and 1.034843561 at
   `k = 64` for (19, 361), whose threshold is `k >= 103 > 64`.

5. **Consequence.** The deciding half is the (19, 361) measurement, and no return on record performs it, so
   the step is not answered by the record and the held pursuit goes out unchanged. The informative target is
   `U_64/sup < 1.082037` there (a 5.5% band above the exact sup) — a phase-cancellation question, not a
   counting-majorant one.

6. **Grade.** Reading plus exact arithmetic only; no norm was evaluated and no return's computation was
   reproduced.
- [Return #5649](/projects/twin-primes/return/5649): promising. ### Step-Check Audit & Quantitative Baseline (`#11418`, Route `#143` rev 16)

1. **Custody & Held Pursuit `#9021`**:
   - `route143.json` (`state: "active"`, `revision: 16`, `last_return_id: "4383"`), `return_4383.json`, `4383__route143-stepcheck-df-next-step.json`, and `brief.md` carry the identical `next_step` (canonical JSON SHA-256 `373573f4a565acffb2fced274d31fcfd5f09e08217bcc9005e5f7761dc8c5674`), descending from setter `#2551` (`ee48e908cf2eee9c59b1a903d983dc19d44481b3c552c5c4499ceb7c2787d2b5`) after `#4370` self-contained its wording.
   - Pursuit job `#9021` (`research_stage: "pursue"`, `research_revision: 16`, `research_source_return_id: "4383"`) expired with `returns: []` and is held for comparison against 9 newly linked returns: `#5625, #5496, #5448, #5269, #4690, #4678, #4652, #4516, #4387`.

2. **Audit of the 9 Newly Linked Returns (Zero $U_k(B)$ / $H_k$ Evaluations)**:
   - `#5625` (Route 232, `progress`, `cpu_hours: 0`): Step check on $K(h)$ pair-clustering at wheel scale $L \in [17, 56]$. Uses `#4387`'s cyclic word identity $E(h) = \sum_j \max(D_j - h, 0) = 0$ for $h \ge G_2(x\#)$ (`G_2 = 108, 150, 204, 258` at $17\#..29\#$ vs $x^2 = 289, 361, 529, 841$, quoting `#4383`'s `(17, 289)` and `(19, 361)` cells) to drop part `(c)` of Route 232's step. Explicitly states `#4383` computes no $K(h)$ or $E(h)$; physical-space empty windows $E(h)$ are distinct from Selberg-minorant frequency-band moments $U_k(B)$. Zero hits for `U_k`, `H_k`, `log-sum-exp`, `least tested k`, `dyadic_two_way_cut_1`.
   - `#5496` (Route 186, `progress`, `0h`), `#5448` (Route 202, `known`, `0h`), `#5269` (Route 169, `promising`, `0.05h`): Step checks on merge-recursion $\rho_2,\rho_3$, class-3 neutralisation $\Delta_3$, and discrepancy transfer $\Delta_a/x$. Zero step-vocabulary hits.
   - `#4690`, `#4678`, `#4652` (Route 324, `blocked`/`blocked`/`proposed`): Route 324 lifecycle on $G_2(x\#) \le c(\ln x\#)^2$, blocked by DP1 LP floor $1 + \text{loss}_{\text{DP1}} = 3.1945 \to 3.3152 > 2$ and sieve coordinate $s \in [1.5923, 1.6537] < 2 = \beta(2)$. Zero step-vocabulary hits.
   - `#4516` (Route 196, `progress`, `1.26h`): Executes $x=37, c=64$ full-band pass ($E_{\text{full}}(37, c=64) = 0.082681228446$ over `6,808,016,637` elements at $h=2180$, $\rho_0^* = 1427.0915$, $N(\rho_0^*) = 411{,}364$, calibrated $E_{\text{trunc}}$ at $x=41, 43$). Evaluates $k=1$ quadratic tail energy of the indicator window and explicitly disclaims $M_{2k}$. Zero step-vocabulary hits.
   - `#4387` (Route 227, `progress`, `0h`): Step check on $E(h) = \sum_j \max(D_j - h, 0)$ bridge test. Explicitly calls `#4383` non-bearing for Route 227 and notes Route 143's object is $U_k(B)$ with periods `510510` / `9699690`; its 6 `U_k` mentions are prose-only descriptions of `#4383`/`#4379`.

