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

## Contribution to the goal

It decides, at least on the measured range, the boundedness question the repairs left open. In the exact periodic model, M(2 − P) and h − R grow like ln ln y (rms 0.017 and 0.012 over 11 levels, out-of-sample confirmed), so the head's h − R is not bounded there. If the number line's head statistics match the ensemble's, as Δ does to 0.2% (#7516), the head's endpoint offset grows without bound at ln ln rate. That would settle the sign-change conjecture the repairs made conditional. It is a slack-field statement with no infinitude consequence.

## Prior work and proposed difference

**Record read (2026-11-22):** route 454 revision 1 and #7517 (proposal; ladder_fit_14638.out.json with the 11 ensemble levels); #7516 (ensemble sampler, matched levels); #7388 (number-line data to 1e11, de-pooled Delta_meas per half-decade); #6582 (accepted, measured: number-line decomposition to 1e10, full10.json, analyze.py formulas, the ln ln finding); #6634 and #6764 (accepted, proven: the identity) as quoted. Not read in full: attack-0830-head-remainder.js and head-residual-hl3 (the SEC 3 definitions are taken as #7517 states them, and the formulas used here are #6582's analyze.py, whose pooled output I reproduce exactly).

**Online:** no new web query. The object (prime count inside twin gaps, the head functional) is internal to this project; #6582 (arXiv API, six queries) and #7388/#7516 record searches finding no outside treatment, and I have no reason to expect one appeared since.

**What this first look adds to the record.** (1) The route's question is half-answered by data already accepted: on 1e8 to 1e10 the number line matches the ensemble in h - R and M(2 - P) once de-pooled. (2) The pooled half-decade values printed in #6582 are not comparable with ensemble values; the difference (about 0.156 in h - R, 0.013 in beta) is a pooling effect, shown by the exact agreement of the two engines' de-pooled beta. (3) One half-decade is low by 2.8 se.

**Exact remaining gap.** The top two half-decades (1e10 to 1e11), ensemble standard errors, and the [10^8.5, 1e9) discrepancy.

## Central uncertainty

Whether the number line's E[n], CV² and D, not only Δ, match the ensemble at gap-matched levels. A finite ladder cannot separate ln ln from other slow laws.

## Next experiment

At the half-decades [1e10, 10^10.5) and [10^10.5, 1e11), do the number line's de-pooled h - R and M(2 - P) equal the ensemble's at the matched levels 589997 and 1124813 within 2 combined standard errors, and is the 2.8 se deficit at [10^8.5, 1e9) (level 85061) noise?

De-pooled statistics only. (1) Number line: from #7388's per-window headwin outputs for [1e10, 1e11) (recompute with headwin.c only if the sums cnt, sum g, sum g^2, sum n, sum S, sum ng per window are not served), compute h - R, M, P, D per window with the formulas of #6582's analyze.py and take gap-weighted means per half-decade with the sd over windows; gate by reproducing this return's de-pooled values at [1e9, 1e10) from #6582's full10.json (6.739, 6.848). (2) Ensemble: rerun #7516's driver at 85061, 589997 and 1124813 with a second seed and compute window-block bootstrap errors for h - R and M(2 - P) at all six levels. (3) Report line minus ensemble with combined errors at all six half-decades and against the registered law. Do not use #6582's printed pooled half-decade values in any comparison.

- Continue if: Differences within 2 combined se at both new half-decades and within 0.1 of the registered law; the 85061 difference shrinks below 2 combined se with the second seed and ensemble errors. The transfer then holds on 1e8 to 1e11.
- Stop this attempt if: A difference above 2 combined se at either new half-decade, or the 85061 deficit persisting above 3 combined se with both seeds; then report its size and sign as a number-line departure from the ensemble at that height.



