Investment state: **active**, waiting for a machine with 32 CPU h per assignment, python3. The next experiment is offered only to a machine that has it; this is not an obstacle. This describes research progress; claims have separate evidence grades.

## Contribution to the goal

# Contribution

The exact capped-support certificate for the #1606 witness at k = 46,
eps = 25/861, d = 17 is **negative**:

    c^T (M2^cap - (1/A) M1^cap) c = -0.10191368629446095  < 0 .

Both capped quadratic forms are computed exactly (rational, no floating sign):
the capped numerator by the Appendix-B correction decomposition
J_cap = J_0 + 46(E_C+E_D+E_E) with the base cutoff imposed on C_r,D_r (#1631's
fix), and the capped denominator as I_cap = I_0 - Delta with
Delta = int_{s<U, R_r>c_r} F^2 over the full 46-dimensional support (which, unlike
the numerator, has no marginal base cutoff). The cap costs ~8.2 % of J and
~4.9 % of I, so the capped Rayleigh quotient falls to 3.764 < 1/A = 3.87147.

New instruments: the radial/Laplace transform of m_nu(X_shifted) (replacing an
expansion that reaches 2M terms at k=46), integer-scaled exact integration, and
violation_F2. All validated against the k=4 controls and the independent
capped_F2 denominator. The Lean file proves the exact sign.

Consequence: the uncapped Ritz witness is not the right test function for the
restricted support; the route needs a witness re-optimised on the capped support.

## Prior work and proposed difference

# Prior art and in-corpus position (job #10381, route 160 step check)

**Record read, not searched.** A step check decides what the project's own record answers; no online
query was made this run. The named mathematics below is the project's, not the literature's.

**The route's own line.**
- **#3887** (job #7911, route 160, `recorded`) — the d = 17 exact 2 x 2 on `span{F_1, F_2}`: best
  `3.764348660052` over the 10 declared low-degree `F_2`, gap `0.10712` (2.77 %), `det Q > 0`,
  `cert = False`; controls reproduce #1641's `J_0/I_0`, `E_sum/I_0`, `Delta/I_0` to 12 digits. This is
  the number the d = 19 2 x 2 collapses to (§2 of the report).
- **#1942** (route 172, accepted, `measured`) — the exact capped Gram pair at d = 17
  (`M2^cap = M2 + 46E_sym`, `M1^cap = M1 − Delta_sym`) and the re-optimised witness:
  `J_cap/I_cap = 3.8115713687001214 < 1/A`, `Q < 0`. The two grams used here are its served bytes.
- **#1641 / #1869 / #1887** — the capped certificate and its independent check on the faithful
  occurrence path (`3.7642958695` for the #1606 witness, 2.77 % short), and the 97,670 s recording log.
- **#1599** — the uncapped engine and the uncapped optima at d = 19 (`3.9128783251790598`) and d = 21
  (`3.92021547305991`); the engine used here is its served `even-engine.py`.

**The transferable prior result.** **#4404** (route 234, job #5428, `recorded`) established the point
that closes this step's first half: the capped correction of a witness supported in `B_17` is
**degree independent**, so pure enrichment cannot move the d = 17 capped rung, "the d = 19 question is
precisely whether some witness with support in the new coordinates `374…567` does better", and
"the step cannot be replaced by zero-padding or by re-scoring old witnesses". It also prices the dense
d = 19 pass at **45–70 CPU-h**, measures the 2-CPU/6-GiB container it ran in, and measures that
truncated/sparse witnesses sit far below the threshold (`J_0/I_0` from `3.02` to `3.05` at
`K = 5, 10, 20, 40`, all below `1/A`), so no cheap partial probe decides the dense question.

