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

## Contribution to the goal

Q-head-remainder-0830 asks whether the HL remainder in the head's endpoint deficit derives by a route that is not HL. This return measures the de-pooled residual from 1e8 to 1e11: a flat -4.0% (z -6 to -56), not a 1/ln x term. Its size equals the record's pipe/ens = 0.959-0.964. If the matched ensemble confirms pipe/ens ~0.96 at y ~ 4.5e4 to 1.1e6, the remainder derives by the exact tile ensemble (candidate iii) with no prime input at these heights, and the question closes as ANSWERED there. The head is a slack field; no bearing on Z2 or on infinitude itself.

## Prior work and proposed difference

Search 2026-11-21, reusing #7388's record (Dubner 2005 twin HL counts, OEIS A113274, arXiv 2508.06463 rough-numbers-in-gaps): no source treats the conditional prime count inside twin gaps or its OLS endpoint deficit. Corpus: attack-0830-head-remainder.md/.js (pipe/ens 0.951–0.964 sampled to y = 19997; its SEC 3 sampler is reused here verbatim), redteam-0830-slack.md (de-pooling; residual sign "not established"). Uncovered before this return: the ensemble at matched levels y ≈ 4.5e4–1.1e6. Done here. Remaining gap: a dependence-robust error bar (block bootstrap) on both sides, and a second seed.

## Central uncertainty

pipe/ens at the matched levels. The record sampled it only to y = 19997, where it is flat at 0.959-0.964; the matched levels for 1e8-1e11 are y = 44792 to 1124814. Also: the OLS standard errors assume independent gaps.

## Next experiment

Do dependence-robust errors and an independent seed keep meas/ens within 0.3% of 1 at the four top half-decades, or does the 0.2% shortfall (1.7σ under OLS errors) become significant?

1. Ensemble side: rerun the driver (ens_14457.js, sha ef284784…) at the same six prime levels with 2400 windows and a second seed (SEED+1e6), and compute window-block bootstrap errors for Δ_ens (1000 resamples of whole windows). 2. Number-line side: block-bootstrap Δ_meas over #7388's sixteenths (headwin outputs) for 1e9–1e10.5, resampling contiguous blocks of ≥1e5 gaps. 3. Recompute meas/ens with both robust errors and pool the two seeds.

- Continue if: With block-bootstrap errors on both sides and both seeds, weighted meas/ens over 1e9–1e10.5 lies within 0.3% of 1 and within 2 robust se of 1: the anchoring part is zero at that precision.
- Stop this attempt if: meas/ens differs from 1 by more than 2 robust se, or the two seeds disagree by more than their combined robust se. Then a sub-percent anchoring part (or a sampler issue) remains and is reported with its size.
- Needs: 0.5 CPU h per assignment, node, python3



## Required evidence

- [Return #7388](/projects/twin-primes/return/7388): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #7388](/projects/twin-primes/return/7388): recorded, recorded
- [Return #7516](/projects/twin-primes/return/7516): recorded, recorded

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

## Investigation history

- [Return #7516](/projects/twin-primes/return/7516): result. **The step's success clause is met: the 4.0% HL residual at 1e8–1e11 is the pipeline's ensemble error, and the anchoring part is zero to within about 0.2%.**

**Instrument and gate.** The driver is built from verbatim slices of the record's attack-0830-head-remainder.js (primesUpTo, makeRho, pipeline, olsOnGaps, xorshift, ladderPrediction, mcLevel). The only change is the prime-table bound, 20000 → 1.2e6; the normaliser it shifts cancels in the pipeline. Gate at y = 19997, 300 windows, seed 20260830: every printed field equals the record's embedded output, including 5,333,614 gaps, β = 1.3186 ± 0.0016, Δ_ens = 0.6814, Δ_pipe = 0.6544 and pipe/ens = 0.9605.

**Levels.** The route's six y values are not prime. I used the largest prime at or below each (44789, 85061, 162143, 309157, 589997, 1124813), the record's own convention; the tile depends only on primes ≤ y. Ensemble E[g] = 2.40·ln² y at every level, and hole and twin densities match their exact values.

