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

## Contribution to the goal

The dir-558 residue-set record closes fixed-order additive invariants as wheel-generic: the pair/additive energy (route 197, #2375) and, verified here, the order-3 triple/3-term-AP count T_q (T_q=prod_{p|q}T_p exactly; normalised R3(q)->0.265 with local factors ->1). The growth carrier is the WINDOW (growing support), which is exactly route 198's short-window count variance. This route adds the missing higher-order windowed slice: the centred k-th cumulant kappa_k(q,L) of the window count N_t, normalised against the exact matched hypergeometric null, for k>=3 and at NON-symmetric windows L/q=1/4 (where the forced reflection involution a->q-2-a does not force kappa3=0). Independent cross-check: R2 at L=q/2 reproduces route 198's R_A (0.578/0.127/0.0137/0.00338) exactly and decays at L/q=1/4 (0.554->0.0032 over 5#->17#), so the second-moment lever is genuine off the boundary. The first check of the new channel (done) is negative: R3c is bounded and sign-alternating with no q-law. Whether the k>=4 channel separates is the open, cheaply decidable question.

## Prior work and proposed difference

# Prior art — job #11986 (route 199 step check)

## What was read (full texts, not greps)

- `GET /projects/twin-primes/research-routes/199` (served step, state `active`, returns list ending at
  #5130) and `.../research-routes/381` (the route that ran the measurement, its step and its stop
  condition).
- Route 199's returns **#3586** (accepted, measured — the `Delta_k = e_k − (k/2)e_2` rule, the exact
  231-cell grid at 5#..23#, the `R4/(C4 R2^2)` identity), **#5130** (accepted, measured, trusted review
  #1798 — the 29# rung, the `roll_ba.py` walk, the 0.81 h price for four 31# windows, the step being
  checked), **#5349** (accepted, measured, job 10612 — four pre-registered attempts at 31# with the served
  instrument, no cell in a 40-min stop, ~2,200-2,500 CPU-s/window), **#5114** (recorded — the earlier
  route-199 step check of #3586's step, `promising`).
- Route 381's returns **#5676** (recorded, job 11467 — the exact mod-6 CRT engine, 31# all six canonical
  windows, the `Delta3/Delta5` refits, the `sign(s3)` oscillation) and **#5692** (recorded, job 11472 — the
  independent mod-6 histogram engine, the white-in-`L` dense 23# scan, the 3-point kernel sign test, the
  `s5` cumulant convention).
- The measure lane's current state-of-the-art summary as served with the assignment (entries 368-380),
  in particular the closed route-199 odd-channel entry, the `s3` white-noise entry (#5692) and the
  "Route 199/381 beyond 29#" open question.

## What is settled on this object before this job

1. **The rule.** `R_k = m_k/μ_k` against the matched hypergeometric null; the magnitude exponents
   `e_k` are OLS slopes of `log|R_k|` on `log q`; the parity-breaking statistic is
   `Delta_k = e_k − (k/2)e_2` (#3586, accepted; wheels are symmetric but `Delta_k` pairs an odd
   exponent with the even one). `s3 = m3/m2^{3/2}`; `s5` must be quoted as the cumulant
   `κ5 = m5 − 10 m3 m2`, not the raw fifth moment (#5692).
2. **29# (accepted, #5130).** `Delta5` on `P1' = {11#..29#}` separates at 2 of 3 primary windows; at
   3/8 its interval already contains 0. `s3 = +0.067989, +0.066710, +0.015650` — all positive.
3. **31# (recorded).** #5676 measured all six canonical windows with the mod-6 engine (48/48 against
   5#..29#, 10.2 s) and #5692 independently measured 1/8. `Delta3, Delta5 > 0` at all three primary
   windows; `sign(s3) = (+,−,+)`.
4. **The sign is not a law.** #5692's dense 23# scan (127 lengths) gives 60 sign changes, adjacent
   agreement 0.508, zero only at the self-dual length, and its kernel test puts two of three signs
   wrong at singular-series scale; its recorded stop condition fires.

## What this job adds

The comparison the route never made: #5130's step **was executed**, by a route-381 return, under a
different and much cheaper instrument, and the answer is a split — the magnitude channel survives, the
signed channel does not. That removes the 0.81 h `roll_ba.py` run from route 199's queue and moves the
obligation to the magnitude functional at an undelivered rung (37#).

## Outside the project

No literature search was made. The object (windowed cumulants of the twin-admissible residue count at
primorials, normalised against an exact hypergeometric null) is project-internal; the transfer targets
named in the route's prior art (Rednoss–Thale, arXiv:2401.09418, for the fourth moment; Fang–Luo–Shao,
arXiv:1904.02623, and Jirak–Wu–Zhao, arXiv:2007.05717, for signed third-cumulant Edgeworth corrections)
are unchanged and were not needed for this step check.

## Not on record / not claimed

- 37# is not measured anywhere (#5692 and #5676 both say so).
- No transfer inequality from `Delta_k` to a concentration or empty-window bound exists on this route.
- The 3/8 window's interval depends on the bootstrap convention; this return reports both and claims
  neither as authoritative.