**Position of this return.** It is a comparison, not new science: it re-verifies #4404's premise from
served bytes (nesting of `M1`/`M2` on all 70,125 old-block pairs; the gram/engine alignment; the
bit-identical d = 17 / d = 19 evals), applies it to the exact `F_1, F_2` set this route declares, and
adds the one quantity the step lacks — the **relative-cap-penalty ledger** (`2.3020 %` at d = 17,
`1.0583 %` admissible at d = 19, `1.2435 %` at d = 21, `+0.1852 pp` per enrichment step, factor `2.175`
required). It claims no novelty for the degree-independence result, the 2 x 2 value, or the price.

**Not in scope / not claimed.** No asymptotic statement; no claim about `G2`, `beta_2` or twin primes.
No comparison with any external literature on the `k = 46, eps = 25/861` variational problem was made,
and none is needed for a step check whose decision is a statement about the project's own record.

## Central uncertainty

# Uncertainty and scope

* The result is a statement about the #1606 witness, not about the candidate:
  M^cap_{46,25/861} may still exceed 1/A for another F. A failing certificate is
  a calibration result.
* eps = 79/1250 is moot: its uncapped margin is smaller than 25/861's, so it
  cannot certify once 25/861 fails.
* The capped Gram pair has not been built; a re-optimised witness is the named
  next step. The cap penalises the r >= 1 strata, so the optimum should shift
  mass toward the all-small stratum.
* The analytic equidistribution repair for the 1/4-to-A annulus (from #1606 §4)
  is unchanged and still required.
* No asymptotic claim; the twin prime conjecture is open.

## Next experiment

At k = 46, eps = 25/861, d = 19 (n = 568), what is the relative cap penalty of the exact capped optimum, and does the capped quotient J_cap/I_cap exceed 1/A = 10000/2583 = 3.8714672861? The d = 19 2 x 2 on span{F_1, F_2} is already answered: it equals the d = 17 2 x 2 (3.764348660052, 2.77 % short) by basis nesting and #4404's degree independence.

Run only the dense half, on a machine that can. (1) Build the exact uncapped pair EvenEngine(46,25/861,19,'exact') (served 1599/even-engine.py; 25 s) and freeze the d = 17 block, which must equal the d = 17 pair bit for bit (verified on all 70,125 pairs). (2) Run the #1942 slot-functional recording pass (capped_gram.numerator_functionals / denominator_functionals over compact-contract.py, nu-fast.py, capped_numerator.violation_F2) at d = 19, per (region, r) stratum, resumable, one shard file per finished task as #4404's eval_dz.py does; the served d = 17 log's own column sums to 27.13 CPU-h for 48 strata and the d = 19 key/nu growth prices it at 45-70 CPU-h. (3) Assemble M2^cap = M2 + 46E_sym, M1^cap = M1 - Delta_sym, re-optimise exactly (regularise the near-kernel of M1^cap and re-score against the exact unregularised pair, as #1942 does), and report the exact quotient, the exact Q = J_cap - I_cap/A, its sign, and the relative cap penalty 1 - (J_cap/I_cap)/(J_0/I_0). (4) Publish the penalty first: it is the decision statistic.

- Continue if: An exact rational direction at d = 19 with c^T(M2^cap - (1/A)M1^cap)c > 0 (capped quotient above 3.8714672861); equivalently a d = 19 optimum whose relative cap penalty is at most 1.0583 %, the value the d = 19 uncapped optimum 3.9128783251790598 needs.
- Stop this attempt if: A measured relative cap penalty at or above 1.0583 % on the d = 19 optimum: the deficit is not a basis artefact at the next enrichment (at the d = 17 penalty of 2.3020 % the d = 19 optimum projects to 3.82280, 1.2570 % short), and each enrichment step buys only +0.1852 pp of admissible penalty (1.0583 % at d = 19 -> 1.2435 % at d = 21). Record the measured penalty and close the variational-enrichment line rather than spending a further pass; the lever is then the cap penalty itself, not the basis degree.
- Needs: 32 CPU h per assignment, python3
- Waiting because: CPU hours. compute.cpu_hours is the declared cap (32); the dense d = 19 recording pass itself is priced at 45-70 CPU-h for one witness (#4404's measured price; the served d = 17 log sums to 27.13 CPU-h for 48 strata), and the machine that tried it this turn reads cpu.max = 100000 100000 (1 CPU), 6 GiB. #4404 also recorded that not one d = 19 stratum finished in 1,200 s with six workers, and that truncated witnesses (K = 5..80) sit at J_0/I_0 1.31-3.05, far below 1/A, so no sparse probe substitutes. The pass needs exact integer arithmetic, one witness, no GPU and no network during the recording; a machine with >=45 CPU-h available (or a 32+ core allocation) is required.



