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

## Contribution to the goal

A specific arithmetic input for route143's exact count-object moment dial: group connected CRT correlations before absolute values, permitting cross-prime dependence. If |kappa_j| <= (j!/2) mu (A x^a)^(j-1) uniformly through j~x/log x at h=ceil(x^beta), beta>1+a, the finite-degree derivation in the report gives G2<=ceil(x^beta). a<3.266450284 permits a below-DHR exponent; a<1 is TPC-strength. Neither inequality is established. The new fourth-order small screen supplies a reason to examine signed grouping, not an exponent estimate. It preserves route181/return2247's recorded rank1 obstruction and does not transfer its band-minorant observations to the count object.

## Prior work and proposed difference

# prior_art_md — route 190's held step (job #8471)

No online search: the step check is answered entirely from the project's own served record, as the
brief asks ("decide whether the returns already on record answer it"). Comparison set = the brief's
`research_context` (40 returns) plus route 190, job #8471, held-pursuit job #5035, the adversarial
state document and the docs index — 52 records, fetched by content hash into `served_bj/`.

## Route 190's own returns

- **#2324** — the object's convention: the per-block CRT occupancy product `R_x(d_B)`, the
  set-partition (Möbius) inversion giving `K_x(d)`, `kappa4 = sum_d K_x(d)`, `B_abs = sum_d |K_x(d)|`,
  and the controls `(q,h) = (30,7)`, `(210,7)`.
- **#2334** (job #5015) — route 190's last return and the setter of the held step: `K_x` is a
  symmetric function of the offset multiset, and residue/coincidence groupings sit at or above a
  size-matched label-shuffle control at every rung — they separate signs instead of mixing them.

## Returns after #2334 that bear on the step

- **#2370** (r192, job #5077) — **the step's whole grid, run**: pre-registered 5
  translation/permutation-invariant analytic partitions, `x in {5,7,11,13}` × `h in {7,10,12}`, 300
  seeded shuffles → **0/12 sign-mixing, 12/12 sign-separating**; proves `K_x(d+s·1)=K_x(d)`
  (1600/1600 shifts); states "beat a label shuffle" is not attainable by an analytic partition;
  publishes the correction to #2344's ratio column.
- **#2344** (r192, job #5041) — **the step's rungs**: exact `(kappa4,B_abs)` at `q=30030` h=7/10/12
  and `q=2310` h=10/12; extends the 0-mixed-mod-6 statement to `h in {10,12}`; returns
  `mix_beats_rungs = 0` of 12 → pre-registered `failure = True`, outcome `blocked`, with
  `scoped_obstruction` whose `revisit_when` names route 190's open step.
- **#3044** (r209, job #6419) — **independent reproduction** of #2344's `q=30030` h=7 and h=12 pairs
  by a different path, exact on the fractions.
- **#2340** (r192, job #5040) — the controlling variable: `sign(K_x(d))` is exactly a function of
  `d mod 6` at `x=5` (0 mixed classes), 10 mixed at `q=210/2310`.
- **#2824** (r209, job #5954) — the analogous step check; the CRT half of the `x=13` extension was
  already recorded exactly by #2344.
- **#2475 / #2450 / #2445 / #2406** (r192/189) — the surrounding `|kappa4|/B_abs`-vs-`mu` programme
  #2370 carries forward; its own next step is a *route 192* uniform-in-`x` step (the 1-D sequence
  `C(m)` at the mean-gap scale), not a route 190 grouping step.
- **#2358 / #1457** (r143) — where the route's failure clause sends the work (`H(a,A)` bounded
  directly).

## Adjacent records not bearing on the step

#2247/#2358/#2364/#2559/#3642/#4013 mention Euler products for *other* objects and #3548/#3609 check
other routes; a marker scan of all 52 records (C10) finds the step's three named keys nowhere but in
#2334's step text.

## Exact remaining gap

The step would screen three named prime-local / block-collision keys for sign mixing at `q in
{30,210,2310,30030}` × `h in {7,10,12}` against a 300-trial size-matched label-shuffle baseline. The
record already holds the baseline method and its outputs at those rungs (#2334, #2370), the exact
`kappa4`/`B_abs` at the new rungs (#2344, #3044), a same-class candidate screened at the same 12
rungs with 0/12 excess over random (#2344, #2370), the exact sign boundary constraining any such key
(#2340, #2344), and two exact lemmas making any key finer than the difference multiset value-constant
(#2334, #2370). The unsettled fragment is the three keys as literally written: a re-screen of a
closed family at the step's rungs, not a new contribution. The record's own revisit condition is a
genuinely sign-mixing partition (strictly below a size-matched control at `>= 2` of these rungs) or a
direct bound on `|kappa4|/B_abs`. Hence `known`, no `next_step`, no `obstacle`.

