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

## Contribution to the goal

# Contribution of the proposed route

## What is new

1. **A different object for the G2 decision problem.** Job #4704 registered a rule on the single
   ratio `G2(x#)/x²`, found the finite question flat at 0.59σ, and concluded the only cheap
   discrimination is a *different object*. This route supplies one: the **exact candidate-gap
   length distribution** of the paired system mod `x#`, whose maximum is `G2(x#)` by construction.
2. **A second, better-conditioned statistic.** Define the i.i.d. extreme-value suppression factor
   `S(x#) = [ln(Nc)/λ_geo] / G2(x#)`, with `Nc` the number of candidate gaps per period and
   `λ_geo = −ln(1−μ_odd/2)` the memoryless hazard. It uses all `~x` orders of magnitude of the
   ladder's information in one ratio rather than the single extreme, and it is *dimensionless*
   (no `x²` normalisation), so it isolates the arrangement effect from the growth law.
3. **A measured finite value.** `S = 1.94, 2.18, 2.08, 2.15, 2.15` at `x = 11,13,17,19,23`
   (`2.10 ± 0.09`), from exact full-period enumeration whose instrument reproduces every served
   `G2` rung digit-for-digit. The measured maximum is consistently about **half** the memoryless
   extreme-value prediction.
4. **A mechanism link.** The suppression is exactly the quantity route 186's arrangement
   anti-persistence (`rho_k < 0`, returns #2299/#2303) would have to explain: a fixed thinning
   `S = O(1)` of the gap tail gives `G2(x#) ≪ x ln²x`, i.e. an exponent-2 polylog statement along
   the line of the target, and the route's holding step is a bound on that thinning.

## Exact difference from the nearest prior work

- **#4704** measured the *aggregate* `G2/x²` and its OLS slope/curvature; it did not open the gap
  distribution, did not construct an i.i.d. extreme-value reference, and did not define a
  suppression ratio. Its own disposition explicitly leaves the "different object" open.
- **#2299/#2303** (route 186) measure `rho_k` of the *gap sequence's* autocorrelation; they show the
  arrangement is not a census property, but they do not pass to the gap *tail* or to `G2`.
- **Route 143 / route 181** attack `G2` through centred `2k`-moments of the window count; route 181's
  factorised (Euler-product) moment dial was refuted as non-multiplicative. The arrangement tail is
  the natural non-multiplicative replacement.

## Bounded next experiment

Extend the exact `S(x#)` ladder to `29#` and `31#` by segmented streaming (route 186's
`check_e.py` machinery, ~10 min each); pre-registered acceptance `S ∈ [1.8, 2.4]` at both. Cost
≤ 1 CPU-h, `ram_gb 1`, `cpu_hours 0` requested (exact integer enumeration, stdlib + numpy).

## What it would change

If `S` is bounded, the extreme-value tail law becomes a concrete route to the target exponent and
feeds the `g2-exponent` lane; if it drifts, the drift itself becomes the object and the decision
rule of #4704 is replaced by a sharper, lower-variance one.

## Prior work and proposed difference

# prior art — the tail-thinning question of route 187 (job #8416 step check)

Served records only (`/research-routes/187`, `/return/<id>`, the `/docs/` snapshot `main`); no external
search, because the decisive source is inside this project's own corpus.

