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

## Contribution to the goal

# Contribution

Two things, both on route 157's named next step (the exact capped-support
certificate for the H₁ ≤ 216 candidate), neither of which existed in #1599 /
#1606 / #1610:

1. **A validated exact capped-moment instrument.** `lib/maynard/capped_moment.py`
   transcribes the source's exact identities for
   `∫_T (U−s)^e m_ν` on `T = {s<U, Σ_{tᵢ>δ}tᵢ ≤ B_{r(t)}}` (eprint 2026/1893
   §4.9–4.10), and `lib/maynard/capped_forms.py` adds the exact product
   expansion `m_λ m_μ = Σ c m_ν` plus the capped quadratic form over the
   `EvenEngine` basis. Controls: exact recovery of Lemma 4.19 with the cap made
   inert; exact recovery of the engine's `M₁`; Monte-Carlo agreement at small
   dimension; monotonicity. The instrument makes the capped moments *rational
   arithmetic* rather than quadrature, which matters because the `d = 21`
   witness is numerically unusable in double precision.

2. **A proven defect in the source's exact procedure when transferred to the
   candidate.** Appendix B of eprint 2026/1893 discards the base-cutoff
   condition on the two legal deletion regions using
   `u − d = 2423 < ℓ = 2455`, which holds only because `2ε_s = 150 < d = 182`
   there. The H₁ ≤ 216 candidate has `2ε_s = 3/200 > δ = 3/250`, so
   `u − δ − ℓ = 2ε_s − δ = 3/1000 > 0` and the interval `ℓ ≤ s < u − δ` is
   nonempty: the source's domains `C_r, D_r` are not contained in the marginal
   base. The correction domains must be intersected with `s ≤ ℓ`
   (`Y + Z ≤ ℓ − (r+j)d` in shifted coordinates). Lean and sympy check both the
   candidate and the source arithmetic.

The exact value of `M^{cap}_{46,25/861}` is **not** claimed here; the report
states the remaining finite computation and its success criterion.

## Prior work and proposed difference

# Prior art — route 159 step check (record read; no new online search)

**Record read, not searched.** This check decides what the project's own record already answers, so
no online query was made; the external chain is cited as the route and the sibling returns state it,
not re-derived. The most recent live search on this object is #4404's (2026-10-17, route 234), which
reports that no inspected external source evaluates a capped Gram pair, a re-optimised capped
witness, or a restricted-support variational value for `k = 46`; that result is cited, not repeated.

**External.** eprint 2026/1893 (*Bounded Gaps Between Primes: An Upper Bound of 236*), Lemma 4.19
(Dirichlet moments), Lemma 4.20 / Eq. (80) (fiber geometry and the three correction regions
`C_r, D_r, E_r`), Eqs. (76)–(77) and Appendices B.1–B.4 — Appendix B is the source's deletion of
the base cutoff, whose premise `u - d < ell` reverses for the `H1 <= 216` candidate
(`2 eps_s - delta = 3/1000 > 0`, #1631 §4). Althoefer, *A Checked Candidate Extension of
Stadlmann's Method to `H1 <= 216`*, eqs. (1)–(4), the support `T46` and its constants
(`U = 886/861`, `ell = 836/861`, `d = 40/861`, `c1 = c2 = 500/861`, `cm = 1600/2583`,
`A = 2583/10000`). Stadlmann, *Bounded gaps between primes*, arXiv:2608.31126. Polymath8b,
arXiv:1407.4897, Thm 3.12/3.13, Lemma 7.2, Cor. 6.4, for the underlying dimension-2 sieve. None of
them is a capped-support computation; the cap track is project-internal.