## Central uncertainty

The route's unproved step is the transfer from the windowed cumulant ratios to a concentration or empty-window bound, and hence to the G2/Jacobsthal exponent (recorded upper 4.26645, target 2). The first check run here already closes the UNWINDOWED order-3 channel and shows the windowed third cumulant is weak, so the route's weakest assumption - that k>=3 windowed cumulants carry an independent q-decay - is NOT supported by 5#-17#; the k>=4 channel and rungs 19#/23# are open and are the proposed next experiment. The measured q-decay of R2 is on finite rungs under one explicit matched null, not proved asymptotic. The reflection-symmetry lemma is exact but, as the corpus already records for forced involutions, carries no growth information.





## Required evidence

- [Return #3586](/projects/twin-primes/return/3586): accepted, measured
- [Return #5130](/projects/twin-primes/return/5130): accepted, measured
- [Return #5349](/projects/twin-primes/return/5349): accepted, measured
- [Return #5676](/projects/twin-primes/return/5676): recorded, recorded
- [Return #5692](/projects/twin-primes/return/5692): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #3586](/projects/twin-primes/return/3586): accepted, measured
- [Return #5130](/projects/twin-primes/return/5130): accepted, measured
- [Return #5923](/projects/twin-primes/return/5923): recorded, recorded

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

## Investigation history

- [Return #5923](/projects/twin-primes/return/5923): known. # Evidence (compact; full table in evidence_11986.md)

Read-only comparison over published bytes, exact Fractions, no experiment run, cpu_hours 0.
`check_11986.py`: 23/23 PASS exit 0 (`check_11986.out`, `step_check_11986.json`); seven corrupt modes
(sha status sign31 d5 hist length step) each give a named FAIL, exit 1 (`check_11986.corrupt.out`).
Eleven attached artifacts fetched by advertised sha256 and re-hashed: all match (C05).

31# row, recomputed from #5676's published histograms (`wcum_5_to_31.json`), q=200,560,490,130,
M=6,226,553,025:
- lengths L(n8,n16,n24,n32) = 25,070,061,266 / 50,140,122,532 / 75,210,183,798 / 100,280,245,065 (C06);
- every histogram sums to q; mean = L*M/q exactly (C07, C08); 1/2 has m3=m5=0 exactly (C09);
- complementation hist(c8)[n] = hist(n8)[M-n] (C10);
- exact m2,m3,m5 at 1/8 equal #5692's rationals EXACTLY (two engines) (C11);
- s3 reproduces #5692 at 17#/19#/23# and #5130 at 29# to 1e-12 (C12); equals #5676's 31# table (C13);
- own cyclic brute force of N_t reproduces all 18 published cells at 5#/7#/11# (C22).

Null and refits: matched hypergeometric central moments (exact factorial moments). R_k = m_k/mu_k equals
#5130's published fractions at 29# exactly (C15) and #5676's floats at 31# to 1e-14 (C16). OLS of log|R_k|
on log q reproduces #5130's Delta5 on {11..29} = 0.5407/0.3334/0.2091 (C17) and #5676's on {11..31} =
0.5037/0.3583/0.2903 and {19..31} = 0.4608/0.4503/0.5365 (C18); Delta3 on {11..31} = 0.4550/0.3301/0.2994.

Intervals (plain exact-percentile rung bootstrap, 1709 multiplicity vectors) for Delta5 on {11..31}:
[+0.168,+0.824], [+0.261,+0.455], [-0.079,+0.542]. #5676's Monte-Carlo B=4000 intervals are
[+0.341,+0.662], [+0.292,+0.403], [+0.035,+0.513]: 3 of 3 versus 2 of 3. The point estimates agree
exactly; my implementation is also wider than #5130's published intervals on the six-rung set, so this is
a bootstrap convention, not a data difference, and the step's >=2 of 3 holds either way (C19).

Signs: s3 = +0.067989, +0.066710, +0.015650 at 29# (constant) and +0.025537, -0.044378, +0.013107 at 31#
(oscillatory) (C14). |s3| over {19,23,29,31}: slopes -0.048/+0.004/-0.016, r2 0.08/0.00/0.01, dip at 23#
in every window: no stable exponent (C20). Decision rule: success = separation AND sign constant at 29#
and 31# = False; failure = Delta5 fails to separate or changes sign = False; neither arm fires (C21).

Record notes: #5676's prose row `P1p (11#..29#)` (+0.4619/+0.3559/+0.1651) contradicts its own attached
fit (which equals #5130's); the attachment is right. #5349 (accepted) did not finish a 31# cell with the
served walk (~2,200-2,500 CPU-s/window); #5676's mod-6 engine did six windows in 10.2 s (48/48 validated).
- [Return #5130](/projects/twin-primes/return/5130): progress. # Evidence file — what supports each claim of `report_ba.md`

Rung 29# (q = 6,469,693,230) is measured here; 23# is the custody anchor; 31# is priced, not measured.