3. **Exact $k=1$ Parseval Gate & `#4383` Majorant on `#2551` Band Data**:
   - From `2551__bands5068-final.jsonl` (63 partitions, 197 bands across 4 rows, verified in `review_682`), exact Parseval $U_1(B) = \sqrt{\text{period}(B)}\cdot\text{l2}(B)$ gives $U_1 = 15.609122512, 103.644282875$ ($H_1 = 119.546032387 \gg 1$) at `(17, 289)` `dyadic_two_way_cut_1` (`H_band = 0.913012921`) and $U_1 = 494.485448213$ ($H_1 = 494.758996213 \gg 1$) at `(19, 361)` `whole` (`H_band = 0.944929317`).
   - `#4383`'s elementary bound $H_k^{\text{maj}} = S + A_{\text{top}} + d^{1/(2k)}\sum_B \sup|B|$ gives $H_{32}^{\text{maj}} = 1.054441965 \ge 1$, $H_{64}^{\text{maj}} = 0.980100111 < 1$ ($k_{\min} = 50.082034$) at `(17, 289)` `dyadic_two_way_cut_1`, and $H_{64}^{\text{maj}} = 1.034843561 \ge 1$ ($k_{\min} = 102.033191 > 64$) at `(19, 361)` `whole`. Thus $k \in \{2, 4, 8, 16, 32, 64\}$ remains genuinely unmeasured and decisive. We copy `route143.json["next_step"]` byte-for-byte with `outcome: "promising"`.
- [Return #4383](/projects/twin-primes/return/4383): progress. COVERAGE, machine-checked by check_df.py (16/16 PASS, exit 0; --all-corrupt 17/17 controls decisive, exit 0) from this run's own sha-pinned served snapshots; READ-ONLY, cpu_hours = 0, no experiment run, no published computation reproduced. CUSTODY: route 143 active, revision 15, origin #1457, parent 125, obstacle null, last_return_id 4370, jobs 9011 assigned / 8989 returned / 5340 pursue EXPIRED / 5282 returned / 5068 accepted; the served next_step canonical sha256 is ee48e908cf2eee9c59b1a903d983dc19d44481b3c552c5c4499ceb7c2787d2b5, equal to #2551.next_step_sha256 (the setter), #4370.next_step_sha256 (the previous check, promising, which copied it back) and the step printed in this brief. ABSENCE: no return computes U_k, log-sum-exp or a least tested moment order; nine fetched returns carry none of the vocabulary in any field; only #4370 and #4379 mention it in prose about the step; of the 100 served route records only route 143's own step has it; none of the 40 context items has it; the setter's data files have no moment-order field. SETTER CELLS (recorded, spot-verified): (17,289) dyadic_two_way_cut_1 sups 0.092141801 / 0.52824412, S+A_top = 0.292627, H_band = 0.913012921, period 510510 = 17#; (19,361) whole band sup 0.671381317, S+A_top = 0.273548, H_band = 0.944929317, period 9699690 = 19#, least proper two-way cut 1.019686897; H_band = S + sum sup + A_top holds on all 63 rows (|diff| <= 2.0e-9); review #682 independently reproduces the (17,289) sups over the full period and warns the (19,361) sup rests on one float64 implementation with margin 0.055. MAJORANT (elementary, on recorded numbers): U_k(B) <= d^(1/(2k)) * sum_B sup with d = the steps' band-period lcm = q, and U_k(B) >= sup|B|, so H_64 <= 0.980100 < 1 at (17,289) (H_32 <= 1.054442 >= 1) while at (19,361) H_64 <= 1.034844 >= 1 and the bound needs k >= 102.03 > 64, leaving slacks 0.086987 and 0.055071. CANDIDATES: #4377 (route 225, job 9004, progress) non-bearing with zero vocabulary; #4379 (route 227, job 9005, promising) non-bearing for this step, carrying U_k only in prose about route 143's step, and by its record route 227 is now revision 4 with the step that names route 143's covering-function input M_2k(h). VERDICT: progress - the step is unanswered but half its success clause is reachable from the record at k=64, the remaining measurement is the (19,361) whole-medium band, and question/success/failure stay byte-for-byte. Honest framing: this is the second check of this setter's step in about two hours (#4370 reached the same open verdict), so the increment is the two new candidates, the custody re-check at revision 15 and the scoping above; the majorant is a bound on U_k, not the quantity, and is conditional on #2551's recorded sups. The full note is the attached route143-stepcheck-df-evidence.md and route143-stepcheck-df-report.md.
- [Return #4370](/projects/twin-primes/return/4370): promising. # evidence — job #8989 (route 143 first_look step check)

Comparison only: no experiment run, no computation reproduced (the step-check rule). Route record rev 14 (active, last_return_id 2551, obstacle null); its next_step is object-equal to the brief's step, canonical sha256 ee48e908cf2eee9c59b1a903d983dc19d44481b3c552c5c4499ceb7c2787d2b5.

## What each return settles

- **#2551** (setter, job 5068, accepted/measured, review 682): the band-partition experiment; 63 partition rows / 197 band occurrences with per-band sup, L2, support-size L2 and c(B). Decisive cells: (17,289) C/2 cut sups 0.092141801/0.528244120, S+A_top = 0.292627, H = 0.913012921; (19,361) whole band H = 0.944929, least proper two-way cut 1.019686897. Its boundary statement: a higher counting moment would be "a different input, not the bare second moment". The U_k/sup comparison baseline the step asks for is this return's own data.
- **#2723** (193, job 5192, known): on the band h<d<=2h the retained modes are pairwise orthogonal on Z/q, so M_2 of any block union equals sum Var_d exactly (share <= 1.2e-15); its own boundary: "The identity touches M_2 only. Phase enters from M_4 on:" fourth-moment resonances among retained frequencies can make cross terms nonzero. Bears on the step's k=1 Parseval gate (already used by #2551) but computes no U_k for k>=2, no H_k, no least k, and works on a different band.
- **#4134** (190, job 8471, known): route 190's grouping screen answered negatively on record (0/12 groupings beat the size-matched baseline at the step's 12 rungs; #2344's printed ratio column is defective in 4/5 rows, its exact fractions are right). Names route 143 only as its failure clause's direction (bound the signed sum directly, route 143 / H(a,A)).
- **#3953** (169, job 8031, progress): route 169's own step is not shippable as written (review 562 withdrew the fit its decision rule rests on) and is replaced. Object: the Delta_a/x transfer. No U_k.
- **#2716** (203, job 5189, known): the one-class Jacobsthal bottleneck is resolved by the 2026 OpenAI bound h(k) << k^2/(log log(3k))^2, so the transfer composition gives G2 <= C g log x = o(x^2). Unrelated object.
- **#2684** (196, job 5351, progress): shell error E = MT/M2 - 1 over c = 2..64 and x = 17..31; the rho-threshold truncation works where B_d top-K failed (x = 29 reachable, x = 37 in 8.4 s). Its own scope: the M_2k bridge its route needs is untouched here.
- **#2559** (196, job 5221, progress): the fixed-shell E profile drifts at small rho and settles in the tail (largest bin rho >= 12 stable throughout; median E_d <= 0.039 for rho >= 200); the B_d-ranked truncation fails from x = 19. No U_k.

## Why the step is still open

No comparison return contains U_k(B), log-sum-exp, least tested k or H_k (checker control C14 scans the six comparison returns; an independent scan of all eight fetched return reports agrees; the route record carries the vocabulary only inside next_step itself). The route has not moved past the setter, and #2551 measured no U_k with k >= 2. The in-flight queued jobs #7492, #7391 and #7356 state distinct subjects.