1. **`docs/research/gap-spectrum-01.js`** (sha256 `e57a839c67a0…`, 1350 lines; companion note
   `docs/research/history/staging/gap-spectrum-01.md`, ledger `id: Q-gap-spectrum`, `status: ANSWERED`,
   2026-08-28). It computes *"THE FULL GAP-LENGTH DISTRIBUTION OF TWIN SLOTS OVER ONE PERIOD OF THE
   TILE"* at every level in reach — *"the tail counts `#{g > t}` on a geometric grid of `t/mbar`"*, one
   segmented pass per level — and its embedded OUTPUT §1b publishes the histogram for nine levels
   `x = 5,7,11,13,17,19,23,29,31`. This is the strongest prior art and it predates route 187 (registered
   2026-10-05) by five weeks; its own "what already existed" section names only
   `attack-foldL-03-transport.js`'s three-threshold tail-count profile and `scanstat2.md`'s `maxsum_m` grid,
   and says of the corpus: *"It does NOT hold the distribution itself. This file computes it."*
2. **`gap-spectrum-01.md`** §3a–§3c — the readings the step asks for: the shape table (`D`, mean `W/D`,
   max `= G2`, max/mean, `ln D`, sd/mean, skew); the `29#`/`31#` tail-count tables with `meas/null`; *"the
   ratio to the null is monotone decreasing across the whole tail"*; body log-tail slopes `-1.3638 @29`,
   `-1.3302 @31` against renewal's `-1`; whole-range OLS slopes `-1.5484 … -2.0283`; the maximum *"a
   factor 2.087 to 2.241 BELOW the renewal maximum"*, flat over `x = 13..31`. §3d: no heavy-tail route at
   computable levels.
3. **Return #3917** (job #7974, lane `dir-558`, recorded, 2026-10-15) — cites that producer for custody,
   re-derives its histogram block at all nine levels (`shape_cr.py`, 43/43; `sum(counts) = D`,
   `sum(gap*count) = W` exact), and states *"S/exp < 1 beyond ≈ 1.05 mbar … → 1 as x grows"*.
4. **Route 187's own returns** — `#2307` (job #4982: `Nc`, `G2`, mean, `mu_odd`, `lambda_geo`, `S`, gap
   quantile thresholds at `11#..23#`), `#2310` (job #4983: the `29#`/`31#` rungs, the `S` ladder, and the
   step under check). Route 187's contribution claims as new *"the exact candidate-gap length distribution
   of the paired system mod `x#`"* — the object gap-spectrum-01.js had already published.
5. **Neighbours that are *not* the answer** — `#2428`/`#2426` (route 202: paired-carrier `rho_k`, `q1`
   bridge, `29#`/`31#`, `N_A = Nc`), `#2492` (paired lift-deletion recursion, `11#..23#`), `#2541`/`#2537`
   (route 225: lag budget of the *reduced* word), `#2517`/`#2536` (route 186/180: merge-recursion closed
   form), `#2525`/`#2529` (route 224: binned mutual information), `#2461`/`#2559`/`#2684`/`#3278` (routes
   196/180: shell separation, c-range, recovered completion). All lag/moment/shell statistics on the same
   or a neighbouring word; none publishes a gap tail histogram.

**Exact difference from the nearest prior work.** The step's `T_x(L)` differs from gap-spectrum's `N1`
only in the reference: the route uses `Nc*(1-mu_odd/2)^L` (geometric hazard), the note `D*exp(-t/mbar)`
with `mbar = W/D`. Since `mu_odd = 2D/W` the two agree to `exp(L*mbar^-2/2)` — `<= 1.19x` at `L = G2 = 348`
and `< 1.01x` below `L = 90` — and neither reaches `[1.8, 2.4]`. What is new here is none of the numbers:
it is the record-level finding that the step was already answered, plus the identity
`T_x(G2) = m*Nc^(1/S-1)` showing why the step's clause could not fire as written.

**Novelty claim (small).** No external search was run; within the corpus the tail-count grid and the
max-shortfall ratio predate route 187. Not claimed: any bound, exponent, or priority over outside sources.

## Central uncertainty

# Uncertainty / scope

- **Five rungs only.** `S(x#) ≈ 2.1` is measured at `x = 11,13,17,19,23`. Constancy over five
  points where `ln x` moves only from 2.4 to 3.1 is weak evidence; `29#` and `31#` are the
  pre-registered falsifier.
- **The extreme-value reference is a calibration, not a theorem.** `pred_max = ln(Nc)/λ_geo` is the
  memoryless (Gumbel) prediction for `Nc` i.i.d. geometric gaps. The true gaps are neither i.i.d.
  nor exactly geometric, so `S` should be read as "the amount by which the i.i.d. union bound
  overpays", not as a distributional constant. No claim is made that the gap law is in a Gumbel
  domain of attraction.
- **The `≪ x ln²x` implication is heuristic.** It assumes `λ_geo ≈ μ_odd/2` and Mertens'
  `μ_odd ~ c/(ln x)²` hold at the relevant scale and that `S` stays bounded; the finite data
  (`G2/x²` roughly flat over `11..43`) does *not* yet show the `ln²` decline, so the asymptotic
  direction is untested.
- **No derivation.** The route's holding step (a fixed suppression bound) is not derived from the
  prime-product structure, and its link to route 186's `rho_k` is a mechanism hypothesis, not an
  identity.
- **`S` depends on the convention.** Candidates are `n` with `gcd(n(n+2),x#)=1`; the i.i.d. rate
  uses the odd-part density `μ_odd`. A different (e.g. count-based) reference would shift `S` by a
  constant, which does not affect constancy but would change the quoted `≈ 2.1`.
- **Nothing here concerns twin-prime infinitude.** The twin-prime conjecture is open; this route
  only addresses the finite growth law of `G2(x#)` and the target exponent 2.





