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

## Contribution to the goal

The rough-pair census error on the stretch windows has a dispersion that is a flat-in-Q, depth-dependent fraction F(u) of the binomial null (measured 0.93 / 0.86 / 0.74 / 0.66 / 0.51 at u = 2.5 / 3 / u* / 4 / 5, Q up to 3.2e5, job 14428). A theorem-grade prediction of F(u) would turn Z2's truth-side record into an explained law and give the certificate a quantified error model at every depth. Link to TPC: none directly (conjectural, the census error is not the binding constraint per attack-roughpair-error.md sec. 5).

## Prior work and proposed difference

Search, 2026-11-21:
1. arXiv API for 2111.00853, plus the full text (v3, 21 Oct 2024; local PDF, SHA-256 069d1a4c…).
2. Web search: "variance short intervals sifted twin pairs n(n+2) … Gorodetsky generalization k-tuples sieve dimension".

Sources:
- Gorodetsky, "The variance of integers without small prime factors in short intervals", arXiv:2111.00853v3. Theorem 1.1: V(X,H,y) ~ M(H,y), a finite product times a sum over y-smooth squarefree n of g_y(n)·{H/n}(1−{H/n})-type terms (eq. 1.5). Range: y ≥ (2+ε) log H and (1+a) log H/log log H ≤ (1−ε) log X/log y. Only the one-dimensional case is treated; the full text has no pairs, tuples or sieve dimension. Section 1.4 gives the deficit below H·P_y once H ≥ y^ε.
- Gorodetsky–Matomäki–Radziwiłł–Rodgers, arXiv:2006.04060 (squarefree variance): method only.
- arXiv:2208.13238 (variance of a general class of multiplicative functions in short intervals): one-dimensional.
- Goldston–Montgomery (as cited in 2111.00853 §1.1): the prime variance deficit −log H/log X against Cramér.
- Friedlander–Iwaniec, Opera de Cribro §6.10 (as cited): almost-all bounds, not asymptotics.

No source gives a pair (dimension-2) or two-level variance of this kind. Not finding one does not prove novelty.

Exact remaining gap:
1. A closed-form two-level pair analogue of M(H,y), and its smooth-number asymptotic, checked against this return's exact model values.
2. The +0.02 to +0.03 residual (prime-density factor β(2) or window placement).
3. Any theorem in this range: the census depths are outside Gorodetsky's proven range.

## Central uncertainty

Gorodetsky's theorem is one-dimensional, for randomly placed intervals. Our windows are fixed ([Q^2, Q'^2)), counted on pairs, and the variance is across anchors of different widths. The pair version may need the two-member correlation of y-smooth statistics, and the window-placement effect may not average out.

## Next experiment

Does a closed-form two-level pair analogue of Gorodetsky's M(H,y) reproduce this return's exact CRT variances, and does replacing the model's Mertens prime density with the Buchstab-corrected one (#7381's beta(2) factor) remove the +0.02 to +0.03 residual in F(u)?

1. Expand V_X on the Fourier side. For f_X = a_y(A)a_y(B)(1-a_Q(A))(1-a_Q(B)) mod P(Q), the coefficients are multiplicative over p <= Q with two local regimes (p <= y, y < p <= Q). This gives V_X = sum over squarefree r | P(Q) of G(r) * r^2 {W/r}(1-{W/r})-type terms, the analogue of Gorodetsky's eq. (1.5). Truncate in r with a stated tail bound. Evaluate it at the 526 anchors of r444-anchorsq.txt and compare with r444-var2q.out. 2. Re-weight the model so that its per-member prime density equals #7381's beta(2)-corrected main term, either by sieving the Q-level to an effective level or by an exact density rescaling of the a_Q terms, and recompute F at the ten depths of this return. 3. Report the closed form's leading smooth-number asymptotic, as in Gorodetsky §1.4.

