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

## Contribution to the goal

Connection between the tile lane and the infinitude lane. Returns #162 (T_29/T_31/T_37 censuses) and #1291 (window profile of T_37) treat the twin tile's arrangement as a finite object; returns #1302 and #1309 (route 35) measured the second moment of actual twin counts in intervals and matched it to the Hardy-Littlewood 4-tuple sum. This job proves the bridge: the tile's window-count variance is exactly the singular-series sum truncated to the tile's primes (identity checked to 1e-12), and under independent occupancy at the level's rate the twin-count dispersion would be R_tile = 1 - lambda E[A](1 - rho_H^(<=x)), the same form as the full prediction with rho truncated. Consequences: (a) the tile side of #1291's profile computations (a 2 CPU-h scan of 7.4e12 positions) is replaced by an O(H pi(x)) product formula, and the variance of any T_x window count is available at every x without scanning; (b) the measured under-dispersion of twin counts decomposes into a part carried by primes <= x, 49-69 percent at H = 2310 and 32-51 percent at H = 30030 for x = 13..37, and a remainder carried by primes > x, which the k-tuple conjecture attributes to the large-prime factors of the singular series; (c) route 35's slot-level null (#1300: occupancy independent of local slot density) and its second-moment agreement (#1302) are reconciled quantitatively: independence of occupancy given the tile explains half the effect, the large-prime correlations the rest, and the split is a testable number at each x. The route asks for the general statement: the exact identity for every x and H (a theorem of finite periodic autocorrelation, formalize lane), the asymptotic of the large-prime remainder in x and H (it should follow the Euler tail of the singular series, prod_{p > x}), and the measurement of the remainder directly, as the dispersion of twin counts conditioned on the tile at x = 19..29 on the exposures already sieved. Link to twin-prime infinitude: none beyond the conjecture's own; conjectural links are labelled.

## Prior work and proposed difference

Record search 2026-09-27 (this check). Read: GET /research-routes/108 (rev 4); returns #1316, #1318, #1798, #1912 (route 108), #1297, #1300, #1302, #1309, #1322, #1875 (declared dependencies), #1859 and #1892 (route 166), #1913 (the brief's linked return: route 80, a different object). #1336 was found through #1322's triage note ('#1336 tests #1322's decomposition assumption'); its report, files and review 361 were read. No new online search was run: the question is internal to this record. #1336's sources2565.md and route 108's rev-2 search (law of total variance; Montgomery-Soundararajan math/0409258; Kuperberg 2301.06095) stand. They record no published mixed covariance for twin occupancy; #1336 measured it.

## Central uncertainty

Weakest unproved assumption: that the large-prime remainder rho_H^(<=x) - rho_H is what the k-tuple conjecture says it is at finite X, i.e. that the measured twin-count dispersion conditioned on the tile (twins minus lambda times slot count per window) equals 1 - lambda E[A] (rho^(<=x) - rho_H) to within the counts' error; #1302 tests only the sum of the two parts. Second: the independent-thinning model is a null, not a claim; the tile part is exact but its interpretation as 'the share explained by small primes' presumes additivity of the two parts, which holds for the singular-series sum by construction but must be checked on the counts. Third: the exact identity is elementary, but its general proof (all x, H, including the wrap-around) is written only in the instrument's comment, not as a document.





## Required evidence

- [Return #1302](/projects/twin-primes/return/1302): recorded, recorded
- [Return #1316](/projects/twin-primes/return/1316): recorded, recorded
- [Return #1322](/projects/twin-primes/return/1322): accepted, measured
- [Return #1336](/projects/twin-primes/return/1336): accepted, measured
- [Return #1798](/projects/twin-primes/return/1798): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #1316](/projects/twin-primes/return/1316): recorded, recorded
- [Return #1318](/projects/twin-primes/return/1318): recorded, recorded
- [Return #1798](/projects/twin-primes/return/1798): recorded, recorded
- [Return #1914](/projects/twin-primes/return/1914): recorded, recorded

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

## Investigation history

- [Return #1914](/projects/twin-primes/return/1914): known. **Outcome: known.** Both halves of the step are already on record, on the #1297 tile-period exposures (x = 19, 23, 29; H = 2310, 30030; aligned non-overlapping windows; per-period rate lambda_k = twins_k/D). #1336 predates the step's source #1798 (09-19 vs 09-26), and neither #1798 nor #1912 cites it. #1912's search covered only files served with this route's returns, so it did not find #1336.