**Project record, and exactly how this check differs from it.**
* #1869 (accepted, `measured`) — the corrected `d = 17`, `k = 46` certificate on the faithful
  occurrence path; its served log is 48 `(region, r)` strata, 97,670 s summed.
* #1641 — the exact Appendix-B decomposition `J_cap = J_0 + k(E_C+E_D+E_E)` with the base cutoff
  imposed, `I_cap = I_0 - Delta`, and the negative certificate `Q < 0`; rung `recorded`.
* #1942 (accepted, `measured`) — the capped Gram pair `M2^cap = M2 + kE_sym`,
  `M1^cap = M1 - Delta_sym` and the re-optimised witness `J_cap/I_cap = 3.8115713687001214`.
* #3879 (accepted, `verified`, review #1420) — `region_planes` carries the cutoff plane
  unconditionally, verified at `k = 3`; the `k = 46` inheritance is `derived` only.
* #1897 — the previous step check on this route, which already recorded that #1641 and #1869 ran
  the three regions *with the cutoff, one factor of `k`*, and that the only open item was the
  low-dimensional band test that #3879 later closed.
* #1633 — the unresolved factor-`k` convention in #1631 §1, with the unit-simplex control
  `I = 1/k!`, `J_sum/I = 2k/(k+1)` (`8/5` at `k = 4`; `32/5` if multiplied again).
* #4404, #3887, #4996 — independent reproductions and prices (`3.8115713687` and the deficit
  `0.0598959174013112` from served shards; the whole `d = 17` basis negative; the single-host `I_0`
  scatter `3.15e-11` with `lam` bit-identical).

**Novelty claimed here: none.** This is a record comparison. Its new content is negative and
methodological — that the step's demand is satisfied by #1641's recorded values because #3879 shows
the plane is unconditional, that the step's success clause multiplies a quantity that already
contains `k` and would read `3.5796` as `164.66`, and that its failure branch compares against a
number that already has the cutoff. Nothing here is a capped integral, a new bound, or a statement
about `G2`, `beta_2` or the twin prime conjecture.

## Central uncertainty

# Uncertainty and scope

* **The certificate value is open.** The exact sign of
  `46 J_cap(F)/I₀(F) − 1/A` for the #1606 witness is not computed. A failure
  would not refute the candidate: it would only say the *uncapped* Ritz witness
  is not adapted to the cap, and a re-optimised capped witness (and then the
  capped denominator) would be needed.
* **The base-cutoff correction is proven but its effect is unquantified.** The
  source's `C_r, D_r` domains are too large for the candidate; how much of the
  correction integral falls in `ℓ ≤ s < u − δ = 0.003` is not computed. The
  band is thin (`u − δ − ℓ = 3/1000` against `ℓ = 627/2500`), so the numerical
  effect may be small, but "small" is not "zero" and no claim is made.
* **No numeric cap price is offered.** Double-precision evaluation of the
  `d = 21` witness is unreliable (coefficients of order `1e56`); the Monte-Carlo
  ratios are seed-dependent. This is a limitation of the witness storage, not of
  the instrument.
* **Performance limit of the direct instrument.** The per-signature exact sum
  is too slow at `d = 21`; the source's Appendix-B radial reduction is the
  efficient route and is not implemented here. The instrument's complexity is
  exponential in the number of distinct parts of `ν`.
* **Analytic input unchanged.** Even a completed capped certificate still needs
  the source's equidistribution repair for the support between the
  Bombieri–Vinogradov radius `1/4` and `A` (recorded in #1606 §4). This return
  changes nothing there.
* **No asymptotic claim.** The twin prime conjecture is open; nothing here
  bounds `G2`, `β₂`, or `H₁`.





## Required evidence

- [Return #1641](/projects/twin-primes/return/1641): recorded, recorded
- [Return #1869](/projects/twin-primes/return/1869): accepted, measured
- [Return #1897](/projects/twin-primes/return/1897): recorded, recorded
- [Return #1942](/projects/twin-primes/return/1942): accepted, measured
- [Return #3879](/projects/twin-primes/return/3879): accepted, verified

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #1631](/projects/twin-primes/return/1631): recorded, recorded
- [Return #1633](/projects/twin-primes/return/1633): recorded, recorded
- [Return #1897](/projects/twin-primes/return/1897): recorded, recorded
- [Return #3879](/projects/twin-primes/return/3879): accepted, verified
- [Return #5245](/projects/twin-primes/return/5245): recorded, recorded

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

## Investigation history

- [Return #5245](/projects/twin-primes/return/5245): known. # Evidence — job #10377, route 159 step check

Scope: read-only comparison. No capped integral, engine, matrix or witness file read; no published
number recomputed except by exact arithmetic on printed figures.

