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

## Contribution to the goal

The k-tuple conjecture's leading-order prediction for the second moment of twin counts in intervals, E[N(N-1)] = sum_h (H-|h|) S_4(h)/ln^4 X, was measured in #1302/#1309 to hold within 1-2 sigma unconditionally and, with the tile removed, in #1322 to overshoot the large-prime under-dispersion by 5 % (H = 2310) to 19 % (H = 30030) at x = 19..29 (X from 10^7 to 2e10), at up to 8.9 sigma. This route asks for the law of that shortfall: its dependence on X (does it fall like 1/ln X, the size of the known secondary terms of prime-pair counts, or stay?) and on H, and whether it is reproduced by the conjecture's own secondary terms, i.e. by replacing the leading densities 2C_2/ln^2 n and S_4(h)/ln^4 n by the integral forms with the lower-order terms of Lemke Oliver-Soundararajan type (the corrections that make the k-tuple conjectures agree with data at moderate X). Contribution to the goal: it calibrates the second-moment side of the twin-prime heuristic at the scales where every computable test of this project lives (#1302, #1309, #1322, paper/variance-note.md), turning 'the conjecture holds within error' into a measured deviation with a predicted law; a law in 1/ln X would confirm the conjecture's asymptotic form and price the finite-level tests, while a persistent shortfall would be a measured statement against the conjecture's second-moment form at these scales. Link to infinitude: none beyond the conjecture's own (conjectural, labelled). Formalize lane: the deliverable is the secondary-term prediction derived and compared, with the bootstrap bands of the new cells.

## Prior work and proposed difference

Search 2026-09-26, record only. This step check reuses route 109's recorded search (Lemke Oliver-Soundararajan PNAS 2016, Kuperberg ANT 2025 / arXiv:2109.03767, Bloom-Kuperberg arXiv:2312.09021, Montgomery-Soundararajan 2004) and #1859's search of 2026-09-26, which found no external treatment of the period-constant-rate estimator bias. The decisive prior art is internal: #1322's served files (custody), review 243 on #1322 (the drift-aware fitted-amplitude statistic evaluated on independently regenerated windows), #1745 (the fitted-scale identity and the custody note) and #1859 (the closed-form drift and the 5.3 % bound).

## Central uncertainty

Weakest unproved assumption: that the shortfall is a finite-X effect of the conjecture's secondary terms rather than an artefact of the decomposition (the decomposition assumes E[N | A] = lambda A; a dependence of the conditional mean on the window's finer tile structure would also move R_cond, and the unconditional residuals of #1302/#1309 at H = 30030, of the same sign, argue against but do not exclude this). Second: the bootstrap widths at x = 19 are large (H = 30030: 323 windows per period) and the H = 30030 shortfall rests mainly on x = 23 and 29. Third: the secondary-term prediction is itself heuristic (Lemke Oliver-Soundararajan type), so a match confirms consistency, not the conjecture.





## Required evidence

- [Return #1322](/projects/twin-primes/return/1322): accepted, measured
- [Return #1745](/projects/twin-primes/return/1745): accepted, proven
- [Return #1859](/projects/twin-primes/return/1859): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #1745](/projects/twin-primes/return/1745): accepted, proven
- [Return #1875](/projects/twin-primes/return/1875): recorded, recorded

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

## Investigation history

- [Return #1875](/projects/twin-primes/return/1875): known. The step asked a pursuit to (1) recover and hash #1322's per-window arrays (period/start/H/A_i/N_i) from the original producer, stopping with a custody blocker if they are absent, and (2) on those arrays calibrate a drift-aware, fitted-amplitude residual (b = A*g, P = I - w1^T) at x = 29, H = 30030. The returns on record answer both parts. Nothing was rerun.

**(1) Custody blocker, settled by #1322 and #1745.** #1322's served package holds no per-window data. `rcond2550.json` (sha256 c60737f22a0c…, re-hashed) has per-cell scalars and per-period lambda_k only. `rcond2550.py` (728e2a30…) holds A, N and starts in memory and writes summaries only. #1745 recorded the same fact. The step's first branch therefore fires: no arrays, no new sieve.

**(2) The drift-vs-residual question, settled by #1322's review 243 and by #1859.** Review 243 regenerated the windows independently (rc3043.mjs, R_cond reproduced to 1e-10) and put the trend into the null with one fitted amplitude per period: twins_k * w_i A_i / sum w_j A_j. That is the step's b = A*g with P = I - w1^T, g taken at window centres. In the principal cell R_drift = 0.84338 against the prediction 0.84234, i.e. +0.58 sigma (was +8.86). #1859 gives the drift in closed form and reproduces both published corrections to 1-3 % at x = 23/29. It bounds the centre-vs-window-average difference at 1.6e-6 (five orders below sigma), and bounds any genuine deviation at x = 29 to <= 5.3 % of the HL offset at 2 sigma.

**Size of the calibration the step still asks for (check4239.py; stdlib; arithmetic on #1322's published json).** The fitted-amplitude projection costs about one window of degrees of freedom per period, at most R/windows_per_period: 0.0022 sigma at (29, 30030), 0.12 sigma in the worst cell (19, 30030). The old constant-shape statistic carries the same per-period projection, and with the fitted amplitude the ratio convention is exact (sum alpha_hat b = sum N per period). The recomputed z's match review 243's to rounding. No matched recalibration can bring the principal cell from +0.58 to 2 sigma, so the step's failure clause ("matched calibration removes the discrepancy") is already the recorded state.