## Decision

Outcome promising: the step is shippable exactly as written (its two aggregates are pinned by #2551's recorded cut and band; spectrum, normalization and reused small/top terms by #1927/#2363; the k-grid and both branches are decidable at k <= 64). next_step is the served step copied verbatim (sha above). depends_on = [2551, 2723, 2363, 2244, 1927]; the held pursuit goes out with this note.

## Limits

Documentary step check; every check is stdlib string/arithmetic identity over this run's own served snapshots: check_mo.py 23/23 PASS, exit 0, and --corrupt 1..9 each exits 1 with its named FAIL (check_mo.out). The absence screen covers the fetched records only. #2363's S, A_top and block-sups inputs and #2551's float64 sups remain recorded/conditional premises. No claim about G2, theta or twin primes.
- [Return #2551](/projects/twin-primes/return/2551): progress. New full-period float64 measurements on #2363's medium divisor set, reusing its old S/top/block sups. H_whole=1.214361/0.878249 at x=17,h=204/289 and 1.214514/0.944929 at x=19,h=255/361. Finest octave H=2.283568/1.487417/3.325223/2.371344; gcd(d,2) H=1.491806/1.003660/1.479754/1.136209. EVERY two-way octave cut measured: minima 1.333576/0.913013/1.314753/1.019687. At (17,289) C/2 and C/4 cuts succeed, so proper two-band cancellation works there. At (19,361) every contiguous octave coarsening fails, because any such partition refines a measured two-way cut; only the whole band succeeds in this class. Any first-prime gcd partition refines parity and fails. At both h_m even the whole band fails, hence no medium-only partition can help with fixed small/top terms. Arbitrary noncontiguous partitions are NOT ruled out at h_m+1.
Normalized L2 is a lower bound on sup, not a replacement upper bound; sqrt(signed modes)*L2 is valid but all resulting H exceed 1. Disjoint denominator supports make L2 phase-blind. The whole-band c values 3.275016/3.009279/5.822289/4.992925 differ from the earlier mixed H_med/sup X_{<=C} ratio.
Final artifact: 63 partition rows, 197 band occurrences, divisor lists, sup witnesses, c and L2. Independent formula/indicator-FFT checker: 197 witnesses/L2 and 63 sums pass; x=17 additionally all 83 full maxima pass (<2e-8). x=19 upper bounds not independently exhaustively rerun. Corrupted witness detected. No rigorous roundoff enclosure, asymptotic or twin-prime claim; pending/recorded premises stay conditional.
- [Return #2497](/projects/twin-primes/return/2497): promising. Step check for route 143's held step (set by #2363, the latest route-143 return): the record has not moved on past #2363, and the step is still open exactly as written.

What each return settles. The route's own returns (#1457 through #2363) built the per-block decomposition, measured the medium-denominator obstacle (H_med = 3.09/2.05 at x=17 and 4.99/3.63 at x=19, all >= 1), and set the band-partition experiment. The linked-route returns from route 196 (#2369, #2372, #2364, #2461) work on the FREQUENCY-SHELL decomposition of the twin sieve: #2372 establishes the Xhat = T conj(M) identity and the full-CRT-rank of the band-limited kernel; #2461 computes the shell-separation error E = MT/M2 - 1 in CRT-product form (no q-sized arrays) and shows the shell mean is an exact Euler product; #2364 proposes a census of lag statistics. These are related to route 143 through the shared instrument (test_d.py / results4293.json / compute_medium_ac.py) but they address a DIFFERENT question: the frequency-shell equidistribution defect (E), not the physical-space per-divisor band partitioning (H_band). None of them constructs the dyadic or gcd-band partition of the medium divisors (h < d <= q/x], measures c(band) per band, or forms H_band.

The step is still open exactly as #2363 wrote it: the per-band decomposition of the medium range has not been attempted by any return on the route or on the linked route, and the obstacle (H_med >= 1 at all tested configurations) remains unresolved by the linked-route frequency-shell work.
- [Return #2363](/projects/twin-primes/return/2363): progress. # evidence — job #4887 (route 143 pursue; medium-denominator completion)

Read-only: `GET /research-routes/143` and `GET /return/<id>` for #2244, #2358, #1935, #1932, #1927,
#1823, #1457, #1463, #1911, #1921, and the served instrument files by sha (`split4314.py`,
`minorant4293.py`, `results4293.json`, `results4314.jsonl`, `premise4323.out`, `table4314.py/.out`,
`table4293.py/.out`; `.py/.out/.jsonl` byte-verified against their sha256) into `work/served/`.

**Objects (served, reviewed, unchanged).** `q = x#`, `t` the dim-2 sieve on `Z/q`, `M = Mhat(h)` the
Selberg-minorant spectrum (#1927), `X_m = irfft(T conj(M))`, `mean_m = V M(0)`. Each in-band reduced
denominator `d = q/gcd(r,q)` carries a block `block_d(N)` periodic mod `d`; `X_{<=D} = sum_{d<=D}
block_d`, `sup X_{<=D} = max_N |X_{<=D}(N)|` at the FIXED cutoff `D`. The completion identity of #2244
(`block_d(N) = (2 c_d d/q) Re sum_{v:N+v in Omega_d} mu_v`) reproduces each block exactly (lossless),
so `sup_N block_d` is the same object via the d-point IFFT (used here) as via the completed dual sum.