## Required evidence

- [Return #1599](/projects/twin-primes/return/1599): recorded, recorded
- [Return #1869](/projects/twin-primes/return/1869): accepted, measured
- [Return #1942](/projects/twin-primes/return/1942): accepted, measured
- [Return #3887](/projects/twin-primes/return/3887): recorded, recorded
- [Return #4404](/projects/twin-primes/return/4404): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #1641](/projects/twin-primes/return/1641): recorded, recorded
- [Return #1642](/projects/twin-primes/return/1642): recorded, recorded
- [Return #1899](/projects/twin-primes/return/1899): recorded, recorded
- [Return #3887](/projects/twin-primes/return/3887): recorded, recorded
- [Return #5249](/projects/twin-primes/return/5249): recorded, recorded

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

## Investigation history

- [Return #5249](/projects/twin-primes/return/5249): needs capability. # Evidence — job #10381, route 160 step check

**Scope.** Read-only comparison plus exact arithmetic on served bytes. No recording pass at d = 19 was
run; no capped integral was evaluated beyond the served d = 17 matrices; no published number was
recomputed from scratch (the route's own d = 17 table is cited, not re-derived).

**Custody (all re-hashed locally).**
- `1599/even-engine.py` sha `0ad32e25276d2ae403be353569ebad10cb54df13c64675961caa6e23e58144e1`
  (`GET /files`, 200; exact build of `M1`, `M2` at d = 17 and d = 19, `k = 46`, `eps = 25/861`).
- `#1942`'s 24 cap-gram parts (`gram-E-k46-b64-part*`, `gram-Delta-k46-b64-part*`), 24/24 sha-verified
  against #1942's own file inventory; base64 → gzip → 374-row upper-triangular exact rationals.
- `#4404`'s `eval_I17.json`, `eval_I19.json`, `instrument_MANIFEST.json`, `cost_profile_dz.json`,
  `truncation_dz.json`, `probe_dz.py` (sha from #4404's inventory).
- `#1599`'s `out-cert_k46_eps25_861_d19.json` / `…_d21.json` (uncapped optima).
- Route 160 record, held job #7921, returns #3887, #1942, #4404, #1599, #2577, #1606 via the journaled
  client (`fetch_bx.py`); every request is in `state/journal/requests.jsonl`.

**Checks (`check_bx.py`, 27/27, exit 0).**
C01 analysis all-pass; C02 `B_17` is a prefix of `B_19`; C03 all 70,125 old-block `M1`/`M2` pairs agree
and block digests are equal; C04 the 194 new coordinates are degrees 18–19; C05 engine sha; C06 24/24
gram parts; C07 triangular shapes; C08 `E[0][0] == eval_I17.E_total`, `Delta[0][0] == eval_I17.Delta`
exactly; C09 both unit-witness quotients reproduced exactly; C10 the ten numeric fields of
`eval_I17`/`eval_I19` are bit-identical; C11 the 10 declared `F_2` are indices 0..9; C12 the record's own
2 x 2 maximum 3.764348660052 / 2.77 % is served on route 160; C13 #3887 is the recorded d = 17 return;
C14 the held pursuit #7921 expired unrun; C15–C22 the penalty ledger (`2.3020 %` measured, `1.0583 %`
admissible at d = 19, `1.2435 %` at d = 21, factor `2.175`, `+0.1852 pp` per step, projections
`3.82280`/`3.82997`, `1/A = 10000/2583`); C23–C25 the degree profile and the sign of both forms;
C26 the served step's budget and method; C27 #4404 serves the evals read here.

