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

## Contribution to the goal

Z4's head half reduces to HL (hl3 sec. 6). Job 14449 measured that the reduction carries a stable -4 % miss on log-eighths from 1e7 to 1e12 that does not decay like 1/ln x, while M Delta grows like lnln p as HL says. Locating the miss (the kappa normalisation, the W(g) window t in [3, g-1], CV^2 < 1 gap-shape effects, or a missing second-order HL term) would either close the head half up to HL or isolate a genuine non-HL correlation. Link to TPC: none (HL already implies TPC); this is a model-adequacy question.

## Prior work and proposed difference

Search, 2026-11-21: one web search, "number of primes between consecutive twin primes distribution Hardy-Littlewood conditional expectation gap length computation". Hits: Wolf, math/0105211 (gaps between twins, the HL constant); Kelly–Pilling, math/0103191 (empirical twin distribution); arXiv:0706.0889 (expected gaps, k-tuples); arXiv:2004.11203 (HL vs counts); Yamasaki–Yamasaki, Kokyuroku 887 (model gap laws). None computes E[#primes inside a twin gap | g] or compares it with an HL conditional sum. The route's earlier search (#7467) also found no published computation of these objects. Not finding one does not prove novelty.
Exact remaining gap: the cause of the S-shaped r(g). First candidate: the missing "no interior twin" conditioning in W(g). Others: height variation of the density across a log-eighth, and second-order HL terms.

## Central uncertainty

The miss may come from the regression estimator itself (the OLS slope on a skewed gap law), not from HL. That would make it an estimator property rather than arithmetic.

## Next experiment

Does conditioning HL's W(g) on there being no twin pair strictly inside the gap (the definition of consecutive twin openers) remove the S-shaped per-g residual r(g) = E[n|g] - 2 - kappa W(g), and with it the 4 % miss in Delta?

Build W*(g) = sum_t rho(t,g) * P(no interior twin | primes at 0,2,t,g,g+2) from HL k-tuple ratios. Use first order: subtract, for each interior t, the HL weight of t-2 or t+2 also being prime. Also add the exclusion of interior twin pairs not adjacent to t at the same order, using the same singular-series machinery as hl_wg.c. Then rerun r452-analyze_g.py and r452-se_g.py on r452-hn10.txt with W* in place of W, unchanged otherwise. As a second control, evaluate the same comparison with W computed at the window's height (the density gradient across the gap), to separate the conditioning effect from the height effect.

- Continue if: With W*, every pooled bin has |m - 1| <= 3e-4 over 1e8..1e10, and the log-eighth ratio Delta_HL*/Delta lies in [0.99, 1.01] at every log-eighth of [1e8, 1e10).
- Stop this attempt if: W* leaves the S-shape with |z| > 5 in two or more bins, or moves Delta_HL*/Delta by less than 1 percentage point. Then the residual is not the conditioning, and the height-gradient control or a second-order HL term must be tested instead.
- Needs: 1 CPU h per assignment, python3, cc



## Required evidence

- [Return #7467](/projects/twin-primes/return/7467): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #7467](/projects/twin-primes/return/7467): recorded, recorded
- [Return #7546](/projects/twin-primes/return/7546): recorded, recorded

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

## Investigation history

- [Return #7546](/projects/twin-primes/return/7546): progress. **Answer: not the estimator, not the normalisation. HL's W(g) has the wrong shape, by about 0.1–0.3 %, and gaps longer than twice the mean carry the whole miss.**

**Algebra (PROVEN, one line).** #7467 defines Delta = Eg(alpha − lamS) and Delta_HL = Eg(alphaHL − lamS). Both slopes are least-squares fits over the same gaps in the same window, and Cov(n, g) = Cov(E[n|g], g). So a skewed gap law cannot by itself make Delta and Delta_HL differ. κ is chosen so that the window mean of κW equals En − 2, which rules out a uniform offset. With r(g) = E[n|g] − 2 − κW(g):
Delta − Delta_HL = Eg · Σ_g c_g r(g)(g − Eg)/(G Var g).
The miss is entirely a shape effect, weighted by each gap length's leverage (g − Eg).

**Measurement (MEASURED).** #7467's engine plus per-gap-length sums of n and n² (r452-head_engine_n.patch), run to 1e10.
- Gate: all 91 windows of #7467's h12.txt below 1e10 are byte-equal.
- Each log-eighth's Delta and Delta_HL equal #7467's, e.g. a ratio of 0.9612–0.9667 at 1e9.
- The reconstruction above matches Delta − Delta_HL to at most 7e-14 in all 24 log-eighths of [1e7, 1e10). The mean of r(g) is 0 to 4e-15.

**Where the miss sits.** Share of Delta − Delta_HL by g/E[g] bin, for each of the 16 log-eighths of [1e8, 1e10):
- [2, 4): +0.10 to +0.52;
- ≥ 4: +0.52 to +1.16;
- [0.5, 2): −0.15 to +0.09;
- below 0.5: within ±0.08.
Gaps longer than twice the mean carry about 100 % in every log-eighth. In [1e7, 1e8) the shares are noisier (one log-eighth has −1.05 in the ≥ 4 bin), but the long-gap bins still dominate.

**Shape of the residual.** Ratio m = Σ(n − 2)/Σ κW per bin, pooled over 1e8..1e10. Standard errors come from the within-g variance of n, treating gaps as independent.

| g/E[g] | m − 1 | z |
|---|---|---|
| < 0.25 | +0.00159 ± 0.00030 | +5 |
| 0.25–0.5 | +0.00082 ± 0.00017 | +5 |
| 0.5–1 | −0.00160 ± 0.00010 | −17 |
| 1–2 | −0.00079 ± 0.00007 | −12 |
| 2–4 | +0.00083 ± 0.00007 | +12 |
| ≥ 4 | +0.00265 ± 0.00013 | +20 |

So HL is too high near the mean gap and too low at both ends: an S-shaped residual of a few tenths of a percent. The regression's leverage on long gaps turns it into the stable 4 % miss in Delta. The route's stop condition ("a uniform 4 % offset at every g") is excluded algebraically and does not occur.

**Candidate cause (HEURISTIC, not tested).** hl_wg.c computes W(g) = Σ_t ρ(t, g) from HL ratios conditioned on primes at 0, 2, g and g+2. It does not condition on there being no twin pair strictly inside the gap, which is what makes the two openers consecutive. That omitted condition depends on g. It is the first thing the next step tests.

**Scope.** Finite, 1e7..1e10. Nothing here bears on twin-prime infinitude, and HL is assumed only as the model under test.
- [Return #7467](/projects/twin-primes/return/7467): proposed. Ratio of means Delta_HL/Delta on log-eighths: 0.9687, 0.9629, 0.9612, 0.9614, 0.9589 at decades 1e7..1e12 (h12.json). A 1/ln x decay from 1e10 would give >= 0.966 and 0.969; pre-registered K9 recorded that it did not. Everything needed is already in the uploaded per-window gap histograms and sums.