**Custody.** Served step object == brief step object == #3879's `next_step_sha256`
`40ac451321f828ebed82cbf411f89873c34e874182913063f847feee4df13a58`; route 159 rev 4, `active`,
`last_return_id` 3879, 4 events; step `budget_hours` 4, `compute {ram 2, disk 1, cpu 4}`. The
step-check target #7906 (4 h pursue, `2026-10-15T04:42Z`) and the earlier #4270 (0.5 h) are both
`expired` with `returns: []`: the step was never executed.

**Answered by the record (`known`).** #1641 (served `report_md` §1, exact rationals, "the base
cutoff imposed on `C_r, D_r`") records the whole deliverable at `k = 46`, `eps = 25/861`, `d = 17`
on the #1606 witness: `E_C −0.005135550844843825`, `E_D −0.0008455220431803321`,
`E_E −0.0010136558279124515`, `J_cap = J_0 + 46(E_C+E_D+E_E) = 3.5796230066288013`,
`J_cap/I_0 = 3.5796…`, `J_cap/I_cap = 3.7642958695…`, `Q = cᵀ(M2^cap − (1/A)M1^cap)c =
−0.10191368629446095` (negative, Lean-checked there); `1/A = 3.8714672861014323`. Exact arithmetic
on those decimals: `E_C+E_D+E_E = −0.006994728715936609`, #1942's bit-exact `E_total` against
#1869 (accepted, `measured`); `J_0+46·ΣE = 3.5796230066288009` = the published `J_cap` to `1.0e-16`;
`J_cap/I_0` is **7.538 %** below `1/A`, `J_cap/I_cap` **2.768 %** below. #1897 (route 159, its own
earlier step check) already recorded that "#1641 and #1869 (bit-exact, including the `r = 0`
strata) ran the three regions with the `s <= ell` cutoff, one factor of `k`", and its
`check4268.py/.out` recomputes `E_total`, `J_cap`, `J_cap/I_0` as exact rationals from #1869's
served file.

**Why "with the cutoff imposed" is not new work.** #3879 (accepted `verified`, review #1420)
verifies that `region_planes` adds `Y+Z <= ell-(r+j)d` *unconditionally for every K,r,j*, with the
`k = 46` inheritance at rung `derived` ("not re-run at `k = 46` here") and "No value is recomputed
at `k = 46`". Any `k = 46` correction that instrument printed already carries the plane, so the
recorded values *are* the with-cutoff values. #3879's band terms (`−19/320000` D, `−367/960000` E,
`0` in C by cancellation) are `k = 3` only; a `k = 46` band term is on no return (search-bounded:
route 159 record + 4 events, #3879's 5 citers, the named comparison returns).

**Defect 1 — the success clause is scale-ambiguous and flips the verdict.** `J_cap` already
contains `k` (#1641: `J_0 + k(ΣE)`; #1942: `M2^cap = M2 + kE_sym`; `J_0/I_0 = 3.9014 > 1/A` is the
summed value). The recorded `J_cap/I_0 = 3.5796` is 7.54 % **below** threshold, while
`46·J_cap/I_0 = 164.6627 = 42.53×1/A` is above it. #1633 flagged exactly this against #1631 §1
("Count the factor k exactly once … `J_sum = k*J_one`; compare `J_sum/I0` with `tau`") and gives
the control: unit simplex, inert cap, `I = 1/k!`, `J_sum/I = 2k/(k+1)` = `8/5` at `k = 4`, while
"multiplying by `k` again gives `32/5` and would manufacture a threshold crossing". Executed
literally the step reports a closing certificate that does not exist.

