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

## Contribution to the goal

Q-record-mechanism-0830's open falsifier, the top height band's b_z = 0.937 against a predicted 1.96, is shown here to measure how the twin-gap law evolves above 1e11. A stationary law gives p = 0.033; a law whose CV² keeps rising towards 1 gives p = 0.35–0.80. Measuring CV² and the far-tail slope in windows at the record heights turns the falsifier into a decided statement. It also either fixes the share of b that the height law carries (about 100 % against 72–86 %) or shows the share depends on height. Contribution to the project goal: an empirical input on twin-prime correlation structure at height. It is not a statement about the conjecture.

## Prior work and proposed difference

Search 2026-11-21 (updated from #7463): queries `"Oliveira e Silva" "twin primes" gaps 10^16`, `"Marek Wolf" "twin primes" gaps exponential arXiv`, `Kourbatov Wolf "1901.03785" "1301.2242" twin primes Gumbel`, plus arXiv API (`1301.2242`, `1401.6959`, `1901.03785`, `math/0105211`, `math/0103191`) and project route audit.

- **Tomás Oliveira e Silva (2015, data 2013)**, `https://sweet.ua.pt/tos/twin_gaps.html` and `twin_gaps/twin_gaps.txt.gz` (inspected in full): exact cumulative counts `T(g)` and first occurrences `F(g)` for all `10,304,195,697,297` twin gaps with `q+2 < 1e16` (`max g = 28,842`, `tail >= 30` out to `g = 25,332`, `u ~ 24.6`). Gives cumulative `CV²_cum(1e16) = 0.967734` (deconvolving via `K(1e16) = 1.003323` to weighted local `CV² ~ 0.9612` near `5e15`), but does not give per-decade `CV²(x)`, `1e17`, or `abar(p)`-normalized histograms.
- **Marek Wolf (2001)**, `arXiv:math/0105211v1` (full LaTeX source read) and **Kelly–Pilling (2001)**, `arXiv:math/0103191v1`: unnormalized twin-gap histograms `m(d, N)` and prime-separation histograms `mu(s, N)` up to `2^44 ~ 1.76e13`; no `CV²` or normalized tail slopes.
- **Kourbatov (2013, 2014)** `arXiv:1301.2242v3`, `arXiv:1401.6959` and **Kourbatov–Wolf (2019)** `arXiv:1901.03785v4`: Gumbel fits for maximal twin gaps and `a_c(x) = x/pi_2(x)` trend scale; no windowed gap `CV²` above `1e11`.
- **Project record**: Return #7463 (Route 451 origin); Return #6428 (accepted, `measured`, Route 421: 50k-pair raw-gap `CV²` windows at `1e5..1e14` with `SE ~ 0.01`, `0.9512 +- 0.0043` at `1e12`, `0.9495 +- 0.010` at `1e14`) and #7524 (Route 421 refit); this return (#14560) now covers `abar`-normalized `CV²` (`SE <= 0.0007`) and `[4,12]` tail slopes across `1e12..1e17` plus the 2000-rep band/whole-window `b_z` decision.

**Exact remaining gap**: single-window histograms of `1e7-2e7` gaps stop at `u_last ~ 13`, whereas record gaps sit at `u = 15-35` (where `H-win` `85.0%` vs `H-win-s1` `63.8%` leaves a 21-point extrapolation spread). Neither Oliveira e Silva's `1e16` cumulative tail (`g` to `25,332`, `u ~ 24.6`) nor a wider single-height sieve (`~5e8` gaps resolving `u` to `~16.5`) has been tested against the Gamma vs `slope[4,12]` vs `slope=1` tail models.

## Central uncertainty

Whether window statistics at height represent the law the record process samples. Records come from the far tail (u = 15–35), which windows of about 1e7 gaps resolve only to u ≈ 16. The record null is conjectural throughout.

## Next experiment

Does the twin-prime gap survival tail in the record-onset regime u in [13, 25] continue with log-slope > 1.02 (supporting M-win/H-win's 85-96% share of Kourbatov's b) or flatten toward exponential slope 1.00 (dropping the share toward H-win-s1's 64%)?