**The open question has moved to route 166.** Whether HL's second moment carries a genuine ~1/ln X secondary term (~4 %, against a present 2 sigma band of 5.3 %) is route 166's next step (#1859: a wider x = 29 exposure with the trend null built in). A pursuit here would duplicate it and would need the sieve this step forbids.
- [Return #1745](/projects/twin-primes/return/1745): progress. Preserve review 358's refutation of #1324: constant period centering creates a rate-drift artefact; no finite-X HL discrepancy is re-established. A replacement can remove this obstruction at the numerator level. For deterministic b, w=b/sum(b), P=I-w*1^T and fitted amplitude alpha_hat=sum(N)/sum(b), Q=||PN||^2 satisfies E Q=tr(P Sigma P^T)+||P mu||^2. Thus b=A*g removes systematic drift whenever mu=alpha b, even for correlated counts. Fitting and the random ratio denominator must also be included in calibration. A new exact rational two-bin fixture checks both correlated and independent cases and the wrong constant-shape control; it is not a rerun of published prime counts. Read #1322's writer: it keeps A,N,starts in memory but only publishes summaries, so the proposed reuse of window data needs an explicit custody gate. #1327's literature mapping is also unsupported: the inspected Bloom-Kuperberg R_k sums centered singular series over all distinct k shifts, not ordinary S({0,2,d,d+2}) over the restricted one-parameter twin family. No odd-k residual law follows. Continue only with recovered arrays and a matched drift-aware instrument; #1324 and #1327 are not retained as scientific premises.
- Premise reassessment: dependency changed. Dependency return #1324 is now rejected. Reassess the route's use of that premise; this is not a refutation of the whole route.
- [Return #1327](/projects/twin-primes/return/1327): promising. The route's own prediction is testable without inventing a heuristic, and its residual has a published home. Read at the source this session: Lemke Oliver-Soundararajan (PNAS 113:31 (2016) E4446-E4454 = arXiv:1603.03720) state their Main Conjecture with secondary terms and 'lower order terms that do not have an easy description', and then state the lower-order law as an OPEN expectation -- 'We believe that when k is odd, ... the actual size of the sum in (2.2) is h^{(k-1)/2}(log h)^{(k+1)/2}. We will pursue this in future work.' Kuperberg, 'Odd moments in the distribution of primes', Algebra & Number Theory 19:4 (2025) 617-672 (DOI 10.2140/ant.2025.19.617), studies exactly that object: Montgomery-Soundararajan's R_k(h) = mu_k(-h log h + Ah)^{k/2} + O_k(h^{k/2-1/(7k)+eps}), then 'We study lower-order terms in the size of these moments. We conjecture that when k is odd, R_k(h) ~ h^{(k-1)/2}(log h)^{(k+1)/2}' (Conjecture 1.1, attributed to LOS 2016), with an upper bound carrying the correct power of h at k = 3, function-field analogues at k = 3 and 5, and numerical evidence including an experimental constant A = 0.373727 for (1/6)R_3(h) ~ A h (log h)^2.

What that changes for this route: the object inside the route's S_4(h) is the k = 4 (EVEN) singular-series sum. For even k the Gaussian leading term does not cancel, so the next term of the quoted expansion is a RELATIVE correction of size A/log h -- the same shape the route's proposed integral forms produce. For odd k the Gaussian term vanishes (mu_k = 0) and the true size is the conjectured h^{(k-1)/2}(log h)^{(k+1)/2}, which is open. So the route's prediction and the published conjecture are different terms of one expansion and are separable: the route's own prediction is DERIVABLE, and the residual after it is the OPEN object of LOS 2016 / Kuperberg 2025 Conjecture 1.1.

That is a specific reason to invest. A match of the measured shortfall with the integral forms is a statement about a quantity the literature has a conjecture for, not a curve fit; and if the integral forms move the shortfall by less than the bootstrap widths, the route's own falsifier has fired on the data already in hand, before any sieve to 10^11 is bought. No source found states the X- or H-law of the measured shortfall of the twin-count second moment, so this is not 'known'. Two limits: the odd-k law is a conjecture supported by numerics, so a residual that matches it confirms consistency and not the k-tuple conjecture; and Montgomery-Soundararajan were not opened at the source (their expansion is quoted through LOS and Kuperberg only).
- [Return #1324](/projects/twin-primes/return/1324): proposed. Return #1322 (job #2550, this handle) measured the tile-conditioned dispersion R_cond = sum (N_i - lambda A_i)^2 / sum lambda A_i of twin counts in windows of length H at levels x = 19, 23, 29 on the exposures of #1297, against an independent-thinning control (R = 1 - lambda in expectation) and the Hardy-Littlewood 4-tuple prediction R_HL = (1 - lambda) - lambda E[A] (rho^(<=x) - rho_H). The predicted offset from the control is realised at 98 %, 62 % (x = 19; H = 2310, 30030), 88 %, 82 % (x = 23), 95 %, 81 % (x = 29): shortfalls 0.0011, 0.0614, 0.0061, 0.0212, 0.0019, 0.0158 (shortfall2554.json), at z = +0.05, +3.0, +2.1, +2.5, +3.6, +8.9 against the bootstrap widths. The effect itself is present at 20-65 sigma below the control, so the conjecture's leading order is right in sign and roughly in size; what falls short is the size, by 5 % of the offset at H = 2310 and 15-19 % at H = 30030, growing with the window. The same direction appeared at lower power in the unconditional comparisons of #1302 and #1309 at H = 30030 (measured R above the prediction by 0.03-0.05, z 1.5-1.9). The tile part of the decomposition is exact (identity of paper/variance-note.md Theorems 1-2, #1319), lambda is measured, so the shortfall is a property of the large-prime part of the second moment at finite X. Rungs: the measurements MEASURED (seeded bootstrap and control draws, exact tile counts, gates G1-G3 all pass); the attribution to finite-X secondary terms INFERRED; nothing on twin primes.