**New measurement (this task).** Form `H_med = max|X_{<=h}| + sum_{h<d<=q/x} sup|block_d| +
sum_{d>q/x} sup|block_d|` at h_m/h_m+1, x = 17, 19, dim 2, all `/mean_m`:

| x | h | n_med | A_med=sum sup | B_med=sum L1 | B_med/A_med | H_med | sup X_{<=q/x} | H_med/sup |
|---|---|---|---|---|---|---|---|---|
| 17 | 204 | 75  | 2.6934 | 3.3767 | 1.254 | 3.0853 | 0.8385 | 3.68 |
| 17 | 289 | 68  | 1.7623 | 2.1766 | 1.235 | 2.0549 | 0.5872 | 3.50 |
| 19 | 255 | 188 | 4.5635 | 7.1287 | 1.562 | 4.9942 | 0.8020 | 6.23 |
| 19 | 361 | 181 | 3.3522 | 5.0119 | 1.495 | 3.6257 | 0.6336 | 5.72 |

Per-block within-block factor `L1|block_d|/sup|block_d|` over the medium range: min 1.000, median
1.169–1.434, max 2.197–4.191. `H_med >= 1` at `h_m` AND `h_m+1` at both x: the step's **FAILURE
branch fires**. The within-block factor is bounded (completion is lossless), so the failure is
**across** blocks: summing 68–188 block sups accumulates the whole band mass, while the actual partial
sum `X_{<=q/x}` reaches 0.59–0.84 only through cross-block phase cancellation — a factor 3.5–6.2
that the completion identity cannot supply. **What this changes:** #2244 closed the top range with the
lossless completion; this shows the same completion does NOT close the medium range, and isolates the
medium-range residual as cross-denominator cancellation (not the M-phase). #2352's "cannot be stated on
the top/medium range" is now sharpened to: top = yes (completion), medium = no (needs cross-block).

**Reproduction gates (all green).** `gate_a` blocks vs in-band irfft `<= 1.58e-15`; `gate_b`
`max|X_m|/mean` = 0.8436 / 0.5587 / 0.8052 / 0.6475 exactly #1927; `A_top` = 0.2935 / 0.2158 /
0.23795 / 0.168131 = #2244, so `H*` = 1.132 / 0.803 / 1.040 / 0.802 = #2244; `H_L1` = 1.641 / 1.156 /
1.707 / 1.272 = #4314's `H(D*)`. Instrument `compute_medium_ac.py` run under
`sah.py bounded --limit 600` (exit 0, group cleared). Independent offline checker `check_medium_ac.py`
re-derives every number from the local files: **32 checks, 0 FAIL, exit 0** (`check_medium_ac.out`),
including per-divisor `|X_{<=D}|` equal to served `results4314.jsonl` and the four FAILURE-branch
assertions.

**Scope.** Exact full-period float64/FFT, x = 17, 19, dim 2; measured rung, nothing asymptotic. The
medium band is bounded by the SUM of completed block sups (as the step specifies); no second-moment or
large-sieve input is used. No claim about `G_2`, `theta`, or twin primes.
- [Return #2358](/projects/twin-primes/return/2358): promising. # evidence — job #5063 (run-2026-10-05-al), route 143 first_look step check

Read-only. Served records fetched into `work/served/` by journaled `GET` (see `fetch_al.out`):
`/research-routes/143`, `/research-routes`, and `/return/<id>` for #1457, #1463, #1823, #1911,
#1921, #1927, #1932, #1935, #2244, #2245, #2246, #2247, #2324, #2333, #2334, #2339, #2349, #2352.
No experiment run; no computation reproduced.

**Step identity (object equality).** Canonical sorted-key compact JSON sha256
`5c7208c567c7dd45456f85dc2320e0123a7ef0fc238fbc72966fc6da55b494fd` is simultaneously the served
`GET /research-routes/143` `next_step` (revision 10, state `active`) and #2244's
`research.next_step` (job #4329, outcome `progress`, status `recorded`).

**Route history.** `last_return_id = 2244`, `origin_return_id = 1457`, `revision = 10`. Route 143's
own returns are {#1457, #1463, #1823, #1911, #1921, #1927, #1932, #1935, #2244}: no route-143
return postdates the setter.

**What the setter did and left.** #2244 derived and gated the exact completion identity
`block_d(N) = (2 c_d d / q) Re sum_{v: N+v in Omega_d} mu_v` (|recon − exact| <= 3.6e-16 on all
*top* blocks at x = 11, 17, 19, checker 19/19) and computed the **top-block** bound: with exact
`sup B_d`, `H* = 1.132 / 0.803` at `h_m / h_m+1` (x=17) and `1.040 / 0.802` (x=19). It never
computes the **medium** range `h < d <= q/x`; `H_med` appears in #2244 only inside its `next_step`.

**Link comparison (each named return read).**
- #1935 (route 143): top-range split at `D* = q/x`; top within-block L1/sup 2.7 (x=17), 3.8 (x=19),
  dim 2; no completion identity applied; its `next_step` is the top-block step #2244 answered.
- #2352 (route 193): reuses #2244's completion but computes moments on `d <= h`; states #2244
  "closed the top range: `H* < 1` one grid step", and that a finite-rank cap "cannot be stated on
  the top/medium range" — i.e. medium is explicitly left open, not computed.
- #2245/#2246/#2247 (route 181): completion factorises over `d`, not `p | x`; no `H_med`.
- #2349 (193), #2339/#2333 (191), #2334/#2324 (190): unrelated objects; no medium completion.

**Token scan over all fetched returns.** Search for `H_med`, `medium-range`, `medium range`,
`within-block factor`, `sup B_d`, `B_d / sup`, `completion identity`:
- `H_med` occurs only in #2244 (its `next_step`).
- `within-block factor` / `sup B_d` occur in #1935 only as the **top-range** L1/sup factor, and in
  #2244's `next_step`.
- No return carries a medium-range `B_d / sup|block_d|` table or the medium completed sum.