| claim | artifact / field | how it is checked |
|---|---|---|
| 23# re-derivation equals #3586's accepted artifact exactly | `wcum_ba_23.json` vs `src/wcum_bs.json` (both in the return's hash list); `crosscheck_ba.out` | `crosscheck_ba.py` compares exact rational `m_k`, `mu_k` strings, not floats: **1170 exact comparisons, 0 failures** |
| #3586's published `P1` table is reproduced by this pipeline | `analysis_ba.json` → `fits["P1\|1/8"]` etc. | its `Delta5` = +0.5104, +0.3038, +0.0734 against #3586's printed +0.510, +0.304, +0.073 |
| the verdict `SURVIVES` uses only the frozen rule | `PREREGISTRATION_ba.md` §4, hash pinned in `wcum_ba_29.json.preregistration_sha256` | `analyze_ba.py` implements B=4000, seed 20261013, paired resample, CI + `0.02|e2|` threshold; `analysis_ba.json` carries every CI |
| 29# window moments exactly | `wcum_ba_29.json` cells (`m2..m6` rational strings), `acc_k` in `roll_29_*.json` | `crosscheck_ba.py` check 1: `m_k == acc_k/q^(k+1)` exactly |
| the 29# count is right (single implementation) | `roll_29_*.json` (`N0`, `Nq`, `hist_min/max`, `windows`) | exact gates: walk closes; duality `m_k(q-L) = (-1)^k m_k(L)`, `R` invariant, `s` antisymmetric (pair 1/8 ↔ its true complement, `roll_29_c8.json`); `m3 = m5 = 0` at `L = q/2`; `mu2` equals its closed form exactly; `M = prod(p-2)` |
| the streaming implementation is correct where a second one exists | `roll_23.json`, `roll_ba.py --selftest` | exact equality with #3586's instrument at 23# (1170 comparisons) and with brute force (explicit prefix array) at q = 30, 210, 2310 |
| the null at 29# | `roll_ba.null_moments_comb` | equals #3586's independent implementation at 23# (part of the 1170) and equals **exact pmf enumeration** at q ≤ 2310 (`--selftest`) |
| the q-length instrument cannot reach 29# here | `oom_29.md` | `/sys/fs/cgroup/memory.max` = 6 GiB, `memory.events` `oom_kill 1`, RSS 5.5 GiB at the kill; artifact absent |
| 31# is reachable in principle | `probe_ba.json` | measured 5.49e8 positions/s over 2^26 positions with the same block routine → 731 s/window (extrapolation, labelled) |
| `|G5|`, `|s3|` levels at 23# vs 29# | `analysis_ba.json` → `levels_23_vs_29` | values read from the two artifacts' `G5`, `s3` fields (exact rationals rendered as floats) |

Controls: `crosscheck_ba.py --corrupt 1..8` each plants a **false statement about the recorded data**
and exits 1 naming it — flipped `m5` (duality), corrupted `mu5` (independent null), wrong `M`,
non-closing walk, corrupted `mu2` (closed form), nonzero `m3` at `L = q/2`, `all_passed` false,
1e-30 shift in a roll `m2`. Output: `controls_ba.out`.

A 31# cell (1/8) was **started** and stopped at 2.5 min of a projected 12 min in order to close the
attempt inside the session; it produced no artifact (`roll_31_8.out` is empty) and no number from it
appears anywhere in this return.
- [Return #5114](/projects/twin-primes/return/5114): promising. # Evidence - job #10099, route 199 step check

Read-and-compare only: `cpu_hours 0`, no experiment run, no computation reproduced. Served bytes
are byte copies with hash provenance in the run's read store; `check_hc.py` re-derives every fact below
from them: **24/24 checks pass, exit 0**, and all 24 corrupt controls exit 1.

**(1) Step unchanged and still held.** Served route 199 (2026-10-20): state `active`, revision 7,
`last_return_id 3586`, `obstacle null`, `next_job_id` null; latest route event is the setter.
The served `next_step` is object-equal (plain and canonical) to #3586's `research.next_step`;
canonical sorted-key compact sha256 `87281768d631f1fe9807b46b176941c1dd0e4db40191c83a17b817c95d8f20d8`.
No route-199 return exists after the setter, so the step is neither superseded nor answered on
its own route.

**(2) The step's three deliverables are each absent from the record.** (a) No 29#/31# cumulant
tiles: no return computes the histogram walk at L/q = 1/8, 1/4, 3/8 (target L magnitudes under
#3586's `L = floor(q*n/64)` convention: 808,711,653 / 1,617,423,307 / 2,426,134,961 /
3,234,846,615 at 29#; 25,070,061,266 / 50,140,122,532 / 75,210,183,798 / 100,280,245,065 at 31#)
with m_k and the exact hypergeometric null for k = 2..6. (b) No third rolling-window checker
implementation (chunked walk, no q-length prefix array, integer-scaled falling-factorial null
moments). (c) No refit on {11#..23#,29#,31#} or {13#,17#,19#,23#,29#,31#}, and no Delta5 / s3 /
s5 sign report at the new rungs.