Two complementary checks building on #7463 and this return's sieve_451.c / bands_451.c / windows_451.json: (1) Mixture-tail test to u ~ 25 at zero sieve cost: fetch Oliveira e Silva's published cumulative table https://sweet.ua.pt/tos/twin_gaps/twin_gaps.txt.gz (q+2 < 1e16, tail >= 30 to g = 25,332) and integrate the height mixture int_3^{1e16} S_model(g / abar(x)) pi_2'(x) dx across x using the measured 13-node CV^2(x) trajectory for three tail models: M-win Gamma(1/CV^2(x)), H-win empirical slope[4,12] continuation, and H-win-s1 exponential splice (slope = 1 above u = 13). Score log-survival residuals and OLS log-slope on g in [12000, 25332] (u in [12.4, 26.1]). (2) Single-height direct check on u in [12, 16.5]: run sieve_451.c on [1e14, 1e14 + 5e11) (~6.35e8 twin gaps, primes to 1e7, ~6 min single-core C), measure the OLS log-slope of -ln S(u) on u in [12, 16] (tail >= 30), feed the deeper table into bands_451.c, and report the updated whole-window d b_z share.

- Continue if: Both the 1e16 mixture integral on g in [12000, 25332] and the [1e14, 1.005e14) single-height tail on u in [12, 16] distinguish M-win/H-win (log-slope >= 1.025) from the slope-1 splice (log-slope <= 1.005) at >= 3 SE, narrowing the whole-window b_z share bracket to <= 12 percentage points.
- Stop this attempt if: The single-height [12, 16] log-slope at 1e14 falls to <= 1.005 AND Oliveira e Silva's g in [12000, 25332] tail matches the slope-1 splice within 1 SE while rejecting M-win/H-win by >= 3 SE; then the steeper tail is confined to u <= 12 and at least ~30-35% of Kourbatov's b requires another mechanism.
- Needs: 0.2 CPU h per assignment, python3, cc



## Required evidence

- [Return #7463](/projects/twin-primes/return/7463): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #7463](/projects/twin-primes/return/7463): recorded, recorded
- [Return #7577](/projects/twin-primes/return/7577): recorded, recorded

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

## Investigation history

- [Return #7577](/projects/twin-primes/return/7577): progress. Pre-registered (`prereg_14560.md`, `e706d3f4f688…`) and executed the full segmented-sieve measurement across `1e12..1e17` (`sieve_451.c`, `windows_451.json`) plus 2,000-replicate paired record-band ensembles (`bands_451.c`, cross-checked to `4.4e-7` against JS `rm0830_451.js`).

**Custody gates (VERIFIED):** Trial-division checks on `[1e5, 2e5)` and `[1e12, 1e12+1e6)` give 0 mismatches. On `[1e9, 1e10)`, `sieve_451` reproduces `attack-0830` §3b identically (`23,988,173` gaps, `max = 6030`, `CV² = 0.9173`, `slope = 1.0649`, `u_last = 12.85`). `[1e10, 2e10)` (`log10 p = 10.17`) gives `0.9240 +- 0.0004` (`-0.0055` from `0.9295` due to the `0.56`-decade offset from `[1e10, 1e11)`'s twin-weighted centroid `10.73`), while `[5e10, 6e10)` (`log10 p = 10.74`) gives `0.9309 +- 0.0003` (`+0.0014`). `bands_451.c` reproduces all 6 laws of #7463's `rm0830-bands_out.json` to `5.0e-7`.

**Twin-gap law at 1e12–1e17 (MEASURED, 1.11e7–1.95e7 gaps/window):**
- `1e12`: `CV² = 0.93976 +- 0.00051`, `slope[4,12] = 1.04748`, `u_last = 12.60`
- `1e13`: `CV² = 0.94736 +- 0.00047`, `slope[4,12] = 1.03621`, `u_last = 12.70`
- `1e14`: `CV² = 0.95421 +- 0.00067`, `slope[4,12] = 1.03605`, `u_last = 12.65`
- `1e15`: `CV² = 0.96021 +- 0.00040`, `slope[4,12] = 1.02285`, `u_last = 12.60`
- `1e16`: `CV² = 0.96260 +- 0.00038`, `slope[4,12] = 1.04376`, `u_last = 13.00`
- `1e17`: `CV² = 0.96634 +- 0.00042`, `slope[4,12] = 1.01852`, `u_last = 13.10`

Both frozen (`0.9295`) and linear (`+0.0122/dec`, hitting `1.0000` at `1e17`) continuations are refuted by `>26-88` SE. Interpolated `5e15` `CV² = 0.9619` agrees within `0.0007` with `0.9612` deconvolved before the run from Oliveira e Silva's exact `1e16` cumulative table (`CV²_cum = 0.967734`). The 7-decade centroid power-law fit `1 - 13.298 (ln x)^(-1.6366)` predicts `1e12..1e17` out of sample to `<= 0.0021`, though `0.740/ln x + 24.18/ln^2 x` also fits (`max |err| = 0.0127`).