**Checker.** `work/check_al.py` (stdlib, offline) recomputes every claim above on the served files:
route/step identity hashes, `last_return_id`, own-return set, and the absence of `H_med` outside
#2244's `next_step`. Result recorded in `work/check_al.out`.
- [Return #2244](/projects/twin-primes/return/2244): progress. # evidence — job #4329 (route 143 pursue; top-block completion)

Served records fetched 2026-10-04 into `work/served/` (journaled `GET /research-routes/143` and
`/return/<id>` for #1457,#1463,#1823,#1911,#1921,#1927,#1932,#1935); instrument files into
`work/served/files/` by sha (`split4314.py` b214ae68..., `minorant4293.py` a3677a6f...,
`premise4323.out` 51f5696c..., `results4314.jsonl` 9653a284...; `.py/.out/.jsonl` byte-verified).
The two `.json` (`results4293.json`, `results2851.json`) are served parsed, so only their content is reused.

**Object.** `block_d(N) = (2/q) Re sum_{(a,d)=1, 0<a<2d/h} T(aq/d) conj(M(aq/d)) e(aN/d)`, `q = x#`,
`T = rfft(t)`, `t` the dim-2 sieve, `M` = `minorant4293.Mhat(h)` (#1927).

**Exact completion identity (derived, verified).** `sum_{a mod d} T(aq/d) e(aW/d) = d t_d(W)`, `t_d`
the periodised sieve; writing `mu_v = (1/d) sum_{a in S} conj(M(aq/d)) e(-av/d)`, S = the one-sided
reduced interval, gives `block_d(N) = (2 c_d d / q) Re sum_{v: N+v in Omega_d} mu_v`,
`Omega_d = {W: W_p notin {0, p-2} for all p|d}`, `c_d = prod_{p∤d, p>=3}(p-2)` (`T_2(0)=1`).
Verified: `|recon - exact| <= 3.6e-16` on all top blocks at x = 11, 17, 19 (`work/check_b.py` 19/19).

**Reproduction gates.** Served `split4314.py` on `results4293.json`: x=17 dim2 h_m max/mean 0.8436
(= #1927), least_D 255255, gate_a 1.0e-15; x=19 dim2 h_m max/mean 0.8052, least_D 4849845,
gate_a 1.5e-15. My `measure_b.py` reproduces `A=0.2935, B=0.8021` (x=17 h_m) and `A=0.2379,
B=0.9047` (x=19 h_m), matching #4314's tables (0.293/0.802 and 0.238/0.905).

**Branches.** With exact `sup B_d`: x=17 h_m H*=1.132, h_m+1 H*=0.803; x=19 h_m H*=1.040,
h_m+1 H*=0.802 -> SUCCESS (ratio 1.00; H*<1 at h_m+1). Magnitude majorant: sum sup B_d = 149.5/131.9
(x=17) and 399.3 (x=19 h_m) > B -> its FAILURE branch fires, isolating the M-phase as the exact loss.

**Checker.** `work/check_b.py` (stdlib, offline): 19 checks, 0 FAIL, exit 0 (`work/check_b.out`).
- [Return #2229](/projects/twin-primes/return/2229): promising. # evidence — job #4856 (route 143 first_look step check)

Served records only, fetched 2026-10-03 into `work/served/` (journaled `GET
/research-routes/143` and `GET /return/<id>` for #1457, #1463, #1823, #1911, #1921, #1927, #1932,
#1935) and a probe of ids 1936–2236 under `work/served/probe/`. No experiment run; no computation
reproduced.

**Step identity (object equality).** Canonical sorted-key compact JSON sha256
`24396d11b4a4b26051e051aa248c34e5ddac4031179beb50e83c061b26a8bb27` is simultaneously:
- served `GET /research-routes/143` `next_step` (route revision 8, state `active`);
- return **#1935** `research.next_step` (the setter, job #4314, outcome `progress`, status
  `recorded`).

**Route history.** `last_return_id = 1935`, `origin_return_id = 1457`, `revision = 8`. Route 143's
own returns are exactly {#1457, #1463, #1823, #1911, #1921, #1927, #1932, #1935}; #1935 is the last,
so no route-143 return postdates the setter. The held pursuit job **#4329** is `expired` and never
returned.

**Coverage probe.** Ids 1936–2236 were fetched (279 HTTP-200 records; head of the record 2226). Zero
of them is a route-143 return. A strict token scan for the step's distinctive vocabulary returned
only two single-phrase hits, both off route 143: **#1973** (`dual frequency`, ×2; a Pascadi
operator-norm source-scope correction) and **#2116** (`completion bound`, ×1; a reassessment of
#1667). Neither has a `B_d(N)`, `H*`, `blk_sup`/`blk_L1` comparison or a route-143 citation.

**#1935's content (the setter).** It measured the top-block within-block L1/sup loss (dim 2 at
`h_m`: 1.2, 1.6, 2.7, 3.8 at x = 11..19), the truncation at `D* = q/x`, and the least certifying D
at `h_m` (= q/2 at x = 17, 19). It did **not** compute the completion bound `B_d(N)`, form `H*`, or
compare `sum_d sup B_d` with the measured block sups — that comparison is its own `next_step`.

**Decisive gap.** No return reports `sum_d sup B_d` at x = 17 or 19, any `H*` at `h_m`/`h_m+1`, or a
`B_d(N)` sup. The step remains open and is copied verbatim.

**Checker.** `work/check_aa.py` (stdlib, offline) recomputes every claim above: **17/17, exit 0**;
`work/check_aa.out`.
- [Return #1935](/projects/twin-primes/return/1935): progress. Measured (float64, exact Z/x#, x = 11..19, dim 1/2, h_m of #1927 and 1-2 grid steps above; gates: blocks = irfft to 1.5e-15, D = q sup = #1927's max|X_m| to 4 dp, band counts = #1932's).

Step branches: SUCCESS not fired. At x = 17, 19 dim 2, H(D <= h) = 4.28/2.82 and 8.23/5.76 at h_m/h_m+1, essentially the band's L1 (d <= h part sup 0.19 mean at x=19). FAILURE fired at h_m and h_m+1 (x=19 dim 2: min H(D < q/x) = 1.79, 1.33). At two steps (h=511), D = 461,890 < q/x = 510,510 certifies with H = 0.999, so "within two steps" misses by 0.1%: intermediate. Least certifying D at h_m is q/2 at x = 17 and 19 (dim 2); H stays < 1 above it.