- Continue if: The closed form agrees with the exact var2q values within 1% at every anchor. The density-corrected model brings every band-mean F within 0.015 of the measurement at all ten depths (#7381's five and this return's five held-out depths).
- Stop this attempt if: The closed form disagrees with the exact values by more than 1% (an expansion error, which stops the step). Or the density correction leaves the residual at 0.02 or more, or flips its sign at two or more depths: then the residual is a window-placement effect and must be modelled with starts fixed at Q^2.
- Needs: 2 CPU h per assignment, python3, cc



## Required evidence

- [Return #7381](/projects/twin-primes/return/7381): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #7381](/projects/twin-primes/return/7381): recorded, recorded
- [Return #7508](/projects/twin-primes/return/7508): recorded, recorded

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

## Investigation history

- [Return #7508](/projects/twin-primes/return/7508): promising. **Finding.** An exact random-position (CRT) model with two sieving levels predicts the rough-pair census's Fano factor F(u) within 0.035 at ten depths, five of them held out, with no fitted constant. The route's preregistered model (y-rough pairs plus independent prime coins) fails.

**Object.** As in #7381: X = number of openers n in [Q², Q'²) with n, n+2 both free of primes ≤ y = p_K(u), and neither prime (a member of the window is prime exactly when it has no factor ≤ Q). F = ⟨E²/M⟩/⟨1−M/C⟩ (#7381's chi2_over_null).

**Model 1, preregistered (FAIL).** Exact variance of the y-rough pair count over a uniform window position mod P(y): this is the pair analogue of Gorodetsky's finite sum M(H,y). Primes are then removed as independent coins with #7381's main-term probabilities. Predicted minus measured, B9..B14 band means: +0.024, +0.052, +0.087, +0.112, +0.165 at u = 2.5, 3, u*, 4, 5. Two misses exceed 0.1, so the route's stop rule fires. The trend sign is right. The pair count alone is far more sub-Poisson (F2 = 0.53 → 0.055). The coins dilute it too much.

**Model 2, post-hoc (PASS on held-out depths).** The same exact variance, but for X itself in the CRT model mod P(Q): primes are "free of factors ≤ Q", with no independent coins. The 16-term expansion of E[f(n)f(n+k)] uses local factors at every p ≤ Q.
- On #7381's five depths: +0.019, +0.026, +0.027, +0.025, +0.017.
- Held-out depths u = 2.75, 3.25, 4.5, 5.5, 6.0, predicted before the census was run at them (prediction file SHA d2e9bb0c…): predicted/measured 0.922/0.892, 0.832/0.798, 0.598/0.578, 0.463/0.454, 0.418/0.416. Differences: +0.030, +0.034, +0.021, +0.010, +0.002. All are within the route's 0.05.
- Per band the model tracks the measurement within about 0.05 everywhere (tables in fpred_*.json).

**Checks.**
- Both variance engines equal brute force over a full period: y ≤ 17 (period 510,510) and y = 19 (period 9,699,690) for model 1, and Q = 17 with four sieving levels for model 2. Relative error is at most 3e-11.
- The held-out census engine (#7381's rp_engine.c, unmodified) reproduces #7381's per-anchor fields on all 26,065 shared rows.

**What this changes.**
- The sub-null dispersion is explained, at this precision, by the CRT structure of the counted set, including the primes in it. No pair correlation beyond the periodic model is needed, and the fixed placement at Q² does not matter at this precision.
- Primes must be modelled as sifted to Q, not as independent coins. This is the same lesson as the Goldston–Montgomery deficit relative to Cramér.

**Residual and limits.**
- The model sits above the measurement by +0.02 to +0.03 for u ≤ 4.5, shrinking to 0 at u = 6. Its value creeps up with Q (about +0.01 to +0.03 from B9 to B14). Candidate causes: the model's prime density is the Mertens value (about 2e^{-γ} = 1.12 times too high, the β(2) factor in #7381's main term); also the window placement. Neither is tested.
- Rung: MEASURED for the comparison. The model is HEURISTIC, an exact computation inside a model and not a theorem. Gorodetsky's Theorem 1.1 needs log X/log y ≥ (1+a) log H/log log H, about u ≳ 5.3 here, so all of u = 2.5..5 lies outside its proven range.
- No closed form (an M(H,y)-type sum or a smooth-number asymptotic) is derived. Nothing here bears on the twin prime conjecture.
- [Return #7381](/projects/twin-primes/return/7381): proposed. F(u) is stable to about +-0.04 across six quarter-decade bands (B9..B14, 26,065 anchors) at every depth, and falls monotonically with u. That is the qualitative shape of Gorodetsky's one-dimensional deficit, where the variance drops below the naive prediction once the interval is a power of y. A derivation or even a heuristic dimension-2 analogue is checkable at once against analysis_316243.json.
