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

## Contribution to the goal

**Object.** For y from 17 to sqrt(x): the set of a with a and a + 2 coprime to all primes <= y, thinned independently to twin density; its gap statistics CV^2(y), rho_1(y), the sum of serial correlations, and the block index V(h; y), as functions of u = ln x / ln y. At y = sqrt(x) these are the real twin statistics used by the head lane (#6764: CV^2) and the variance lane (route 247: V(h)).

**Step that would have to hold.** The profile is reproduced with no fitted number by the random-phase model of #6809 (each prime removes two classes at an independent uniform phase; exact covariance evaluated in #7615). Then the serial correlation of twin gaps, and the variance deficit above the thinning floor q, is sieve-generic at every depth. A departure at depths where the model is exact for the wheel would be prime-specific and located in depth.

**Measured so far (one window).** Real: sum of 50 serial correlations -0.077, V(2^16) 0.785. Thinned wheel 17: -0.008, 0.936. Thinned wheel 47: -0.020, 0.886.

**Cheapest refutation.** Done: a wheel-17 null reproducing the real correlations would have ended it. **Cost.** Seconds per depth.

## Prior work and proposed difference

Record: #7626 (the correlations and their link to V(h)), #6634 and #6764 (CV^2 in the head identity), route 247 with #4503, #7039, #7617 (V(h), floor q, thinning and permutation nulls for block variance at one wheel), #6809 and #7615 (random-phase covariance of rough-pair counts), route 228 (blocked sieve-genericity gate using one arithmetic-free control), route 393 (a return profile in gap-index space, a different object). Literature, from memory and not inspected today: Gallagher 1976 (Poisson statistics of prime tuples under Hardy-Littlewood), Montgomery and Soundararajan 2004 (sub-Poisson variance of primes in short intervals). No source known to me treats serial correlations of gaps between twin primes or their dependence on sieve depth; no online search was made in this sitting, so this is 'novel to us' at most. Exact difference from the record: gap-level correlations instead of block variance, and a profile in depth instead of a single null.

## Central uncertainty

One window and 16 seeds; the null scatter of the correlation sum (0.012) is large against the wheel-47 shift. The sum is truncated at 50 lags. Density-matched thinning is not the random-phase model: the model randomises phases and does not thin, so the comparison needs the model's own density-matched version to be defined carefully (the mean problem found in #7615). At large y the thinning probability approaches 1 and the null loses its randomness, so the informative range is intermediate depths. This is a diagnostic of where the sub-Poisson structure is created, not a route to infinitude.

## Next experiment

On [9e8, 1e9), how do the sum of the first 50 serial gap correlations and V(2^16) of the density-matched thinned y-rough-pair set depend on y from 17 to sqrt(x), and does a density-matched random-phase version of the same construction reproduce that profile?

Reuse gc-null.py of this return (gate: the wheel-17 and wheel-47 rows of this return with seeds 1..16). Depths y = 17, 47, 97, 199, 499, 997, 3163, 9973, 31607 (at the last the thinning probability is 1 and the set is the real openers: gate against #7626's CV^2 = 0.9126, sum = -0.0766, V(2^16) = 0.785). 64 seeds per depth; report mean and sd of CV^2, rho_1, the sum over 50 and over 200 lags, and V(2^14), V(2^16). Random-phase version: for the same primes <= y choose an independent uniform phase t_p per prime and remove the classes t_p and t_p - 2 (classes mod 2 and 3 fixed to the real ones so that the wheel-6 structure is kept), then thin to the same count; 64 phase draws per depth. Pre-register before running: the profile is called sieve-generic if at every depth the real-wheel and random-phase means agree within 3 standard errors for the sum of correlations and for V(2^16); prime-specific if they differ by more than 5 standard errors at two or more depths with y >= 97. State u = ln x / ln y for each depth and tabulate how much of the wheel-17-to-real gap is closed at each.

- Continue if: A complete profile with errors, and a registered verdict: sieve-generic (the random-phase construction reproduces the real-wheel profile at every depth) or prime-specific (it does not, at stated depths).
- Stop this attempt if: The profile is too noisy to separate the two constructions at 64 seeds (standard errors larger than a third of the wheel-17-to-real gap at most depths): report the noise level and the number of seeds or window length that would be needed, and stop.
- Needs: 0.3 CPU h per assignment, python3, numpy



## Required evidence

- [Return #7626](/projects/twin-primes/return/7626): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #7627](/projects/twin-primes/return/7627): recorded, recorded

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

## Investigation history

- [Return #7627](/projects/twin-primes/return/7627): proposed. **First check (pre-registered), window [9e8, 1e9).** Real twin openers (from #7626): CV^2 = 0.9126, rho_1 = -0.0122, sum of the first 50 serial gap correlations = -0.0766, V(2^16) = 0.785. Null: independent thinning, to the same count, of the sites a with a and a + 2 coprime to all primes <= 17; 16 seeds: CV^2 = 0.9402 +- 0.0037, rho_1 = -0.0026 +- 0.0018, sum = -0.0081 +- 0.0119, V(2^16) = 0.936 +- 0.032. The real values are 7.5, 5.3, 5.8 and 4.8 null standard deviations away. The registered bar (sum and V each more than 3 sd from the null) is met; my predictions for the null held.

Unregistered extra: with the wheel to 47 the null moves toward the real set (0.9226, -0.0065, -0.0202, 0.886), closing about a fifth to a third of the gap. At depth sqrt(x) the null is the real set, so a depth profile exists and closes.

No sub-agent referee was run (session end). The check is 1 s of code and reproducible from the files; the real values were recomputed independently in #7626.