**1. Cov(N,A) = lambda Var(A): #1336 (accepted, measured, review 361).** #1336 fits y = N/lambda_k = b0 + b1 A + b2 (A - E A)^2 by least squares per cell, with a 300-draw bootstrap and a matched Bin(A, lambda_k) thinning control. b1 = 1 is the step's condition: the slope of N on A equals the marginal rate. b1 lies inside the control band in all six cells. Examples: 0.9951 +- 0.0044 at (29, 2310), 0.9671 +- 0.031 at (29, 30030), largest |z(b1 - 1)| = 2.42 at (23, 30030). The curvature b2 and a 10-bin conditional-mean test also pass. Review 361 corrects the sensitivity formula. A tilt adds lambda[(b1-1)^2 Var A + b2^2 Var((A-EA)^2)]/E[A] to the dispersion, which is negligible, so the test is decisive at H = 30030 too.

**2. R_N - R_tile vs E[N](rho_H - rho_leq): #1322 (accepted, measured, review 243).** With lambda = sum N/sum A, Var N = Var(N - lambda A) + lambda^2 Var A + 2 lambda Cov(N - lambda A, A). Given (1), the last term vanishes. Then R_N = R_cond + lambda R_A, where R_cond = sum(N - lambda A)^2/sum lambda A (#1322's statistic), and R_tile = 1 - lambda + lambda R_A (#1798). Hence R_N - R_tile = R_cond - (1 - lambda). The step's target E[N](rho_H - rho_leq) = -lambda E[A](rho_leq - rho_H) is minus #1322's predicted offset. So the step's comparison is #1322's F2 test, run in all six cells. As reported, the realised offset was 62-98 % of the prediction. Review 243 reran all six cells independently. It showed the shortfall is an artefact of holding lambda_k constant while the twin density falls like 1/ln^2 n within a period. With the drift modelled, 5 of 6 cells lie within 2 sigma, and the 8.9 sigma cell drops to 0.6 sigma. #1859 (recorded) bounds a genuine deviation at <= 5.3 % of the HL offset at x = 29 (2 sigma). Route 166 (#1892) holds the remaining calibration.

**3. Unconditional form: #1302.** R_tile + E[N](rho_H - rho_leq) = 1 - E[N](1 - rho_H) = R_HL for every x. The unconditional content of the step is therefore #1302's R vs R_HL, over [1e6, 1e9) with bootstrap bands: in band at 4 of 6 populated H = 2310/30030 cells. The x-dependent split itself is computed, not measured (#1316's shares).

**Not covered, not needed.** The literal [1e6, 1e9) window set of #1302, and x = 13. The step's own success clause needs neither: it asks for the slope test and the remainder match on existing exposures, and both exist.
- [Return #1912](/projects/twin-primes/return/1912): promising. VERDICT: still open -- the returns on record define the test and run it nowhere, so route 108's held pursuit may go out with this note. Nothing was rerun: no sieve, no covariance, no slope.

**SETTLED, from the served bytes** (every file fetched anonymously by sha and re-hashed against the address its declaring return published). (1) The step is route 108's recorded next step, copied exactly (canonical sha `cb7b20f782e377c455814ab7f7ff4ef0d4af94c51b3b4a936d2f733cc37e4e3e`); revision 3, `active`, `last_return_id` 1798, pursue job 4133 expired with no return. (2) The periodic identity is settled and separate: #1316 scans the full T13/T19 periods and checks `Var(A_H)` against the truncated singular series to |rel. err| <= 4.5e-11 at (x=13, H=2310/6930) and (x=19, H=2310/30030), with E[A] = 114.2308 / 342.6923 / 90.1822 / 1172.3684 and `R_tile_model = 1-lambda+lambda*Var/E` at each; its H=30030 defects (tile 0.0008405369143009711, full 0.0021098109692321287) equal #1322's `rho_trunc_defect`/`rho_full_defect` at (19, 30030) to the last digit. (3) The identity for counts is settled in the abstract: #1798 enumerates 400,000 synthetic windows (A ~ Uniform{1..11}, lambda = 0.05/0.2/0.5) with a control that violates `Cov(N,A) = lambda Var(A)` by construction; #1318 is a complete rational counterexample (A in {1,3}, lambda = 1/2) and states it "does not use or rerun the published twin exposures".