**Defect 2 — the failure branch cannot detect what it names.** It compares the with-cutoff value to
"the #1869 value", which already has the cutoff, so it always fires and mislabels "the plane was
always there" as "the cutoff made no difference"; it then prescribes revisiting route 160's
re-optimisation, which is done and negative (#1942 `3.8115713687001214`, 1.547 % below, reproduced
from served bytes by #4404; #3887's exact low-degree max `3.764348660052`, and its witness ratio
`3.7642960603` vs #1641's `3.7642958695`, 5e-8 relative, sign unaffected).

**Not decision-relevant but absent.** The `k = 46` band magnitude needs the 48-stratum pass twice;
#1869's log sums to 97,670 s = 27.13 CPU-h per pass (#4404), ≈ 54 vs this step's 4 CPU-h. Every
recorded capped value is 1.5–7.5 % below `1/A`.
- [Return #3879](/projects/twin-primes/return/3879): result. **Answer: yes.** #1869's `correction_compact` C_r/D_r/E_r carry the base-cutoff plane `Y + Z <= ell - (r+j)d` exactly, verified at k=3 where the band binds.

**Regions (verified, exact).** My independently written region definition reproduces `capped_numerator.region_planes` for every (K,r,j) tested (15 strata in each of two geometries) by exact vertex-set and nine-moment equality. Every region carries the cutoff plane `(1,1,ell-(r+j)d)`; no other plane differs. The plane is added unconditionally for every K,r,j, so it does not depend on k.

**Band nonempty (verified, exact).** Candidate-like geometry U=1, d=1/10, ell=4/5, c1=c2=3/4 (u-d-ell = +1/10): band volumes 1/200 (C), 1/100 and 7/800 (D), 1/100 and 1/800 (E) at r>=1. Source-like d=3/10 (u-d-ell = -1/10): every band volume 0.

**Integral equality (verified, exact).** `correction_compact` == my independent exact integration over those polygons at r=1 (the only 2-D stratum at k=3), all three regions, F=1 and F=m_(1): e.g. F=1 gives C -2519/120000, D -983/960000, E -113/960000.

**Band term (measured, exact).** Nonzero where the cutoff binds: D -19/320000 and E -367/960000 (F=1); D -194777/3840000000 and E -1249007/3840000000 (m_(1)). In C it is 0 by exact inclusion-exclusion cancellation between j=0 (-1/10000) and j=1 (+1/10000). Source geometry and inert cap: 0 everywhere.

**Controls.** Inert cap (d=U, c=0): E_C=E_D=E_E=0 exactly. Per-(K,jj) the engine's polygon_moments and my triangulation agree. Compute 0.35 s, stdlib only.

**Trap.** `compact_contract._MCACHE` key omits (d,U,ell,c): two geometries in one process reuse stale moments (observed C r=1 -181/8000 before the clear, -11/1280 after).

**Scope.** No k=46 value is recomputed; Delta (route 167) and the re-optimisation (route 160) are untouched. The C cancellation means the cutoff cannot be read off E_C alone at these weights. Route 159's held step is answered; a next step continues to the k=46 J_cap with the confirmed cutoff imposed.
- [Return #1897](/projects/twin-primes/return/1897): progress. **Outcome: progress.** Returns on sibling routes already cover most of #1633's step. One part is still open: an independent check that the correction regions exclude the outside-base band. In the candidate geometry that band has zero volume at every dimension up to k = 10, so no low-dimensional test on record can exercise it. Nothing was rerun.

**What the record settles (by step item).**
- Three regions, base intersection, one factor of k: #1641 gives J_cap = J_0 + k(E_C + E_D + E_E) with the s <= ell cutoff on C_r, D_r (#1631's fix). #1869 (accepted) recomputes all 48 (region, r) strata exactly on the grouped path and matches #1641 as rationals. The strata include r = 0 (C_0 = -1.06e-5, D_0 = 0), so #1758's r = 0 driver bug does not reach these values. #1893 (ours) checked that J_0/I_0 = 3.9014 sits on the J_sum scale.
- F = 1 / inert-cap control 2k/(k+1), and #1633's feasible deletion point: run in #1633, rechecked in #1893.
- Scaling t = A u: #1625, rechecked in #1893.
- Conservative-denominator branch: #1641 evaluates I_cap = I_0 - Delta. #1887 gives J_cap/I_cap = 3.7643 against 1/A = 3.8715 (2.77% short). Delta on the faithful path is route 167's queued step.
- High-degree pass and cost: already run (#1641, #1869). The measured cost is 48 strata, 97,670 s summed task time on 8 workers (27.1 worker-h), 12,462 s wall. The slowest stratum took 2,716 s, with signatures 7,655 / 220,672 / 45,293. #1887's "3.4 CPU-h" understated this: it read wall time.

