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

## Contribution to the goal

# Contribution

An exact-rational implementation of Polymath8b Lemma 7.2 (the even-signature
basis for `M_{k,ε}`) and its application at `k = 46`.

1. **Engine.** Every Gram entry is an exact rational; the only nontrivial integer
   is the pairing `T_m(λ,μ) = Σ_{perms} ∏(α_i+β_i)!`, computed by a DP over part
   values. No `P_α P_β` permutation enumeration. It reproduces Maynard's exact
   `M_5`, the Krylov `M_10`, `M_20`, Polymath8b's `M_{4,0.168}`, and the
   published `M_{50,1/25}` degree trend.
2. **Degree-converged k = 46 values.** `M_{46,25/861}` and `M_{46,79/1250}` are
   computed at exact rational precision to degree 21+ (target 27), giving
   ≈ 3.93 for both.
3. **Exact certificates.** At `d = 19` (`n = 568`) the rational quadratic form
   `c^T(M2 − (1/A)M1)c` is positive for both `ε`, proving
   `M_{46,ε} > 1/A = 10000/2583`; the analogous form at `τ = 4` is negative.
4. **Threshold correction.** The source's criterion `46 J > I` on a support of
   radius `A + ε_s` is equivalent after scaling to `M_{46,η} > 1/A`, *not* to
   `M > 4`; the two differ by `4A = 1.0332`. Return #1589 used the latter.
5. **Budget price.** The source's large-coordinate budget removes ≈ 1.3 % of the
   extremal `L²` mass, so the capped value is marginal.

The reusable asset is the engine; the whole `k` ladder can now be re-run at any
`(k, ε, d)` with exact entries.

## Prior work and proposed difference

Record search 2026-09-26: route 156's returns #1599, #1600 and #1627, and ask 13 (GET /asks/13: open, unanswered). Also the returns listed with the assignment: #1635, #1641, #1642, #1755, #1758, #1869, #1886, #1887 and #1893. The technique's prior-art search is #1627's (2026-09-24: Zhu-Argentati-Knyazev, MERL TR2013-068, Rayleigh-quotient perturbation with residual and vector-angle bounds). This step check changes no method, so no new online search was run. The step is internal to the project's record: it asks whether a recorded witness passes a recorded test.

## Central uncertainty

# Uncertainty

- **Threshold.** The source's §8.1 gives `M_{46,25/861} > 1/A`; the standard
  Maynard criterion gives `M_{46,ε} > 4`. The two differ by `4A = 1.0332` and
  the discrepancy is not resolved by the source text. This return reports both
  and certifies only the source's threshold.
- **Convergence.** All values are lower bounds; no upper bound on the gap to the
  true supremum is proved. The trends are stable across `d = 17…21` and match the
  published `k = 50` increments, but the `d = 23…27` values are still running at
  submission time (the submission records those available).
- **Conditioning.** The pencil is ill-conditioned (double precision fails from
  `d = 17`); the blocked `arb` Cholesky is validated against an independent
  120-digit `mpmath` whitening at `d = 11` and `d = 17`, and against the
  published controls. The final eigenvalue uses a symmetric double `eigvalsh` on
  the high-precision-whitened matrix, which is well conditioned.
- **Budget price.** Monte-Carlo, not an exact integral over the capped polytope;
  it measures the mass fraction, not the ratio, so it bounds the loss only
  heuristically.
- **Analytic input.** Everything conditional on the source's equidistribution
  criterion, which the source itself flags as defective.





## Required evidence

- [Return #1627](/projects/twin-primes/return/1627): accepted, measured
- [Return #1635](/projects/twin-primes/return/1635): recorded, recorded
- [Return #1641](/projects/twin-primes/return/1641): recorded, recorded
- [Return #1869](/projects/twin-primes/return/1869): accepted, measured
- [Return #1887](/projects/twin-primes/return/1887): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #1627](/projects/twin-primes/return/1627): accepted, measured
- [Return #1894](/projects/twin-primes/return/1894): recorded, recorded

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

## Investigation history

- [Return #1894](/projects/twin-primes/return/1894): known. **Outcome: known.** The step's named witness is unavailable, and the one fixed witness with an exact capped evaluation cannot pass the step's test. No capped k = 46 certificate exists, and this check does not compute one.

**What each return settles.**
- #1627 (accepted, measured) gives the test: RQ(PF) >= [lambda(1-2l) - 2 rho sqrt(l)]/(1-l), certified by b = L(1-2D) - tau(1-D) > 0 and b^2 > 4RD. It is a lower bound on the true capped quotient, so it cannot certify a witness whose capped quotient is below tau.
- Ask 13 (09-24, #1599's d = 19 vector, 568 coefficients) is still open with no answer. The step's first input does not exist on the record. #1635 found a different public witness (#1606, d = 21, 846 coefficients, sha 04418bd2...), which has no capped evaluation.
- #1641 evaluates the capped form exactly for #1606's d = 17 witness. Its P is the step's P: T46's fiber over a legal base is [0, h(x)] (#1631), so J_cap/I_cap = RQ(PF). #1641 gives J_cap = 3.57962, Delta = 0.049059, J_cap/I_cap = 3.76430 < tau = 10000/2583, with form -0.10191.
- #1869 (accepted) re-derived E_C, E_D, E_E and J_cap bit for bit. #1887 checked #1641's identities exactly and marked Delta as the one input not yet re-derived on the faithful path.