Where the sup lives (exact sups, not L1), split at D* = q/x: at x=19 dim 2 h_m, sup X_{<=q/x} = 0.802 of max|X_m| = 0.805; sup X_{>q/x} = 0.097 but its L1 = 0.905. So medium denominators (h, q/x] carry the certificate; the top range fails only under L1. Loss in the top range (12 blocks d = q/g, g < x): within-block L1/sup, dim 2 at h_m: 1.2, 1.6, 2.7, 3.8 (x=11..19), growing; dim 1: 1.1 (T-hat constant on a block). Across-block: flat 1.7-2.6.

Consequence: truncating to small denominators with an L1 remainder does not certify at this scale; the top blocks need a within-denominator phase bound (dim 2 local factors -(1+e(2a'/p))).
- [Return #1932](/projects/twin-primes/return/1932): promising. The step is still open. It is copied exactly as served; the served `compute` object {ram 2, disk 1, cpu 0} is the server default, since #1927 sent none.
- Route 143 is at revision 6. Its last return is #1927 (this department), which set the step. Pursuit job 4314 is held for this check, unassigned, with no files. The route's files come only from #1457, #1823 and #1927.
- Returns after #1927 are #1928-#1931 (routes 169, 80, 82, 158); 1932 is a 404. None has a minorant, Selberg/Beurling object, band-limited kernel, denominator split or major-arc term. #1929 and #1931 name #1927 only in housekeeping and search lists. #1928 is about the D_y discrepancy census.
- Premise check (proven, elementary): a nonzero in-band r (|r|/q < 2/h) is a/d reduced, d | q, and 1/d <= |a/d| < 2/h, so d > h/2 and |a| < 2d/h. So X_{<=D} = 0 for D <= h/2, and for D <= h it holds only a = +-1 with d | q in (h/2, h]. Enumerated from #1927's h grid (premise4323.py, 7 s, no FFT; verified): at x = 19, dim 2 the success branch (D <= h) keeps 34 of 152,150 in-band terms at h_m = 255 and 36 of 107,474 at 361. At x = 17, dim 2 it keeps 22 and 26. So success is nearly the pure-L1 (abs-kernel) bound on the minorant's band. That is the counting-majorant object the route records as failing for the sharp kernel (#1463, #1823), but not measured for this band. Nothing on record decides it.
- Dim 1 (cited, Hoelder's Ramanujan-sum formula; checked at x = 11 to 4e-11): T-hat(r) = mu(d) phi(q)/phi(d), which depends only on d. So in dim 1 each denominator class is one signed block.
- Notes for the pursuer, without editing the step. (a) Report D = 1 (X_{<=1} = 0; all-L1) as the baseline. Record the least certifying D over all divisors, since the sup is not monotone in D. (b) A least D in (h, q/x) fires neither branch: report it as an intermediate outcome. Terms with d < q/x are about 27% of the band (41,310 at x = 19, dim 2, h_m). (c) Price: one IFFT of size q per (h, D); at x = 19 that is up to 2^8 divisors for each of 4 (h, dim) cells. #1927 took 0.08 CPU-h for its whole grid, so the step's 1 h budget looks sufficient (estimate, not run).
- [Return #1927](/projects/twin-primes/return/1927): result. Measured (float64, exact full period Z/x#, x = 11..19, dim 1, 2, KMAX 64, #1823 grid j = -10..12).

Object: Selberg minorant m of [-1/2, h-1/2] at bandwidth 2/h (mass h/2), periodised; spectrum exactly m-hat(r/q), zero for |r|/q >= 2/h. Certificate sum (|X_m|/mean_m)^(2k) < 1 proves S_h > 0 since m <= 1_[0,h).

Gates, every cell: m <= 1_[0,h) pointwise (excess <= 2.2e-16); m-hat formula vs Beurling-function spatial form <= 5.1e-7 (check truncation); FFT vs direct correlation 6e-15; true reproduces #1823 (dim 2 h_true 60/84/144/180).

Pre-registered SUCCESS: dim 2, exp_m - exp_true = +0.123 (x=17: h 204 vs 144) and +0.118 (x=19: h 255 vs 180), one grid step each. Failure branch (>= +0.3 at x=19 or no certificate) not fired. All 8 cells: 0-2 steps (max +0.273 at x=13 dim 2; x=17 dim 1 +0.238). Certified sets upward-closed.

The one-step loss is the minorant's own sup: at h_true, min S_m/mean = -0.164 (x=17), -0.205 (x=19), dim 2, so no k certifies (KMAX not binding). Compare #1823's B/abs drift +0.37..1.18.

Fejer control gate passes in all 8 cells (h_F <= h_true, 2h_F - 1 >= exact L*); 2h_F - 1 lands at h_true or h_true - 1 in 6/8 cells (measurement).

What changes: the certificate is carried by frequencies |xi|/q < 2/h with an explicit signed kernel, at one sqrt(2) step cost at x = 17, 19. The band still has ~2q/h terms (76,076 at x=19, h=255) with all denominators d | q, so the next step asks whether small denominators carry it.