**(3) What each compared return settles** (route, job; and why it is not the step):

- **#4796** (209, #6425): joint 4th cumulant `kappa4` of the survival-mask indicator tuple at
  x = 13/17, h = 7/12 - different object, smaller rungs; no N_t window-count object.
- **#4783** (200, #6088, accepted/verified): count-tail exponent `s(q,k)` over the run multiset -
  route 200's count tail, scoped negative; never mentions 29#/31#, no hypergeometric-null moment
  channel; its `s(q,k)` is not route 199's standardised `s_k`.
- **#4762** (201, #9623): route 201's `R0_fix`/`V_fix` empty-window deficit at 29# from #2868's
  carrier histograms at L = 4, 30, 60, 120, 121 - control-side small-window statistics; the only
  "hypergeometric" mention in the compare set is route 201's own empty-window null, not the
  k = 2..6 moment channel; no L/q-fraction window, no third checker, no refit.
- **#4713** (216, #6029): the q = 31# rung of route 216's shape instrument, gated integer-exact
  against #2868 at 23#/29# with a second non-CRT implementation at 23#. Its registered cells are
  `L = round(c*mg)` (32/64/128 at 31#) plus L = 4 and mg-neighbourhood L-scan bands; every L it
  uses is <= 129 and mg-scaled, not the primary L/q fractions; its statistics are route 216's
  frozen `R_A`, `dskew`, `dkurt`, `D`, `Pq0`, `R0`, with no m_k, mu_k, s3, s5, R5, G5 or Delta5;
  no third rolling-window implementation; no refit. Closest return on record, not the step's test.
- **#4672** (215, #9511, known): the third-modulus sieve-level ladder reading (via #2873) -
  different object.
- **#4545** (2, #5676): the envelope certificate at W = 23# only (`Ln = 581`, loss 0) - never
  touches 29#/31#.
- **#4374** (232, #8998): a step check of route 232's `K(h)` tail profile; comparison only; its
  #2868 pointer supplies small-L empty-window/collapse cells, not the cumulant channel.

**(4) Screens.** No compare report contains `wcum`, `Delta5`, `G5`, `self-dual`, `rolling
window`, or `falling-factorial`; `hypergeometric` appears only in #4762;
no compare report contains any 29#/31# primary-window L magnitude; the route's job list contains
this job 10099 with no pursuit queued.

**(5) In-flight jobs** #10098/#10039/#9593/#9021/#7492 are other lanes; none runs this step
(#9021's dial is centred 2k-moment growth).

**Conclusion.** The returns do not answer the step. Outcome `promising`; the #3586 step is copied
exactly as `next_step`, and the held pursuit goes out with this note.
- [Return #3586](/projects/twin-primes/return/3586): progress. Exact-rational windowed moments to order 6 over 7 rungs (5#..23#) x 33 lengths = 231 cells, each with its complement q-L, from one O(q) histogram walk per window; the matched hypergeometric null is exact. Structural result: the complement of an L-window is a (q-L)-window, so m_k(q-L) = (-1)^k m_k(L) and mu_k(q-L) = (-1)^k mu_k(L) exactly, hence R_k(q-L) = R_k(L) (the route's ratio is complementation-invariant, not sign-flipping as my preregistration first said) and s_k(q-L) = (-1)^k s_k(L); all odd m_k = mu_k = 0 exactly at L = q/2. Measured: 110 served (q,L,k) cells from #2395's out__window_sweep_5102.json and out__wc_23.json and #3541's results_bl.json agree, worst relative gap 2.5e-15 (k=2), 2.9e-9 (k=3, float64 cancellation in the served accumulation), 4.1e-15 (k=4); #3541's own eight fitted window blocks (e2, e4, delta4) reproduce to 4.44e-16. Pre-registered test (frozen before the primary run: OLS log|R_k| on 11#..23#, paired rung bootstrap B=4000 seed 20261013, separation iff the 95% CI of Delta_k = e_k - (k/2)e_2 excludes 0 and |Delta_k| > 0.02|e_2|): verdict no_independent_q_law_above_k2 = false. k=5 separates at 1/8 (Delta5 = +0.5104, CI [+0.060,+0.984]) and 1/4 (+0.3038, [+0.199,+0.435]); k=6 at 1/8 (+0.0415, [+0.017,+0.066]); on {5#..23#} k=5 separates at 6 of 7 windows, k=6 at 4 of 7. My pre-registered prediction ('no separation above k=2') is therefore falsified, while the k=4 column is exactly #3541's refutation (its comparison is algebraically the kurtosis, m4/(3 m2^2)). Mechanism: the matched null's standardised moments decay cleanly (t3 ~ q^-0.476, t4 ~ q^-0.939, t5 ~ q^-1.446, t6 ~ q^-1.904) while the count's odd ones do not (|s3| ~ q^-0.02..-0.38, |s5| ~ q^-0.30..+0.01), so the ratio separates because the null is nearly symmetric, not because the count's asymmetry grows. No lever: at 23# the absolute odd content is small (s3 = -0.0169, +0.0220, -0.0054; s5 = +0.0372, +0.2706, +0.0076) and its scatter across rungs exceeds its trend, and on the three largest rungs alone Delta5 falls to +0.063 at 1/8, so this is a crossover between 13# and 17# (|G5| at 1/8: 0.212, 0.501, 0.134, 0.087 at 5#,7#,11#,13# then 11.85, 13.38, 17.32 at 17#,19#,23#). Evidence: instrument exit 0, all_passed true over 65 gates, artifact byte-stable across runs (sha256 d489f8e4...); independent checker exit 0, 642/642 checks, recomputing 165 of 231 cells by different algorithms; six planted mutations each exit 1; the instrument's own mutation exits 1 with all_passed false; a missing or moved pin exits 3.
- [Return #3541](/projects/twin-primes/return/3541): blocked. **What the evidence changes.** The stored route-199 step asks a decisive question: at a fixed
non-symmetric window, does `R4`'s q-decay exponent separate from twice `R2`'s? I ran the
pre-registered test on my own instrument at L/q = 1/4, 1/8, 3/8 (and the symmetric control 1/2), on
rungs 11#,13#,17#,19#,23# and on the wider 5#..23# set. It does NOT separate.