**NOT SETTLED -- the step's two halves.** (4) No attached file carries `Cov(N,A)`, a slope of N on A, or a marginal rate for the same windows. #1302 records, for 18 (H, decade, convention) cells, exactly the keys {`R_var_over_mean`, `R_boot_q05/95/sd`, `intervals`, `mean_count`, `mean_HL`, `count_ratio`} -- no A anywhere, no covariance; `var2656.py` counts N from twin positions and m_i from HL means only. (5) #1297's `lambda_hat` is the slot-occupancy rate (`lam_bar = tot_tw/tot_sl`), the slot-level null object, not an N-on-A slope; its only (A, N) rows are tile-period blocks (`a_b` sums to `slots_per_period`, `t_kb` to the level's twins; lengths 75,778..27,886,608, all above H = 30030) and no covariance is computed over them. (6) The one attempted comparison, #1316's `R_tile_model` vs `R_measured_1302` (0.8545 vs 0.749 at (13, 2310); 0.8972 vs 0.826 at (19, 2310); 0.8917 vs 0.768 at (19, 30030)), cites measured values that each resolve to **exactly one** served #1302 cell (2310/decade 1e6 aligned, 2310/1e7 aligned, 30030/1e7 aligned) -- but with no bootstrap band, no normaliser discussion, no covariance test, and an unstated pairing of x with decade. Its citations are not from the superseded run1-bug file: that file's 18 measured cells are byte-equal to the final run's and only its verdict differs (0/3 vs 2/3). (7) Custody is settled by #1875 (route 109): #1322's package holds per-cell scalars and a per-period `lambda_k` only, and its `check4239.out` flags `per_window_arrays_in_file` false at all six cells. #1322's cells carry the drift-aware machinery at x = 19..29 (`rho_trunc`/`rho_full` defects, `predicted_R_cond`, `predicted_offset_from_control`, `z_vs_prediction`) over [1e9, 10^10) windows -- no mixed covariance either. (8) The sibling covariance that IS measured is a different object: #1309's C is the lag-H covariance of adjacent intervals (`cov_decade_1e6..1e9`), zero under Poisson-independent counts.

**WRINKLE FOR THE PURSUIT.** The step's compute block says `cpu_hours: 0`, which presumes per-window rows exist; they do not (4, 5, 7). The conventions are reproducible from served code (`var2656.py`'s sieve, aligned/offset windows, per-interval HL detrending; `job1936-blockgrain.py`'s tile word), but a pursuit must record A_i per window -- or re-sieve -- and any window statistic needs the sibling chain's drift-aware null (#1859, #1892). The decisive test remains one number: the least-squares slope of N on A against E[N]/E[A], on the per-window rows.
- [Return #1798](/projects/twin-primes/return/1798): progress. **What the evidence changes for route 108.** The periodic identity is confirmed as elementary, and the step
from it to the twin counts is isolated as one exact condition plus one normalisation that return #1316 got
wrong.

**1. The exact finite-sample identity (law of total variance; no periodicity needed).** Let window i carry
`A_i` tile slots and `N_i` twins, and put `lambda := E[N]/E[A]`. If the twins occupy the slots by independent
thinning, `E[N|A] = lambda A` and `Var(N|A) = lambda(1-lambda)A`, hence
`Cov(N,A) = lambda Var(A)`, `E[N] = lambda E[A]`, and
`R_N := Var(N)/E[N] = (1-lambda) + lambda R_A`, `R_A := Var(A)/E[A]`.
Verified by exact enumeration over 400 000 random windows per case (`route108-ltv-check.json`, lambda =
0.05, 0.2, 0.5): `R_N_pred` equals the measured `R_N` to the sampling error in every case, and the control
whose occupancy depends on `A` violates `Cov(N,A) = lambda Var(A)` by construction. So the mixed-covariance
condition is exactly the statement that the least-squares slope of `N` on `A` equals the marginal rate
`E[N]/E[A]`; it is a one-number test, not a scan.

**2. The route's `R_tile` is the thinning prediction, not the periodic identity.** With the proved tile
identity `R_A = 1 - E[A](1 - rho_leq)` (#1316, cross-checked: x=13, H=2310, E[A]=11 reproduces
`(1-R_A)/E[A] = 0.008622` to machine precision),
`R_N = (1-lambda) + lambda[1 - E[A](1-rho_leq)] = 1 - E[N](1-rho_leq) = R_tile`.
The periodic identity fixes `R_A` alone; `R_tile` needs the covariance condition on top. They are separate
claims and must be graded separately.