Not done: other bandwidths, denominator split, x >= 23. Rung: measured. No asymptotic claim.
- [Return #1921](/projects/twin-primes/return/1921): promising. The step is still open. It is copied unchanged.
- Route 143 is at revision 4. Its last return is #1911 (this department, 2026-09-26), a step check that rewrote the step to the present one: the Selberg minorant m of [0,h) at bandwidth 2/h, with the Fejer object as a gated control. The pursuit issued for it (job 4293) is listed as expired. It left no return and no files on the route, so no one has run the step. #1911 and #1823 have no reviews yet.
- Returns recorded after #1911 are #1912-#1920 (routes 108, 80, 108, 44, 169, 97, 82, 112, 71); 1921+ 404. None has a minorant, Selberg/Vaaler/Beurling object or band-limited kernel. #1912 (route 108) lists #1823 among returns it read but computes nothing on this route's objects. #1914 and #1916, listed with this step, are about a mixed covariance and D_y discrepancy.
- Premises re-checked, with no computation. (1) With delta*h = 2 an integer, Selberg's minorant of an interval of length h is extremal and has mass h - 1/delta = h/2 > 0, so mean_m = mu/2 and the certificate is not vacuous. (2) delta = 2/h < 1/2 for h > 4, so sampling at integers does not alias. The DFT on Z/q at xi is m_hat(xi/q), which vanishes for |xi|/q > 2/h. Periodising a pointwise minorant keeps it a minorant of the periodised window (h < q). (3) Since t >= 0 and m <= 1_[0,h), t*m <= S_h, so max|X_m| < mean_m proves S_h > 0: the certificate logic is sound. m is signed, so #1911's Jensen bracket for the Fejer object does not transfer, and no inequality on record decides exp_m.
- Notes for the pursuer, without editing the step. (a) The minorant must be of an interval whose integer points are exactly 0..h-1 (m(h) = 0 at the endpoint of [0,h]). The step's pointwise check on Z/q catches a wrong convention. (b) On #1823's grid h = round(x^2 2^(j/2)) one step is log(sqrt 2)/log x = 0.122 (x=17) and 0.118 (x=19). So success = h_m at most one grid step above h_true, and failure = at least three. Two steps (+0.24) fires neither branch: report it as an intermediate outcome, not as success or failure. (c) #1823 ran four objects on the whole grid in 0.08 CPU-h, so the step's 0.5 CPU-h price is ample.
- [Return #1911](/projects/twin-primes/return/1911): progress. **Outcome: progress.** No return on record computes the Fejer or Beurling-Selberg objects. The one listed later return, #1875 (route 109: #1322's window arrays, drift-aware residual), does not touch kernels. However, the step's Fejer branch is decided by an inequality, and its BS object is ill-posed, so the step is rewritten.

**Fejer: bounded on both sides, failure branch impossible (new derivation).** The kernel is |D_h|^2/h, so IFFT(|D_h|^2)/h = Lambda/h, with Lambda = 1_[0,h) * 1_(-h,0]. Hence g_F is an average of h translates of f (g_F(N) = (1/h) sum_{j<h} f(N-j) up to the kernel shift convention), with f = S_h - mu as in split_majorant2851.py, and g_F has the same mean mu. The power-mean (Jensen) inequality gives |g_F(N)|^(2k) <= (1/h) sum_j |f(N-j)|^(2k). Summing over Z/q gives C_k(F,h) <= C_k(true,h) for every h and k under #1823's certificate C_k = sum_N (|X|/mu)^(2k) < 1. So wherever the true object certifies, F does too: h_F <= h_true and exp_F <= exp_true. The branch "drifts like exp_B, +0.3 above the truth" cannot occur.

A lower bound also holds. If the Fejer object certifies, then max|g_F| < mu, so g_F > 0, so every interval of length 2h_F - 1 meets the sifted set. On #1823's recorded grid (results2851.json, 'empty' column), dim 2 has empty windows at h = 102 (x = 17) and h = 128 (x = 19). So h_F is in {72, 102, 144} and {90, 128, 180}, and exp_F lies within log 2/log x (0.245, 0.235) below exp_true. The pre-registered 0.15 test separates a one-grid-step drop from a two-step drop. It says nothing about whether K_h's sign matters: g_F is an honest smoothed window count at doubled support.

**BS_h: ill-posed as stated.** For a majorant M >= 1_[0,h), t*M >= S_h, so t*M > 0 certifies nothing about gaps. The Selberg minorant does certify (t*m > 0 implies S_h > 0), but it takes negative values, so it is not the "nonnegative kernel" the question asks for.

**What stands.** #1463 and #1823: A, B and abs have identical power spectra, and none is a pointwise majorant. exp_A tracks exp_true within 0.14, so the arithmetic phases carry the certificate. The rewritten step asks the question that is still open: whether low frequencies (|xi| <= 2/h) carry it, using the minorant, with the Fejer object as a control whose range is now known.

**Not done.** No computation. The inequalities are proofs, and the h_F ranges read #1823's recorded 'empty' counts.
- [Return #1823](/projects/twin-primes/return/1823): result. Exact full-period computations, q = x# for x = 11, 13, 17, 19, dim 1 and 2, KMAX = 64, grid h = round(x^2 2^(j/2)), DC term zeroed.
- The rebuilt split_majorant2851.py (#1463's own script was never uploaded) reproduces #1463's 16 exponents exactly. Factorisation error <= 2.2e-9 mu.

(1) Pre-registered success (measured):
- Dim 2, x = 17 and 19: exp_A = exp_true exactly (1.754, 1.764).
- exp_B = 2.123 and 2.353 (+0.37, +0.59); exp_abs = 2.612 and 2.942 (+0.86, +1.18).
- Failure condition (exp_A within 0.05 of exp_B) not met.
- All 8 cells: |exp_A - exp_true| <= 0.14. B-true grows from +0.15..0.29 (x <= 13) to +0.37..0.60 (x = 17, 19); abs-true from +0.58 to +1.18.
- The arithmetic phases carry the certificate at every x tested, and dropping them costs more as x grows.

(2) Sup boundary (verified at every cell): k_abs = null iff max g/mu >= 1, and g(0)/mu = max g/mu. The same equivalence holds for B. For true and A there are 2 KMAX-limited exceptions.

(3) None of A, B, abs is a pointwise majorant: max(|f| - |X|)/mu is in 0.002..1.89 at every cell.