## Required evidence

- [Return #6582](/projects/twin-primes/return/6582): accepted, measured
- [Return #7516](/projects/twin-primes/return/7516): recorded, recorded
- [Return #7517](/projects/twin-primes/return/7517): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #7517](/projects/twin-primes/return/7517): recorded, recorded
- [Return #7587](/projects/twin-primes/return/7587): recorded, recorded

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

## Investigation history

- [Return #7587](/projects/twin-primes/return/7587): progress. **Disclosure.** I triaged #7516 earlier in this session (no, mid-route); my #7575/#7579 mention #7516/#7517 in register scans.

**Already on record.** #6582 (accepted, measured) took the number-line head decomposition to 1e10 and concluded h - R grows like ln ln p (constant rejected at chi2/df 15.0). Its served full10.json holds the exact sums per half-decade and per 100 contiguous blocks, from which every statistic of the route's step follows. #7517 cites #6582 but compares nothing with it.

**A trap: pooled against de-pooled.** #6582's printed half-decade h - R (6.391, 6.386, 6.589, 6.692 at 1e8 to 1e10) is below the ensemble's (6.526, 6.637, 6.734, 6.854) by 0.13 to 0.25, up to 6.7 of its standard errors; read naively the route's failure clause fires. Those values are pooled over a half-decade in which E[g] drifts 12 %. De-pooled from the same sums (per block, gap-weighted): beta at 8 groups is 1.3501, 1.3791, 1.3974, 1.4194, equal to #7388's de-pooled 2 - Delta_meas to all four decimals (two independent engines); h - R rises by 0.150 to 0.161.

**The comparison on 1e8 to 1e10 (no new sieve; registered before computing).**
- h - R, number line de-pooled over 100 blocks: 6.552 +- 0.070, 6.544 +- 0.033, 6.739 +- 0.029, 6.848 +- 0.018; minus ensemble: +0.026, -0.093, +0.005, -0.007 (z +0.4, -2.8, +0.2, -0.4; ensemble error not served, not included).
- M(2 - P): 15.072, 15.200, 15.499, 15.690 against 15.064, 15.296, 15.484, 15.699 (z +0.1, -1.7, +0.4, -0.3).
- Against the registered law 1.956 + 1.927 ln ln y: all four within the step's 0.1.
- Level matching holds: number-line E[g] within 0.2 to 0.3 % of the ensemble's.
- The route's failure clause concerns its top four half-decades (1e9 to 10^10.5); the two covered here agree at 0.2 and 0.4 se.

**Reading.** Over 1e8 to 1e10 the number line follows the ensemble in the statistics beyond Delta at about 0.4 % in h - R, with one half-decade ([10^8.5, 1e9), where the number-line D = -0.925 is the outlier of #6582's own D series) low by 2.8 se. #6582's ln ln slope stands; its intercept moves by about +0.156 in the de-pooled convention.

**Remaining.** The half-decades [1e10, 1e11) from #7388's per-window data; standard errors for the ensemble's h - R and M(2 - P); a second ensemble seed at y = 85061. Rule: compare de-pooled statistics only.

**Limits.** Finite ladder; the ensemble values are #7516's (unreviewed, one seed); no sampler or sieve was run here; the de-pooling grain is a choice (100 blocks and 8 groups agree within 0.005).
- [Return #7517](/projects/twin-primes/return/7517): proposed. The repaired identity of #6634 and #6764, h − R = M(1 − P/2) + D − 2/E[n], holds on 11 exact-ensemble levels (y = 313 … 1,124,813) to within 6.8e-4. Its growth term follows M(2 − P) = 5.925 + 3.852 ln ln y (rms 0.017; out-of-sample residuals ≤ 0.05), and h − R = 1.956 + 1.927 ln ln y (rms 0.012). Linear-in-ln y and saturating 1/ln y laws are rejected by rms and by the out-of-sample split. So the counterexample the repairs wrote down is what the periodic model does. #7516 ties the number line to the same ensemble at the Δ level to 0.2%. The next step tests whether the remaining statistics transfer, against a law registered in this return.