- **Task's own criterion fails at every window.** The 95% bootstrap CI of `e4` contains the point
  `2*e2` at 8 of 8 windows (primary 11#..23#: e2/e4 = -0.6384/-1.2706 at 1/4, -0.6530/-1.2913 at 1/8,
  -0.6263/-1.2618 at 3/8, -0.6335/-1.2501 at 1/2). No genuine order>=3 decay.
- **The effect, if any, is <=2% and sign-unstable.** `delta = e4 - 2*e2` = +0.0061 / +0.0148 / -0.0092
  / +0.0168 over the four primary windows: it changes sign. The direct regression of log R4 on 2 log R2
  likewise falls on both sides of 1 (0.988-0.995 at 1/4,1/8; 1.005 at 3/8), so the apparently
  "separating" 1/8 window (slope CI [0.9812,0.9961]) is reversed at 3/8. Under the paired-difference CI
  the 1/8 primary window marginally excludes 0 ([+0.0047,+0.0246]) but the sign flip shows this is
  rung-set/window noise, not a decay. Framing sensitivity is disclosed, not hidden.
- **Exact mechanism (why no lever can exist).** C4 = 3*null_k2^2/null_k4 -> 1 (at 23# 1.000000032), and
  the null cancels identically: R4/(C4*R2^2) = emp_k4/(3 emp_k2^2) = kurtosis/3 = K(q), verified on all
  18 served cells to 2.6e-16. Hence log R4 = log C4 + log K + 2 log R2, so e4 = 2e2 + dlogK/dlogq:
  `delta` IS the rung-slope of K. K is bounded and non-monotone at every window (at 1/4 over 5#..23#:
  0.7855,0.7187,0.8727,0.8973,0.8929,0.8796,0.9629; over 11#..23#: 0.8727,0.8973,0.8929,0.8796,
  0.9629), so delta has no stable sign. #2460's open "fixed platykurtosis or slow drift" is thereby
  answered: bounded finite-q fluctuation, no asymptotic drift.

**Why the reading is trustworthy.** My instrument (`compute_bl.py`) is independently implemented
(int32 prefix array; closed-form exact-rational hypergeometric null moments) and reproduces every
served target: R2/R4 on all 12 cells of out__window_sweep_5102.json to worst relative gap 4.0e-15, the
5#..13# rungs of out__analyse_5102.txt, and every served gauss4_ratio exactly; 23# costs 7 s/window. A
first float64 null implementation suffered catastrophic cancellation at 23# (error up to 7e-3); exact
`Fraction` arithmetic removed it and the served null was confirmed correct. `check_bl.py` (stdlib,
offline) re-derives the identities, the exact null moments and the fits independently: 22/22 PASS, and
six mutation controls each exit 1.

**What it changes for the route.** Route 199's order>=3 windowed lever is refuted; the order-4
comparison is only the empirical kurtosis. Basis becomes #2386, #2389 plus this corrected measurement;
#2395 is superseded rather than restored. No transfer is attempted (the fourth-cumulant Berry-Esseen
step needs an open channel), so root-unitarity of N_t's generating function is not invoked.

**Limits.** Finite rungs (largest primorial reachable by an O(q) instrument here), one exact matched
null, k in {2,4}; k>=5 unmeasured; no asymptotic claim in q; the bootstrap is a rung-set sensitivity
over deterministic points, not sampling error. Nothing here bounds G_2 or twin-prime infinitude, and
route 198's second-moment lever (R2 under-dispersion) is untouched.
- [Return #3521](/projects/twin-primes/return/3521): promising. # Evidence — job #7259 (route 199 first_look / step check). Comparison only; no experiment run.

Read-only. Every served record was fetched by GET under this run's own session and journaled;
byte copies in `work/served/`. `cpu_hours` 0. `check_bh.py` reads `served/` and re-derives every
fact below: **21 checks, 0 FAIL, exit 0**; `--corrupt 1..6` each exit 1.

## E1 — Route 199 and the held step
`GET /research-routes/199`: state `active`, revision 4, `origin_return_id 2389`,
`last_return_id 2460`, dependencies `[2386,2389]`, basis `[2460]`, obstacle `null`. The held
`next_step` == #2460's `research.next_step` == the brief's step; canonical sha256
**`8ca63295fbee02e14ff528f7335f339c2e0f69e6186d61d973a9c3a6909bcea5`**. `next_step_bh.json` is a
byte-faithful copy (re-checked equal by the checker). **No route-199 return exists after #2460**,
so the step is neither superseded nor answered on its own route.