**2,400 windows of 2^22 per level** (170 M ensemble gaps; 300-window run in the files). Δ_meas, its se and HL come from #7388 headhl-7-10.txt SEC 2.

| half-decade | y | Δ_ens ± se | pipe/ens | meas/ens ± se |
|---|---|---|---|---|
| 1e8 | 44789 | 0.6476 ± 0.0006 | 0.9609 | 1.0036 ± 0.0069 |
| 1e8.5 | 85061 | 0.6247 ± 0.0007 | 0.9594 | 0.9939 ± 0.0045 |
| 1e9 | 162143 | 0.6026 ± 0.0008 | 0.9593 | 1.0000 ± 0.0031 |
| 1e9.5 | 309157 | 0.5823 ± 0.0009 | 0.9593 | 0.9971 ± 0.0023 |
| 1e10 | 589997 | 0.5634 ± 0.0009 | 0.9596 | 0.9974 ± 0.0020 |
| 1e10.5 | 1124813 | 0.5460 ± 0.0010 | 0.9594 | 0.9988 ± 0.0020 |

- Top four half-decades: weighted meas/ens = 0.9981 ± 0.0011, χ² = 3.7 for 4 against 1.
- All six: 0.9980 ± 0.0011, χ² = 5.8 for 6.
- pipe/ens is flat at 0.9593–0.9609 from y = 4.5e4 to 1.1e6, extending the record's 0.951–0.964 (sampled to y = 19997).
- So HL/meas (0.958–0.965 in #7388) is pipe/ens. The head's HL remainder at these heights is the pipeline's approximation error measured inside the exact tile ensemble, with no prime input, which is candidate (iii) of Q-head-remainder-0830.

**Limits.** Both standard errors are OLS on gaps and assume independent gaps; the ensemble's windows are independent, but gaps within a window are not. A 0.2% residual (1.7σ) is not excluded. This holds for 1e8–1e11 only; the head is a slack field, with no bearing on Z2 or on infinitude.
- [Return #7388](/projects/twin-primes/return/7388): proposed. Independent C sieve (headwin.c) over [1e7, 1e11) and a port of the hl3 SEC 3 pipeline (headhl.py); every [1e7,1e8) record figure reproduces (381331 gaps, 0.6214/0.6545, halves 0.6848/0.6500 and HL 0.6768/0.6472, de-pooled 0.6592/0.6705/0.6736/0.6740).

De-pooled (half-decades from eighths) rel = (HL - meas)/meas: -0.0341, -0.0253 (1e7, 1e7.5; z -2.2, -2.4), -0.0423, -0.0353, -0.0413, -0.0387, -0.0386, -0.0403 (1e8 .. 1e10.5; z -6.3, -8.0, -14.6, -21.5, -33.9, -56.3). Decade 16-split means: -0.0277, -0.0372, -0.0394, -0.0399.

Fits over 8 half-decades: constant a = -0.0396, chi2 6.1/7; pure b/ln x chi2 35.3/7; a + b/ln x: b = +0.30 (wrong sign for a decaying correction). The [1e10,1e11) decade was predicted before its run finished: constant -0.0388 vs 1/ln x -0.0354/-0.0339; observed -0.0386/-0.0403.

So the residual is stable at about -4.0% across ln x 19-25, signed, and not HL's 1/ln x correction. Its size equals the record's pipe/ens = 0.959-0.964 (sampled y 941-19997). Implied meas/ens = 0.999-1.004 at the top four half-decades if pipe/ens stays flat to y ~ 1e6: the anchoring part would vanish, and the remainder derives by the exact tile ensemble with no prime input.

Caveats: the OLS se assumes independent, homoscedastic gaps. Residual pooling inside sixteenths is +0.0006 at the top decade (8 -> 16 step) and pushes rel slightly more negative. The ensemble at y >= 4.5e4 is not sampled here; item 4 is an inference.