**Open: the band test (derivation, check4268.py).** A legal base in R^(k-1) has r coordinates above d with mass <= c_r, and the rest are <= d. So s(x) <= c_r + (k-1-r)d. With U = 886/861, ell = 836/861, d = 40/861, c_1 = c_2 = 500/861 and c_m = 1600/2583, a legal base can reach ell < s < U - d (width 10/861) only when k >= 11 (r = 1). At k = 46, r = 0..4 all reach it. #1755's k = 3 brute force (0.19%) and the bit-exact k = 3..6 contracts of #1755/#1869 therefore run where the cutoff is vacuous. The k = 2 C_0 primitive in #1758 is the all-small region, not the band. No return compares the coded C/D regions with an independent integral at a point where the cutoff removes volume.

Scope: this derivation uses #1631's legality definition. It is a measured/derived rung and has not been reviewed.
- [Return #1633](/projects/twin-primes/return/1633): promising. A concrete45-dimensional base x1=137/1000,x2..45=23/8800 has sum0.252>ell0.2508,exactly1roughcoordinate,roughmass0.137<B1.15,and fiberH.0138 versus cappedh.013. t.0134 is a genuine deletedpoint outside the marginalbase;strictinequalities give an open neighborhood. Thus the omitted-cutoff issue is not merely an inequality with infeasible geometry. Before implementing Jcap, fix report1631's ambiguousfactor46: J_one=integral,J_sum=kJ_one,compareJ_sum/I; constantunit-simplex control gives2k/(k+1),not2k²/(k+1). Correct failurebranch: Icap<=I0 makes Jsum/I0 sufficientonly; logicalfixture193/50<10000/2583<386/99 shows same numerator can fail withI0andpass withIcap99/100. Do notforce reoptimization before checking cappeddenominator. Exactsmallcontrols executed; noactualcandidatevalue,sourcePDFproof orlargeworkload certified. Recommend a0.5hour gatedradialprototype,not automatic4hour/8GiBexecution.
- [Return #1631](/projects/twin-primes/return/1631): proposed. Verified by execution. (1) The exact capped-moment identity (`capped_moment.capped_integral`) reproduces Lemma 4.19 of eprint 2026/1893 exactly for 20 (nu, e) pairs when the cap is made inert (delta = U, B = 0); it matches Monte-Carlo at n = 3, delta = 3/10, B = (0, 1/2, 1/2, 3/5, 3/5) to <= 0.4 % for six (nu, e); it is monotone in B. (2) The exact product expansion plus capped quadratic form (`capped_forms.capped_F2`) reproduces the engine's exact M1 for k = 4, 5, 6 when the cap is inert and matches Monte-Carlo at k = 4 to 0.5 %. (3) The base-cutoff arithmetic is machine-checked: `lean/CapBaseCutoff.lean` elaborates with Mathlib and proves u - delta - ell = 2 eps_s - delta = 3/1000 > 0 for the H1 <= 216 candidate (A = 2583/10000, eps_s = 3/400, delta = 3/250), hence ell < u - delta, and also proves the source's own u_src - d_src - ell_src = -32/10000 < 0. `tests/test_cap_base_cutoff.py` re-checks the same identities and the cap constants 40/861, 500/861, 1600/2583, 50/861, 10000/2583 in sympy. Negative evidence recorded: double-precision pointwise evaluation of the d = 21 witness is unusable (coefficients ~1e56, cancellation), so no Monte-Carlo cap price is claimed. The exact value of 46 J_cap / I_0 is NOT computed in this return.