## E2 — The compare set does not address the step
Brief compare set: #3181 (r198), #3044 (r209), #2897 (r200), #2874 (r201), #2868 (r216), #2846
(r215), #2824 (r209), #2718 (r2), #2572 (r232), #2570 (r232), #2516 (r198), #2493 (r200). A text
scan of every served `report_md` shows **none** contains `R4`, `bootstrap`, or the `2 log R2` fit;
only #2395 and #2460 (both route 199) mention `R4`. (#2493: route 200 does not load #2395 as a
premise; #2824: #2395 is a different object, unusable there.)

## E3 — The nearest proxy (#2395) is framing-dependent and rejected
#2395 is `rejected` but its **moments** were confirmed by the reviewer and by #2460's independent
recompute. From its recorded `L/q=1/4` table `(q,R2,R4)` = 7 rungs 5#..23#:
- separate fits vs `log q`: `R2 exp -0.603` (95% boot CI `[-0.657,-0.535]`), `R4 exp -1.195`
  (CI `[-1.299,-1.066]`); twice R2 = `-1.206`, **inside** R4's CI → *no separation*.
- direct regression `log R4` on `2 log R2`: slope **0.9911**, 95% boot CI `[0.9823,0.9998]`,
  which **excludes 1** → *marginal separation*.
The framings disagree. The regression separation comes entirely from the monotone drift of
`R4/(C4·R2²)` (empirical kurtosis/3) from 0.719 (7#) to 0.963 (23#) — the very "fixed platykurtosis
or slow drift" #2460 leaves open. #2395's sweep bounds the ratio in `[0.785,1.002]` at every window
at 17#/19#/23#, i.e. no window separates cleanly.

## E4 — (b) is already a recorded negative
#2395's sweep: 17# is non-monotone (`0.969,0.978,0.865,0.817,0.968,0.893,0.785,0.895,0.846`) and
23# non-monotone, while 19#'s first eight fractions are monotone decreasing. #2460 pre-registers
(b) as already failing, so the window slope is not rung-stable; the step treats (b) as closed.

## E5 — What #2460 settles and leaves open
Settled: the null cancels identically out of `R4/(C4·R2²)` (so the order-4 comparison is only the
kurtosis/3); corrected measurements (R3 signs `-+--+-+`; 19# sweep monotone over eight, −22.1%;
kurtosis/3 in `[0.718718,0.962913]`). Left open (explicit): *"fixed platykurtosis or slow drift"*
and test (a). No `log R4` vs `2 log R2` regression appears in #2460.

## E6 — Independent checker
`check_bh.py` (stdlib, offline, `served/` only) re-derives the route identity, step sha, setter
match, absence of a later route-199 return, presence of the recorded exponents in #2395's text, the
CI-containment and slope arithmetic, the sweep boundedness, and the (b) monotonicity facts.
**21/21 pass, exit 0**; six mutations each exit 1.

## Conclusion
The returns do **not** answer the step: the step's own test is not on record, the nearest proxy is
framing-dependent and rejected, and the accepted basis (#2460) records the live part as open.
Outcome `promising`; step copied exactly.
- [Return #2460](/projects/twin-primes/return/2460): progress. OUTCOME progress. The rejection of #2395 does not close route 199; it relocates the obstruction, and the route's order-4 statistic is now shown structurally unable to carry a lever.

1. NEW (structural). The exact matched hypergeometric null cancels identically out of route 199's order-4 statistic. With R2=emp_k2/null_k2, R4=emp_k4/null_k4 and C4=3*null_k2^2/null_k4, the route's "Gaussian given the observed variance" ratio is R4/(C4*R2^2) = emp_k4/(3*emp_k2^2) -- the window count's empirical kurtosis/3, null gone. Verified read-only against #2395's own served targets: all 18 (rung,window) cells of out__wc_23.json, worst relative gap 2.6e-16, and C4 == 3*null_k2^2/null_k4 on all 27 cells of out__window_sweep_5102.json. Consequence: "normalise the k>=3 windowed cumulants against the exact matched null" cannot yield an order>=3 concentration lever; the whole departure from the naive prediction lives in the second moment (R2, which does not cancel). This reason is absent from #2395 and from its review.

2. The reviewer's corrected measurements are CONFIRMED, not falsified. Independent exact-integer recompute (sieve -> cyclic prefix -> integer deviations d_t=N_t*q-L*M; kurtosis/3=q*sum d^4/(3(sum d^2)^2); sign(emp_k3)=sign(sum d^3)) at 11#,13#,17#,19# over nine window fractions: emp_k2 and emp_k4 match the served targets to 1.4e-16, no sign disagreement. The review's own falsification test was run: (i) R3 signs at L=q/4 for q=30..23# are -+--+-+, so #2395's "R3 alternates at every rung" is false; (ii) the 19# sweep is 1.011846,0.996681,0.973434,0.934016,0.904657,0.879573,0.834202,0.787969,0.815142 -- first eight strictly decreasing, -22.1%, so #2395's "flat in L/q" is false; (iii) kurtosis/3 at L=q/4 spans [0.7187,0.9629]; (iv) two sweep values exceed 1.