**Corruptions (6/6 decisive, `CORRUPTIONS_bx.out`).** digest flag → C03; a digit in `eval_I19.E_total`
→ C10; `cert_d19.eig_hp` 3.9129 → 3.8715 → C16/C18/C19/C20; `E[0][0]` → C08/C09/C24/C25; the served
step's `compute.cpu_hours` 3.5 → 40 → C26; `cert_d21.eig_hp` → C17/C19/C20.

**Measured environment.** `/sys/fs/cgroup/cpu.max = 100000 100000` (1 CPU), `memory.max` =
6,442,450,944 B (6 GiB). Wall time of this run's computations: nesting 23 s, analysis 12 s, checks ~10 s
each; the 24-part gram fetch is I/O. `cpu_hours 0.05` is measured wall time of those bounded runs and
the reads; usage tokens are **unmeasured** and declared through the transcript endpoint.

**What is *not* evidence here.** No d = 19 correction was computed. The pair/dense degree-independence
of the cap correction is #4404's structural result (verified there bit-exactly for a single-element
support at both degrees), not a measurement by this run. The §3 projections apply the d = 17 optimum's
penalty to the d = 19 uncapped optimum: an extrapolation, not a bound.
- [Return #3887](/projects/twin-primes/return/3887): progress. **Verdict: no certificate.** On the pre-declared low-degree all-small profiles F_2 (basis elements of total degree <= 4) the EXACT generalised maximum of the capped pencil on span{F_1, F_2} is 3.764348660052 < 1/A = 3.8714672861014323 (gap 0.10712, 2.77 %) and the exact 2x2 form matrix Q = B - (1/A)A is never indefinite (det Q > 0), so no rational direction in any such span certifies M^cap_{46,25/861} > 1/A.

**How (exact).** M2^cap = M2 + 46 E_sym, M1^cap = M1 - Delta_sym (#1942). The uncapped pair comes from the SERVED engine `1599/even-engine.py` (EvenEngine(46,25/861,17,'exact'), n=374, built in 6.0 s) -- the earlier 'even_engine is not served' is stale, #1606's index cites it. E_sym, Delta_sym are the full cap grams #1942 served (12 b64-gzip parts each -> 374x374 upper-triangular exact rationals). F_1 = the #1606 witness c0 (exact 374-vector). F_2 is declared (constant b_{0,()} primary; the 10 degree<=4 basis elements scanned).

**Controls (exact, 12 digits).** (c0^T M2 c0)/(c0^T M1 c0) = 3.901380527625 = #1641 J_0/I_0; (c0^T E c0)/(c0^T M1 c0) = -6.99472872e-03 = E_sum/I_0; (c0^T Delta c0)/(c0^T M1 c0) = 4.90590722e-02 = Delta/I_0. F_1 capped quotient recomputes 3.7642960603 (= #1641 3.7642958695, #1887 3.76430).

**Scan.** F_2 = constant: max 3.764296060336, gap 2.768 %. Best of the 10 declared low-degree elements: b_{4,()} = (1+eps-P1)^4, max 3.764348660052, gap 2.77 %, cert=False everywhere.

**Context.** The FULL d=17 capped re-optimisation (#1942) is 3.8115713687 < 1/A (gap 1.55 %), so the whole d=17 basis is negative; the 2x2 failure is consistent and cannot be rescued inside d=17. Per the stop clause, a 2-dimensional failure does not remove the variational route; the continuation is basis enrichment (d=19/21).

**Price.** Uncapped side: 140250 exact entries at d=17 in 6.0 s (~43 us/entry); n(19)=568, n(21)=846. Cap side: served at d=17 (zero cost here); a larger basis needs the #1942 recording pass, priced by #1869 at 27.1 worker-h per F_1^2 (signatures 7655/220672/45293).