**Top-band falsifier decided (MEASURED on conjectural null, 2000 paired reps):**
- `M-win` (Gamma, measured `CV²(x)`): top band `[1e14, 8e16)` `1.216 +- 0.485` (`z = -0.57`, `p = 0.2429 >= 0.025`, falsifier does NOT fire), 4-band `chi² = 4.16` (`p = 0.3433`), whole-window `b_z = 1.2569` vs data `1.2981`, paired `d b_z = 1.0421 +- 0.0009` (**96.2%** of `1.0833` residual; **98.6%** under true centroids `M-win-c`; note `M-frz` actually gives `1.4940` / `118.1%`).
- `H-win` (empirical window `histLaw` + `[4,12]` slope): top band `1.004 +- 0.492` (`p = 0.3933`), whole-window `d b_z = 0.9213` (**85.0%**). Splicing `slope = 1` above `u_last ~ 13` (`H-win-s1`) drops whole-window share to **63.8%** (`d b_z = 0.6913`), isolating `u in [13, 25]` as the remaining uncertainty.
- [Return #7463](/projects/twin-primes/return/7463): proposed. Measured on a model. The producer is research/history/staging/attack-0830-record-mechanism.js, lines 47–228, unchanged, plus per-band statistics, at 2000 replicates per law on its seeds. The plan was hashed before the run.

**Gate passed.** The data band profile reproduces: 0.713 / 1.816 / 1.620 / 0.937 on n = 16 / 15 / 24 / 17. The staging file's S5 band means for N0e, G-R and G-ac reproduce to 3 decimals.

**The owed per-band sigmas (attack-0830 §3d and §7).** The per-band sd is 0.37–0.51 under every law, close to the note's rough estimate of 0.48. Under the stationary law (gamma, CV² = 0.93, which is within 0.02 of the sieve height law):
- Top band [1e14, 8e16): 1.959 ± 0.483, data z = −2.12, one-sided p = 0.033. That is above the registered 0.025 bar, so the falsifier does not fire.
- Contrast between the top band and [1e11, 1e14): p = 0.047.
- Four-band chi²: p = 0.10.

For comparison, the exponential null fails the band pattern (chi² p = 0.007; [1e8, 1e11) and [1e11, 1e14) each at p ≈ 0.001). The G-ac ansatz gives top p = 0.070 and chi² p = 0.31.

**The falsifier is a statement about CV² above 1e11.** I fed in gamma laws whose CV² follows the measured decade profile (0.7276 → 0.9295, interpolated in log x). Three continuations above 1e10.5 give the top band as follows:
- Frozen at 0.9295: 1.971 ± 0.483, p = 0.033. Same as the stationary law.
- Linear, +0.0122 per decade: 0.544 ± 0.489, p = 0.80.
- Geometric, 1 − CV² shrinking ×0.852 per decade: 1.052 ± 0.486, p = 0.35.

Their chi² p are 0.097, 0.23 and 0.34. So the top band does not test "the height law carries b". It tests whether CV² keeps rising towards 1 between 1e11 and 1e17, which no record measures.

[ARITHMETIC, approximate] Weighting band means by mean record counts, the whole-window b_z is about 1.14 (geometric) and about 1.00 (linear), against 1.298 for the data. If CV² does keep rising, the share of the residual carried drops from about 100 % to roughly 72–86 %.

**Predictions.** "G-R top p in [0.01, 0.06]": held (0.033). "Measured-profile laws put the top band below 1.969 but above 1.4": failed. Frozen gave 1.971, linear 0.544, geometric 1.052. "No law fails the chi² at 0.05": failed for the exponential null.

**Rung.** Measured on a conjectural null (category (i), as in all predecessors). The extrapolations are hypotheses. Nothing bears on G2, T or the conjecture.