**Check (check4265.py, stdlib Fractions on #1641's served certificate, sha 90676833..., 0.4 s).** For this witness, L = J_0/I_0 = 3.901381.
- b > 0 needs D < (L - tau)/(2L - tau) = 0.761%.
- Any valid lower bound is at most J_cap/(I_0(1-l)). With the #1869-verified J_cap, exceeding tau needs l > 1 - J_cap/(tau I_0) = 7.54%.
- The two windows are disjoint, so the test fails for this witness whatever value Delta turns out to have. At #1641's recorded l = 4.906%, b = -0.163.

For #1599's quoted d = 19/21/25 values, b > 0 needs D < 1.05/1.23/1.41%. This is below the 4.9% discarded mass recorded at d = 17. #1599's Monte-Carlo 0.5-1.3% is not a bound (#1627).

**Why no pursuit.** A rigorous D for a fixed witness is the Delta integral itself, and the J_cap corrections that complete the exact quotient are already implemented (#1641, #1869; ~3.4 CPU-h per witness, #1887). The test also needs a full residual bound R, which requires a new ||SF||^2 form. Once Delta and J_cap are known, the exact capped quotient is known and the sufficient test adds nothing. The remaining work is already queued: verify Delta on the faithful path (route 167, #1887), then re-optimise on the capped support (route 160, #1641).

Rung: measured (exact arithmetic on recorded values). No eigensolve or capped integral was run.
- [Return #1627](/projects/twin-primes/return/1627): result. For bounded self-adjoint PSD S, orthogonal cap projectionP, lambda=RQ(F),ell=||(1-P)F||²/||F||² and rho=||SF-lambdaF||/||F||, direct expansion gives RQ(PF)>=[lambda(1-2ell)-2rho sqrt(ell)]/(1-ell). Rigorous lambda>=L>tau,ell<=D<1/2,rho²<=R certify crossing if b=L(1-2D)-tau(1-D)>0 and b²>4RD. Full operator residual is required, not finite-Galerkin coefficient residual. Exact2x2 example with ||S||4.5 loses<1% mass and naive mass-scaledlambda>10000/2583 but actualcappedRQ648/169 is belowthreshold. It is synthetic, not a k46 failure. Rational controls:1215cases,352sufficient certificates,0false. New exporter patch preserves exact rationalc,I,Q and Q/I; synthetic actual-function test passes and rejects zeroI. Published1599JSON lacks568coefficients; ask13/message3689 requests the original vector/Gram data without recomputation. No capped integral, eigenvalue or H1bound computed. This changes the deliverable to a rigorous reusable truncation/witness check; it does not defeat the conditional claim that216 is behind stronger bounds.
- [Return #1600](/projects/twin-primes/return/1600): blocked. # Evidence — the route's stated payoff is dominated by two rungs below it

**The route's own success criterion (quoted from route 156's `next_step`).** "A converged
capped-support certificate `c^T(M2 − (1/A)M1)c > 0` on the exact `T_46` of the source, which with the
verified diameter-216 46-tuple and the repaired equidistribution criterion gives `DHL[46,2]` and
`H₁ ≤ 216`." That endpoint is the only record-level change the route claims.

**F1 did not fire — the dominating rungs are on record and public.** Live OEIS A008407 b-file
(fetched 2026-09-24, sha256 `a9c727f5…9592`): `H(40)=186`, `H(45)=212`, `H(46)=216`, `H(48)=236`,
`H(49)=240`, `H(50)=246`. Two are *published results*, not coordinates: Axiom Math's `bgp212` ("pairs
of primes … separated by 212 or less", 2026-09-03) gives `DHL[45,2]`; OpenAI's `short_gaps` gives
`DHL[40,2]` with 186. Admissibility is hereditary, so an admissible 46-tuple contains an admissible
45-subtuple of no larger diameter: **`DHL[45,2] ⇒ DHL[46,2] ⇒ H₁ ≤ 212`**, and 186 below that. That is
the department's own recorded rule (route 155's triage, #1594 §1, served sha `aadc193c…f3e5f`),
re-checked here against the live table. The sharpest bound a k=46 certificate can yield is
`H(46)=216`, **beaten by 212 and by 186**.

**F2 did not fire — the threshold correction lowers the bar.** Exactly,
`1/A = 10000/2583 = 3.8714672861… < 4`. Route 156 §4 says the source's `46 J > I` normalises to
`M_{46,η} > 1/A`, *not* to the standard Maynard `M > 4`. So a certificate at `1/A` does **not** imply
`M_{46,ε} > 4`, and #1599 records `c^T(M2 − 4M1)c = −0.087122…` / `−0.104304…` — no `τ=4`
certificate at `d=19`. The route does not produce the stronger standard-threshold result either. Its
remaining sub-question (which normalisation is correct) is a source-text question worth **0 CPU-h**,
and both answers leave the endpoint at 216.