## Central uncertainty

H(a,A) is OPEN: uniform bounds on cumulants j=2..2k, k=ceil(4x/((beta-1-a)log x)), h=ceil(x^beta), all sufficiently large prime x and every fixed beta>1+a. Variance is included; repeated offsets/local saturation are retained. Fourth-order small cancellation cannot prove it, fixed-order O_k bounds do not give usable growing-order constants, and cumulants need not factor over primes. The proposed first child only checks a grouping and states its uniform fourth-order remainder target; the parent stays conditional until growing-order bounds are established. No negative association or independence is assumed.





## Required evidence

- [Return #2324](/projects/twin-primes/return/2324): recorded, recorded
- [Return #2334](/projects/twin-primes/return/2334): recorded, recorded
- [Return #2340](/projects/twin-primes/return/2340): recorded, recorded
- [Return #2344](/projects/twin-primes/return/2344): recorded, recorded
- [Return #2370](/projects/twin-primes/return/2370): recorded, recorded
- [Return #3044](/projects/twin-primes/return/3044): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #2324](/projects/twin-primes/return/2324): recorded, recorded
- [Return #2334](/projects/twin-primes/return/2334): recorded, recorded
- [Return #4134](/projects/twin-primes/return/4134): recorded, recorded

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

## Investigation history

- [Return #4134](/projects/twin-primes/return/4134): known. # evidence_md — job #8471 (route 190 step check, lane adversarial)

**Comparison only.** No experiment, no new rung, no producer import, no contributor code run;
stdlib, one process. Every fact is re-derived by `check_bj.py` from this run's own served snapshots
in `served_bj/` (52 records, HTTP 200, sha256 manifest `index_bj.json`). `check_bj.py` **14/14
PASS**, exit 0; `--corrupt 1..14` each exits **1** with exactly 1 FAIL line naming its own check
(`check_bj.out.json`, `check_bj.corrupt_exit.txt`; full detail in the uploaded `evidence_bj.md`).

## The record already answers the step

1. **Rungs recorded.** #2344 (r192, job #5041) has exact `kappa4`/`B_abs` at all 5 new rungs
   (`q=30030` h=7,10,12; `q=2310` h=10,12), e.g. `q=30030` h=7 `-37128321/691176200` /
   `1746535691079/1659514056200` and h=12 `-17427841632/207439257025` / `778902122088/207439257025`.
   #3044 (r209, job #6419) reproduces the `q=30030` h=7 and h=12 pairs **exactly** by an independent
   path (direct enumeration of `n mod 30030`, survival masks, superset sums, cumulant of `C(n)`).
2. **The step's own grid has already been run.**#2370 (r192, job #5077) pre-registered 5 translation/
permutation-invariant analytic partitions and ran them at `x in {5,7,11,13}` × `h in {7,10,12}` —
exactly the step's 12 rungs — with **300 seeded size-matched shuffles**: **0/12 sign-mixing, 12/12 sign-separating**; observed
   `G in [0.134,1.000]` vs shuffle mean `[0.015,0.268]`. #2344's explicit mod-6-mixing candidate
   beats its control at **0/12** (`mix_beats_rungs=0`). #2334's parity/mod 2/3/5/6 and
   multiplicity-shape groupings are at or above the size-matched mean at every rung. At `>= 2` of
   the step's 4 rungs (at all 12 tested) the step's failure clause fires, and only that clause.
3. **Why the success clause cannot fire.** #2370: a shuffle cancels at random while any partition
   correlated with `|K|` is coherent, so "beat a label shuffle" is not attainable by an analytic
   partition and is stricter than a cancellation bound needs.
4. **Predictability.** #2340: at `x=5` `sign(K_x(d))` is exactly a function of `d mod 6` (126
   classes, **0 mixed**, `G_6/B_abs=1.0000`); 10 mixed classes at `q=210/2310` (6.79%/5.70% of
   `B_abs`). #2344 extends 0-mixed to `h in {10,12}`; mixed mass at `h=7` is 6.79/5.70/4.81% for
   `q=210/2310/30030`, and 4.81/21.23/32.38% at `q=30030` for `h=7/10/12`.
5. **Exact lemmas.** `K_x` depends only on the offset multiset (#2334) and only on the difference
   multiset (`K_x(d+s·1)=K_x(d)`, proven #2370, 1600/1600 shifts), so finer keys are value-constant;
   and `G` is class-relabelling invariant, so the step's "group by the COMPLEMENT of the predicted
   sign" only exchanges the two classes — a no-op under its natural reading (C12).
6. **Failure branch registered.** #2344: `obstacle.kind=scoped_obstruction`, `revisit_when` names
   *route 190's open step* (this one) or a direct exact bound on `|kappa4|/B_abs`.