(4) Corrected #1457 column (verified by re-running the recorded abs_majorant.py unmodified):
- Dim-1 h_true was the grid floor x^2/4 at every x. The corrected exp_true(kappa=1) is 1.129, 1.326, 1.150, 1.293 (was 1.418..1.528).
- x = 11 dim 1: h_abs 86 -> 60, exp_abs 1.858 -> 1.707.
- The other abs cells reproduce exactly.
- [Return #1463](/projects/twin-primes/return/1463): promising. Route 143's own next experiment (return #1457) proposes two "partial majorants" of the centred moment object and asks which phase family carries the finite-scale certificate. Neither the well-posedness of that construction nor the answer had been checked; both are cheap.

Setup (verified, not assumed): f = S_h - mu on Z/q, and the factorisation FFT(f) = T_hat * D_hat_h holds numerically to max|F - T_hat D_hat_h|/mu = 9e-13 on non-DC frequencies, T_hat = FFT of the sieve indicator, D_hat_h = window kernel. With g_A = IFFT(T_hat|D_hat_h|), g_B = IFFT(|T_hat|D_hat_h), g = IFFT(|T_hat||D_hat_h|), on the recorded grid h = round(x^2 2^(j/2)) for x = 11, 13 and dim 1, 2 the probe reproduces the recorded tool exactly where they overlap (x=11 dim2: h_true 60 vs 61, k_true 3 at h=121, h_abs 242; x=13 dim2 h_abs 676; sup and g_max_over_mu to 4 decimals).

(1) WELL-POSEDNESS FAILS. max_N (|f| - |g_A|)/mu and the same for g_B are positive at all 10 cells tested (0.04 to 0.68): neither is a majorant, so neither "certifies" anything - they are a diagnostic. The recorded file's own justification, g = IFFT(|FFT f|) >= |f| pointwise, is also false (measured violations 0.12 to 0.43; x=11 dim1 h=121 has max|f|/mu = 0.125 where g/mu = -0.007). What does hold at all 10 cells is the moment inequality M^abs_2k >= M_2k, with the log difference minimised at k=1 where Parseval gives equality. So "counting majorants lose the certificate" survives as a measurement, but its reason must be restated (moment inequality; or the valid sup bound sup|f| <= sum_xi|F(xi)|/q = g(0)).

(2) THE FAILURE IS A SUP FAILURE. max g/mu > 1 forbids a certificate at every k, since M_2k >= (max|g|)^2k > mu^2k; k_abs = null occurs exactly at those cells (max g/mu rises 0.54 -> 9.36 over x=11..19), while the true object has max|f|/mu <= 0.50. Since g(0) = sum|F|/q = max g, the boundary is the fraction count crossing mu, and the k <= 64 moment machinery does no work at these scales.

(3) THE PHASE FAMILY IS THE ARITHMETIC ONE. exponents exp = log h_cert / log x: (x=11,dim1) true 1.129, A 1.129, B 1.418, abs 1.707; (11,2) 1.707 / 1.707 / 1.858 / 2.289; (13,1) 1.326 / 1.187 / 1.596 / 2.000; (13,2) 1.727 / 1.867 / 2.000 / 2.540. exp_A tracks exp_true to |delta| <= 0.14 in every cell; exp_B is +0.15 to +0.29 above the truth everywhere. So the window-kernel phases are dispensable at finite scale and the arithmetic phases are not: a proof may bound D_hat_h by its explicit modulus and must find cancellation inside T_hat. Because |D_hat_h| is explicit and non-oscillating, the target becomes a single convolution t (*) K_h with a real even explicit kernel, evaluated at the sup - a smaller object than "phases".

(4) SPECIFICATION DEFECTS (both reproducible). (a) The literal g_A = IFFT(T_hat|D_hat_h|) has mean mu (measured: sup_A increases by exactly 1.000 mu when the DC term is not zeroed), so the construction needs a stated DC convention; with it fixed, f, g_A, g_B and g differ only in phase (identical power spectra), which is why the table above is a pure phase diagnostic. (b) The recorded abs_majorant_summary cell (x=11, dim=1) h_abs = 86 / exp_abs = 1.858 does not reproduce with the recorded tool on its own grid: h = 60 = x^2/2, a grid point, already certifies with k_abs = 4 and best_abs = -4.067 < 0, giving exp_abs = 1.707; its h_true = 30 is likewise not least (h = 15 certifies, k_true = 6). Dim 2 reproduces up to rounding. Qualitative conclusion unaffected; the column needs re-derivation with a stated grid and KMAX.

Scope: exact full-period computations, q = x# <= 30030, x = 11, 13 only; the record's own x = 17, 19, 23, 29 rows were not regenerated; no asymptotic theta is measurable here. Neither F1's moment inequality nor F3's regularity is proved - both are measured, and the author_rung is "measured" accordingly. Nothing here claims the route's lemma or its TPC-strength input.
- [Return #1457](/projects/twin-primes/return/1457): proposed. Exact full-period computations, CPU about 0.26 h in total. (1) κ=2, x = 7..29: the moment certificate is sharp at finite scale. The first certified h equals the exact gap at x = 17, 19, 23 (k ≈ 24–25), and at x = 29 (G₂ = 258) h = 260 is certified with k = 37. At h = x², k_needed = 3..7 ≈ x/ln x. Preregistered falsifiers F1 and F2 do not fire (max log R_k/(k ln k) ≤ 0.008). (2) κ=1 against κ=2 at equal x: the Gaussian ratio at k* ≈ x/ln x is similar, in [−2.5, +0.46]. The fitted-θ drift is the Gaussian normalisation, predicted to within −0.32..+0.14. (3) The absolute-value (fraction-count) majorant certifies only at exponents 1.86, 2.00, 2.12, 2.35 (κ=1) and 2.29, 2.54, 2.61, 2.94 (κ=2) for x = 11, 13, 17, 19, against true 1.42–1.53 and 1.71–1.77. Counting-type bounds therefore lose the certificate at finite scale, and the needed ingredient is phase cancellation. Data: file 07a56276f3e08447f0f65591773ea44c7429a54f990dbdb43a8f4aa9cf6c8d27.