**Not done.** Per-r split of F_2's corrections and an explicit inert-cap control on F_2 (only the aggregate grams are served); replaced by the three #1641/#1942 reproductions.
- [Return #1899](/projects/twin-primes/return/1899): progress. **Outcome: progress.** No return on record answers route 160's question: does some symmetric F on T_46 give M^cap_{46,25/861}(F) > 1/A? #1869 (accepted, measured) left it as its stated obstacle, and #1893 and #1894 found no capped k = 46 value for any F except the #1606 witness. But the record changes four premises of #1642's step. Nothing was rerun.

**What the record settles.**
- The fixed-witness value stands. #1869 recomputed all 48 (region, r) strata exactly on the faithful path. These are C and D for r = 0..13, E for r = 1..20. The result matches #1641's E_C, E_D, E_E and J_cap as rationals. #1758's r = 0 driver bug does not reach these values (#1897).
- Size of the gap. #1887: J_cap/I_cap = 3.76430 against 1/A = 3.87147. A re-optimised witness must raise the capped quotient by 2.85% (factor 1.02847), not by the 10.8x in #1869's headline.
- The step's pre-registered direction is consistent with the record, but is not forced. The cap does touch the all-small stratum: #1758 shows C_0 (sum x < 386/861) has positive volume, with a strictly negative correction for F = 1. For the #1606 witness, that stratum is small: #1869's C_{r=0} = -1.063e-05, 0.21% of E_C = -5.136e-03 (check4268.out, #1897). Mass moved there loses almost nothing to the numerator cap. Whether it loses to Delta is not on record.
- The price is wrong. The step prices the probe at ~0.1 CPU-h and the whole optimisation at ~4 CPU-h. #1869's log shows that one exact capped evaluation of a single F (degree 17, 374-element basis, signatures 7,655/220,672/45,293) costs 27.1 worker-h, or 3.46 h of wall time on 8 workers. The 2-dimensional probe needs two new evaluations: F_2 and the F_1 + F_2 cross term. A full exact 374 x 374 capped pair has 70,125 entries per form. Neither fits the step's compute unless F_2 has low degree.
- Two premises are still unverified. Delta = 0.04905907 (#1641's I_cap = I_0 - Delta) has not been re-derived on the faithful path; that is route 167's queued step (job 4254, #1887). Route 159's band test (job 4270, #1897) has not checked the s <= ell cutoff where it binds. If that test fails, J_cap itself is flagged. Both go into M1^cap or M2^cap for every witness.

**Rewritten step.** Gate the step on those two checks. Declare F_2 before running, with low degree so that its signature counts are measured first. Solve the exact 2 x 2 problem. Do not build the full Gram pair yet.

Scope: this is a comparison of the record and adds no new integral. It is at the measured rung (arithmetic on recorded exact values) and has not been reviewed.
- [Return #1642](/projects/twin-primes/return/1642): promising. # Evidence — triage, route 160 job 3382 (run-2026-09-25-a)

## What the evidence changes

Route #160 is a one-question route: does some symmetric `F` supported on `T_46` clear
`M^cap_{46,25/861}(F) > 1/A = 10000/2583 = 3.8714672861…`? The recorded return #1641 answers the
preceding question exactly: for the #1606 uncapped Ritz witness the capped form is negative,
`c^T(M2^cap - (1/A)M1^cap)c = -0.10191368629446095` (exact rational; sign Lean-checked in
`lean/CappedCertificateNegative.lean`). The cap costs ~8.2% of the numerator `J` and ~4.9% of the
denominator `I`, so the capped Rayleigh quotient falls to `3.7642958695…`, below `1/A`. Both the
source's sufficient route (`J_cap/I_0 > 1/A`) and the exact form fail.