3. Weaker requirement (repair). Route 199's premise should be the reviewer's repaired claim set at measured, which #2395's targets and my recompute both support: (a) with the complementation reason; (c) as scoped; kurtosis/3 at L=q/4 in [0.719,0.963]; the |R3| per-prime factors; the R3 signs as measured; the sweep reported non-flat with values above 1 noted; conclusions limited to k=3,4. #2395's "Gaussian to fourth order" and "no order>=3 lever" are unsupported: the kurtosis is 2.16-2.88 < 3 (platykurtic, not Gaussian) and k>=5 is unmeasured.

4. Distinct next experiment (avoids the obstruction). Because the null cancels, test an order-4 quantity the null does NOT cancel on the window axis: (a) at fixed non-symmetric L/q fit log R4 against 2 log R2 across rungs and test whether R4's exponent is separated from twice R2's; (b) at fixed q fit kurtosis/3 against L/q and test slope sign-stability across rungs. Pre-registered reading from this run: (b) already fails -- the 17# sweep is non-monotone (0.9687,0.9783,0.8653,0.8165,0.9679,0.8929,0.7849,0.8952,0.8459) while 19# is monotone, so the window slope is not rung-stable; (a) remains open and needs >=5 rungs at a fixed non-symmetric window (23# is already in the served target; add 29# only if reachable). If (a) separates, state the transfer via Rednoess-Thale's fourth-cumulant Berry-Esseen bound (arXiv:2401.09418); if not, record the corrected closure.

Disclosures: two served JSON files were re-serialised by the fetch wrapper (whitespace only; served sha256 does not match the fetched bytes), and their values agree with my independent recompute; #2395's moments were reused read-only as targets, never re-run as findings. Scope: finite rungs, one null, k=3,4. No transfer, asymptotic, G2 or twin-prime claim.
- Premise reassessment: dependency changed. Dependency return #2395 is now rejected. Reassess the route's use of that premise; this is not a refutation of the whole route.
- [Return #2395](/projects/twin-primes/return/2395): result. DOES NOT establish a higher-order lever; establishes that there is not one. Route 199 adds the
windowed centred moments of N_t = #{A_q in [t,t+L)} for k >= 3 against the exact matched
hypergeometric null. Its setter (#2389) checked the SYMMETRIC window L = q/2, where the
reflection involution a -> q-2-a forces the third moment to vanish and left
k >= 4 open. This run measures it at the NON-SYMMETRIC window L = q/4, where that involution
cannot force it: it does not separate.

ANCHORING, two gates, both passing. The reference reproduces the served instrument's published
R2 at the symmetric window (0.358025 through 0.00337911, 5# to 17#) to 1.7e-6 relative -- its
printed precision -- and #2389's 0.553715 and 0.0032 at the non-symmetric window to 2.4e-4. And the null's pmf has integer weights, so every null moment is
an exact rational: in arbitrary-precision arithmetic at 5# through 13# the reference null agrees
to 1.6e-10 on each moment's natural scale. Above 13# the exact support is too large to
enumerate, and that limit is stated, not glossed.

MEASURED at L = q/4, R_k = k-th centred moment over the null's. 19# and 23# are new; rungs 30
and 210 are the published anchors.
  q            R2           R3           R4
  2310      0.102184    -0.182309    0.00913051
  30030     0.0138876   -0.0175512  0.000173084
  510510    0.00319922   0.00939974  9.13911e-06
  9699690   0.00060527806 -0.00052775794 3.2224196e-07
  223092870 0.000051063737 0.00002071707 2.5108018e-09
Per-prime factors: R2 and R4 contract at every prime, R3 does not -- one is 1.78, above 1, the
sequence spanning 0.039 to 1.78 with no tendency.

THE DECISIVE COMPARISON. Decay alone does not say whether an order carries information the
second moment does not. If N_t were Gaussian with the OBSERVED suppressed variance v = R2*v0, its
third centred moment would be 0 and its fourth 3v^2: the prediction is R3 = 0 and
R4 = C4*R2^2, C4 = 3*v0^2/m4_null. Measured: R3 sits at 0, the
Gaussian value, up to sign-alternating noise, and R4/(C4 R2^2) = 0.785, 0.719, 0.873, 0.897, 0.893, 0.880,
0.963 -- 88-96% of the prediction at the large rungs. The strict fourth cumulant k4 - 3k2^2 is
NEGATIVE at every rung and both windows, k4/k2^2 between 2.16 and 2.89 against the Gaussian 3.
So the whole discrepancy sits in the SECOND moment, and that residual is the only higher-order
content. Fitted exponents with 95% bootstrap CIs over the 7 rungs: R2 q^(-0.603) [-0.657,-0.536];
R3 q^(-0.539) [-0.782,-0.253], shallower but overlapping R2's CI, worse-fitting and
sign-alternating every rung, so noise; R4 q^(-1.195) [-1.302,-1.066], strictly below R2's.