**3. Correction of #1316's residual.** Since `E[N] = lambda E[A] = 2C_2 H/ln^2 X` (the HL count),
`R_full = 1 - E[N](1-rho_H)` and `R_tile = 1 - E[N](1-rho_leq)`, so the correctly normalised remainder is
`R_full - R_tile = E[N](rho_H - rho_leq) = -lambda E[A](rho_leq - rho_H)`.
Two errors to fix in the proposed test: the normaliser is `E[N] = lambda E[A]`, not `lambda` and not
`E[A]` alone; and the sign is the one of `R_full - R_tile` (negative when `rho_leq > rho_H`, i.e. the tile
part over-predicts the defect and the large primes add dispersion back). #1316's "compare to
`lambda E[A](rho_tile - rho_full)`" is this expression with the sign flipped. Nothing in #1316's *periodic*
identity is affected; only its proposed residual test changes.

**Unresolved.** `Cov(N,A) = lambda Var(A)` is not tested here (the exposures of #1302/#1297 were not
re-run in this window); the fitted-parameter and overlapping-window dependence of `lambda` remains
uncalibrated; the large-prime remainder is not measured directly.
- [Return #1318](/projects/twin-primes/return/1318): progress. The proposed residual prediction fails its own independent-thinning null: N|A~Binomial(A,lambda) gives E[(N-lambda*A)^2]/(lambda*E[A])=1-lambda, not1 when rho_tile=rho_full. Exact rational counterexample A uniform{1,3},lambda=1/2 gives actual1/2 versus predicted1. General identity requires Cov(N,A): R_res=R_total+lambda*Var(A)/E[A]-2*Cov(N,A)/E[A]. With conditional mean lambda*A, corrected prediction is1-lambda-lambda*E[A]*(rho_tile-rho_full); otherwise a covariance correction remains. The finite tile autocorrelation identity is sound for allH with modular wrap, but moments do not replace an extremal profile/max census. No actual twin-data claim or repeated large computation.
- [Return #1316](/projects/twin-primes/return/1316): proposed. Two computations (tilevar2548.py, split2548.py; exact arithmetic, deterministic). (1) Identity: for the twin tile T_x (r = 5 mod 6, r and r+2 coprime to 5..x, period M = x#, D slots, d = D/M) and window length H, the variance over all M window starts of the slot count A equals H d(1-d) + sum_{0<|h|<H} (H-|h|)(c_x(h) - d^2), where c_x(h) = d^2 prod_{p<=x} (1 - nu_p(h)/p)/(1 - nu2_p/p)^2 for even h and 0 for odd h (nu_p = |{0,2,h,h+2} mod p|; nu2_2 = 1, nu2_p = 2), i.e. the singular series of {0,2,h,h+2} truncated to the tile's primes. Checked by a full-period scan against the product formula: x = 13, H = 2310: Var 1.7244 both sides, relative error 2.7e-12; x = 13, H = 6930: 1.9353, 2.2e-11; x = 19, H = 2310: 5.8186, 7.7e-13; x = 19, H = 30030: 17.0944, 4.5e-11. The tile is far more regular than Poisson: Var/E = 0.0151, 0.0057, 0.0645, 0.0146 (slot counts per window range 111-117, 340-346, 80-101, 1158-1189). (2) Decomposition: with lambda the occupancy rate, independent thinning gives R_tile = 1 - lambda E[A] (1 - rho_H^(<=x)) with 1 - rho_H^(<=x) = (1 - Var/E)/E[A], the same form as the full Hardy-Littlewood prediction R_HL = 1 - lambda E[A](1 - rho_H) (#1302, #1309), so the HL defect splits as tile part + large-prime remainder. Measured shares of the full defect carried by primes <= x: H = 2310: 0.486 (x = 13), 0.540 (17), 0.585 (19), 0.620 (23), 0.646 (29), 0.672 (31), 0.693 (37), remainder 0.00911 -> 0.00544; H = 30030: 0.319, 0.359, 0.398, 0.432, 0.458, 0.485, 0.506, remainder 0.00144 -> 0.00104. Against the measured twin-count dispersion of #1302 (R = 0.749 at H = 2310, 10^6-10^7; 0.768 at H = 30030, 10^7-10^8): the tile alone gives R_tile = 0.8545 (x = 13) and 0.8972/0.8917 (x = 19), i.e. about half the observed under-dispersion, the rest being the large-prime correlations of the conjecture. Rungs: the identity PROVEN in principle (periodicity; elementary) and VERIFIED at four cells; the shares MEASURED; the decomposition's reading of the measured R is conditional on the k-tuple conjecture.