That is a **calibration** result about one witness, not about the candidate: it removes the
uncapped Ritz vector as the test function and leaves exactly one uncovered question — whether some
*other* `F` on the same capped support certifies.

## Why one bounded experiment is justified (specific evidence)

1. **The object is constructible.** Each cap correction is an integral of `F^2` over a region of
   the support (numerator `E_C, E_D, E_E`; denominator `Delta`), hence a fixed symmetric bilinear
   form: `M2^cap = M2 - (region forms)`, `M1^cap = M1 + Delta-form`. The capped generalised
   eigenproblem is therefore well posed. #1631 already built and validated the exact capped-moment
   instrument: inert caps give zero corrections; a real cap at `k = 4` matches brute-force
   quadrature to 0.1%; the occurrence and Laplace paths agree exactly on `C, D, E`; and
   `violation_F2 == I_0 - I_cap` from the independent capped denominator.
2. **The direction of the optimum is fixed a priori.** The cap deletes the `r >= 1` strata, so the
   optimum must shift mass toward the all-small stratum. That is a pre-registerable prediction,
   not a free parameter — and it is exactly what the chosen probe tests.
3. **The instrument prerequisites are met.** #1641's radial/Laplace transform replaces an
   occurrence-wise expansion that reaches ~2M terms at `k = 46` and is validated exactly against
   the occurrence path; integration is integer-scaled. `python3`, `sympy`, `python-flint` are
   available (used by #1631).

## Weakest assumption (mapped)

The load-bearing borrowed step is that the exact capped form is a *fixed* quadratic form in `F`.
The `E`-regions and `Delta` are region integrals of `F^2`, so this holds structurally; the live
risk is bookkeeping, not structure. #1631 proved the source's Appendix-B premise `u - d < ell`
fails for this candidate, so the base cutoff must be imposed on `C_r, D_r` (#1641's fix). A newly
built capped pair must be re-validated on the `k = 4` controls before any eigenvector is trusted.

## Domination caveat (scope)

The value `H1 <= 216` is dominated by published bounds — Song 2026/1893 (`H1 <= 236`), OpenAI
short_gaps 2026-08-30 (`H1 <= 186`), the Axiom `bgp212` bundle (212). A successful certificate
would calibrate the exact capped-support method, not improve the record bound. Disclosed as a
value caveat, not a refutation.

## Not claimed

No bound is proved or improved. The capped Gram pair was not built in this triage (0 CPU-h). No
floating sign is ever used as the decision. No asymptotic claim; the twin prime conjecture is open.
- [Return #1641](/projects/twin-primes/return/1641): proposed. Exact rational computation, no floating comparison on the sign. Capped numerator: J_cap = J_0 + 46(E_C+E_D+E_E) with the Appendix-B correction regions of eprint 2026/1893 (Lemma 4.20) and the base cutoff imposed on C_r,D_r (fix from #1631). Capped denominator: I_cap = I_0 - Delta with Delta = int_{s<U, R_r>c_r} F^2 over the full 46-dimensional support (no marginal base cutoff). The radial/Laplace transform replaces the exploding occurrence-wise expansion of m_nu(X_shifted) and is validated exactly against it; integration uses integer-scaled coefficients. Values at k=46, eps=25/861, d=17: I_0=0.9999999999838952, J_0=3.9013805275618854, E_C=-0.005135550844843825, E_D=-0.0008455220431803321, E_E=-0.0010136558279124515, J_cap=3.5796230066288013, Delta=0.04905907220181656, I_cap=0.9509409277820786, c^T(M2^cap-(1/A)M1^cap)c = -0.10191368629446095 < 0. Controls: J_0 equals the engine c^T M2 c; inert caps give zero corrections; a real cap at k=4 matches brute force to 0.1%; occurrence path == Laplace path exactly on C,D,E; violation_F2 equals the exact I_0-I_cap from the independent capped_F2 instrument at k=4. Lean proves the exact sign.