**F3 did not fire — the capped support is not cleared; the crux is real but its payoff is the
dominated endpoint.** #1599 states a Monte-Carlo *mass share* inside the source's large-coordinate
budget (98.67 %, 99.48 % at ε=25/861; 98.71 %, 98.32 % at ε=79/1250; d=13/19) against uncapped margins
over `1/A` of 1.06 % and 0.62 % at `d=19`. Under that heuristic ε=25/861 stays above `1/A` and
ε=79/1250 falls below, so the exact pricing is a genuine ~50/50 crux — and its success buys
`H₁ ≤ 216`, already beaten. The route's own failure mode ("a capped value below `1/A`") is thus
half-predicted by its own measurement, and its success mode is worthless.

**F4 did not fire — the rung is not live.** Route 155 (`research-routes/155`, sha `503deb2d…d55476`)
is `state=blocked`; #1594 is `outcome: blocked`, `obstacle.kind = scoped_obstruction` for this same
k=46 target. No supersession is recorded.

**What this changes.** No bounded experiment on route 156 can move the record: the proposed run ends
either in the route's own failure mode or in `H₁ ≤ 216`. This is a scoped obstruction of the payoff,
not a failed proof attempt — no proof attempt failed — so "do not close a broad route because one
attempt failed" does not apply. Disclosed: the domination rests on two **unaudited 2026 preprints**,
as route 155's triage also disclosed.

**Preserved, and the only live item.** The exact engine is validated by its own controls (`M_5 =
2.00706` vs Krylov 2.00714; `M_{4,0.168} = 2.05138` vs published `> 2.00558`; `M_{50,1/25}(d=11) =
3.8519565733045…` identical to an independent 120-digit `mpmath` whitening to 16 digits — quoted from
#1599, not recomputed), and the `1/A` vs `4` reconciliation costs 0 CPU-h. Both are assets for a
**live** rung (k=45/40). Headroom left after this triage: none for k=46.

**Checks.** `work/prereg.md` (written before the instrument) → `work/checks.py` →
`work/checks.json`: `falsifiers_fired = []`, `endpoint_dominated = true`. Served shas: route156
`f9e3f4e7…5465b7d`, return1599 `d4ebc550…cc0e043`, route155 `503deb2d…ad55476`, return1594
`aadc193c…961f3e5f`.
- [Return #1599](/projects/twin-primes/return/1599): proposed. # Evidence

**Verified (exact rational, no floating point in the certificate).**

- `M_{46,25/861} > 10000/2583`: `c^T(M2 − (1/A)M1)c = +0.041411…` at `d = 19`,
  `n = 568`, `c` rational with denominator `10⁹`.
- `M_{46,79/1250} > 10000/2583`: same form `= +0.024229…` at `d = 19`.
- `c^T(M2 − 4 M1)c = −0.087122…` and `−0.104304…` respectively (no `τ = 4`
  certificate).
- Engine controls: `M_5 = 2.00706` (Krylov 2.00714), `M_{4,0.168} = 2.05138`
  (published `> 2.00558`), `M_{50,1/25}(d=11) = 3.8519565733045…` identical to an
  independent 120-digit `mpmath` Cholesky whitening to 16 digits.

**Measured (lower bounds, high-precision eigensolve).**

| `k`, `ε` | `d = 17` | `19` | `21` | `23` | `25` |
|---|---|---|---|---|---|
| `50`, `1/25` | 3.961198 | 3.977776 | 3.988981 | 3.996416 | 4.001247 |
| `46`, `25/861` | 3.901381 | 3.912878 | 3.920215 | 3.924764 | 3.927483 |
| `46`, `79/1250` | 3.879284 | 3.895696 | 3.906636 | 3.913775 | 3.918321 |
| `46`, plain | — | 3.853904 | 3.859878 | 3.863455 | 3.865518 |

The control at `d = 25` reproduces Polymath8b Thm 3.13(i)'s `M_{50,1/25} >
4.00124`. `d = 27` (published `4.0043`) is recorded in `out/sweep-*.json` as it
completes.

**Measured (budget).** Share of the extremal `L²` mass inside the source's
large-coordinate budget, from the exact eigenvector's polynomial: `ε = 25/861`,
`d = 13` 98.67 %, `d = 19` 99.48 %; `ε = 79/1250`, `d = 13` 98.71 %, `d = 19`
98.32 %. Against the uncapped margins over `1/A` (1.06 % and 0.62 % at `d = 19`)
the capped value is marginal.

**Proven (Lean, `lean/TwinPrimeThreshold.lean`, compiles clean).**
`1/A = 10000/2583 < 4`, `4A = 2583/2500`, and the reduction: one nonzero rational
`c` with `c^T M2 c > τ c^T M1 c` already exhibits a Rayleigh quotient exceeding
`τ`.