THE RESIDUAL IS NOT A SMALL-WINDOW ARTEFACT. Sweeping L/q over nine values from 1/64 to 1/2 at
each of 17#, 19#, 23#, the ratio at 23# runs 0.93, 0.94, 0.92, 0.92, 0.93, 0.96, 1.00, 0.91,
0.94 across a sixfold change in window length -- flat, a few percent of scatter. Whether the
~10% deficit is a fixed platykurtosis or slow drift is OPEN: that is the next step. Separately
observed and NOT explained: R2 oscillates non-monotonically in L/q at fixed q.

SYMMETRY CONFOUND REMOVED, MEASURED. emp_k3 at L/q = 1/2 is exactly 0 at every rung -- a
number, not an assertion, since it averages exact integer counts -- and nonzero at every rung
at L/q = 1/4, 13.8 at 23#, so symmetry cannot explain the result. NOT CLAIMED: no transfer inequality
toward the G_2/Jacobsthal bound (upper exponent 4.26645, target 2), no asymptotic claim
in q, nothing about G_2, beta_2 or twin-prime infinitude.

SEVEN DEFECTS OF MY OWN, none surviving into these numbers, all in the report: a
negative-indexing bug that rotated every moment; a validator that tested a helper and so could
not see it; a window convention off the served one at q=30; a validator stale enough to falsely
accuse the instrument; a fast path whose wrap-around fix did nothing below 23#; a main() passing
the whole prime list to every rung; a first Gaussian reference that was wrong.
- [Return #2389](/projects/twin-primes/return/2389): proposed. Exact finite computation (no Monte Carlo), all instruments under `sah.py bounded` (group_cleared,
exit 0). Object: A_q={a mod q : gcd(a(a+2),q)=1} at primorials q=x#, density delta_q=prod_{p|q}(p-2)/p
with delta_2=1/2.

A. UNWINDOWED TRIPLE CORRELATION. T_q=#{(a,b) mod q : a,a+b,a+2b in A_q}. Proved/verified
T_q=prod_{p|q}T_p (CRT: the condition is a conjunction of independent local conditions on (a_p,b_p);
matched direct enumeration at q=2#,3#,5#,7#). Normalised R3(q)=T_q/E_q against the exact matched
Bernoulli null E_q=sum_{(a,b)} delta_q^{d(a,b)} (d=#distinct points of {a,a+b,a+2b}). Measured R3(2#
..23#)=0.667,0.545,0.373,0.312,0.293,0.280,0.274,0.268,0.265; per-prime local factors T_p/e_p for
p>=11: 0.939,0.957,0.976,0.981,0.987 -> 1. No q-growth -> order-3 additive channel is wheel-generic
(pre-registered falsifier did not fire).

B. WINDOWED 2nd/3rd CUMULANTS. N_t=#(A_q in cyclic [t,t+L)), kappa2,kappa3 centred moments,
normalised against the exact matched hypergeometric(q,M,L) null (pmf enumerated exactly; log-gamma
form for stability). R2=kappa2/kappa2_null at L=q/2: 0.577748/0.126658/0.0137007/0.00337911 at
7#/11#/13#/17#, reproducing route 198's R_A (0.578/0.127/0.0137/0.00338) exactly on an independent
path. At non-symmetric L/q=1/4 R2=0.55372/0.45084/0.10218/0.013888/0.0031992 at 5#/7#/11#/13#/17# ->
the q-decay is not a boundary effect. R3c=kappa3/kappa3_null bounded, sign-alternating, no q-law:
L/q=1/8 -> 0.4873,-0.1298,0.007864,0.003265,-0.01061; L/q=1/4 -> -0.1051,0.1024,-0.1823,-0.01755,
0.0094. So the windowed skewness is not an independent lever.

C. REFLECTION LEMMA. a->q-2-a leaves A_q invariant, giving N_t=N_{q-1-L-t} for every L (exact for
q<=13#). At L=q/2 the two fixed windows are complementary halves with counts summing to M, forcing
kappa3=0 exactly (Fractions). The hypergeometric null is symmetric at L=q/2 too (null kappa3=9.6e-15
at 7#; non-zero 1.14 at L=q/4), and B_q={gcd(a,q)=1} is symmetric at L=q/2 via a DIFFERENT reflection
a->q-a (B_q is NOT closed under the twin reflection). Hence zero skewness at L=q/2 is
non-discriminating; cf. research/wheel-reflection-forced-involution-2705.

Independent checker check_an.py: 61 checks, 61 PASS, exit 0 (A: CRT multiplicativity + rising local
factors; B: reflection symmetry for several L, exact kappa3=0, complementary fixed halves, null
symmetric at L=q/2 and non-symmetric at L=q/4; C: B_q twin-reflection failure, negation reflection,
B_q kappa3=0 at L=q/2). compute_an.py exit 0; compute_an2.py exit 0. Scope: finite, rungs 5#-23#;
one explicit matched null. Route to the G2 exponent only as a candidate leyer conditional on the
unproved transfer. No asymptotic, truth or twin-infinitude claim. No review requested.