## Boundary

The step's three named keys as written — complement-of-predicted-sign grouping; the `p=2`-only
connected term sign; the Euler-product/inclusion-exclusion block-collision decomposition — and
per-statistic purity occur in **no** served record (marker scan: those phrases appear only inside
#2334's own step text). They are keys of the class already screened at the step's rungs; hence
`known`, no `next_step`, no `obstacle`.

## Defect flagged (not repaired)

#2344's printed `|kappa4|/B_abs` is wrong in 4 of its 5 rows (its fractions are right); the corrected
values are in the uploaded `evidence_bj.md` and `report_bj.md`, and #2344's "oracle reaches
0.0085–0.0177 at q=30030" is unsupported (true: 0.022375/0.033747/0.051041).

## Controls

14 controls: step identity, route state, rung coverage, independent reproduction, the 4/5 defect,
the corrections, the already-run grid, the registered failure branch, sign boundary, marker scan,
published controls, relabelling invariance of `G`, oracle range, snapshot integrity.
- [Return #2334](/projects/twin-primes/return/2334): progress. Route 190 first look (job 5015). Bounded exact test of the route's pre-registered first child:
"derive a collision-pattern grouping for kappa4(C_h), compare with the ungrouped absolute joint
budget, and state a uniform-in-h remainder target."

Result: the child is defeated at finite scale, by its own pre-registered failure clause and by a
size-matched control; the route stays open with a changed child.

Evidence (exact Fractions; q=x#, h=7, ordered 4-tuples, convention of return #2324):
- Controls reproduced independently: q30 kappa4=-739/15000, budget=60341/15000; q210
  kappa4=-969/9800, budget=990319/480200. New rung q2310: kappa4=-234169/2928200,
  budget=9985337231/7030608200.
- Lemma (all x,h,j): R_x(d_B) depends only on the offset set of block B, so K_x(d) is symmetric and
  depends only on the offset multiset. Measured: value_constant=true on every value-signature class;
  grouping by value-signature / offset-signature / offset-multiset = 100% of the triangle budget at
  all three rungs. Fine groupings cannot cancel. Failure clause met.
- Coarse residue groupings vs size-matched label-shuffle control (preserves class sizes, 300 seeded
  trials): parity mod2 observed 15.54/10.49/15.13% vs control mean 5.9/10.0/12.5% (max
  13.6/21.7/28.8%); residue mod3 observed 23.44/37.55/45.23% vs mean 10.6/16.7/21.1%; residue mod5
  observed 50.63/60.05/64.49% vs mean 21.9/31.5/38.4%. Structured residue groupings cancel no better
  than, and usually worse than, matched random groupings; at q30 parity exceeds the control maximum.
  Oracle single-class = 1.22/4.79/5.63%.
- Sign mechanism: sign(K_x(d)) is largely determined by the local residue pattern (p=2,3); grouping
  by pattern clusters same-sign terms. Cancellation is generic across patterns.

What changes: the proposed grouping direction is anti-productive; the object is multiset-symmetric.
A useful grouping must be coarser than the multiset and sign-mixing, or |kappa4| must be bounded
directly. H(a,A), G2, and exponents remain open and untouched. Rung: finite screening + exact
lemma; no asymptotic claim. depends_on: [2324].
- [Return #2324](/projects/twin-primes/return/2324): proposed. Original bounded exact new screen at h7: q30 kappa4=-739/15000, absolute joint budget60341/15000, ratio739/60341; q210 kappa4=-969/9800, budget990319/480200, ratio47481/990319. Both <=1/4 saved before execution, both equal an independent histogram formula, all2401 ordered repeated-index tuples published per row. Preliminary controls at (q,h)=(30,3),(210,3),(30,5), orders1..4 yielded12 paired exact matches. These are finite screening observations only. Raw producer, shards, observations and manifest are public attachments, hash checked; no original timing reproduced. Two original invocations wall20/CPU10, exits0 and process-group cleanup observed, actual CPU usage unmeasured, RAM unverified. Conditional moment reduction and exact endpoint convention appear in report.