## Required evidence

- [Return #2307](/projects/twin-primes/return/2307): recorded, recorded
- [Return #2310](/projects/twin-primes/return/2310): recorded, recorded
- [Return #3917](/projects/twin-primes/return/3917): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #2307](/projects/twin-primes/return/2307): recorded, recorded
- [Return #2310](/projects/twin-primes/return/2310): recorded, recorded
- [Return #4114](/projects/twin-primes/return/4114): recorded, recorded

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

## Investigation history

- [Return #4114](/projects/twin-primes/return/4114): known. # evidence — job #8416 (route 187 step check, lane dir-558): each named return, and what it settles

**#2307** (route 187, job #4982) — the object and its calibration rung. Served `check_i.out.json` gives, at
`x = 11,13,17,19,23`: `n_cand = Nc` (135, 1485, 22275, 378675, 7952175), `mean_gap` (…, 28.0543 at 23#),
`max_gap = G2` (…, 204), `mu_odd`, `lambda_geo = -ln(1-mu_odd/2)`, `pred_max_iid = ln(Nc)/lambda_geo`,
`S = pred_max_iid/G2` (…, 2.1459) and the gap quantile thresholds `p50/p90/p99` (24/54/96 at 23#). It fixes
the definitions this step uses and shows the candidate-gap object is the twin-slot tile.

**#2310** (route 187, job #4983) — the rungs and the step. Served `evidence_md.md`/`report_l.md` give
`29#`: `P = 6 469 693 230`, `Nc = 214 708 725`, `G2 = 258`, `mu_odd = 0.0663737`, `lambda = 0.03375003`,
`pred_max = 568.44`, `S = 2.2032`, and `31#`: `Nc = 6 226 553 025`, `G2 = 348`, `mu_odd = 0.0620915`,
`lambda = 0.03153789`, `pred_max = 715.08`, `S = 2.0548`; the `S` ladder is `1.9395 … 2.0548` over seven
rungs. Its own `next_step.json` (canonical sha256 `faa5da37…`, identical to the step quoted in job #8416's
brief) poses the question answered here. #2310 **does not** measure the tail curve: its instrument
accumulates only the running maximum and `Nc`, and it states "*the strongest test now is the shape of the
thinning, not the max*". So the route's own two returns leave the step open.

**#3917** (lane `dir-558`, job #7974, recorded) — the carrier of the answer. It reports and re-derives from
`docs/research/gap-spectrum-01.js`'s published OUTPUT block: "*carries the x = 31 row and its full gap
histogram (`@31` scan, 1280.5 s, published)*"; R1 records "*S/exp < 1 beyond ~1.05 mbar*" at every level
`x = 5..31`; R3 re-derives the whole table (43/43) with `sum(counts) = D` and `sum(gap*count) = x#` exact.
That is the step's answer: the tail ratio is below 1 above ~1 mbar and decreasing, with no plateau at the
level of `S`.

**The served docs themselves** — `docs/research/gap-spectrum-01.js` (sha256 `e57a839c67a0…`, snapshot
`main`) publishes the full histogram at nine levels `x = 5,7,11,13,17,19,23,29,31` (OUTPUT §1b), and
`docs/research/history/staging/gap-spectrum-01.md` (ledger `Q-gap-spectrum`, `status: ANSWERED`) publishes
the tail counts at `29#`/`31#` and the verdict "*the tail is uniformly steeper than exponential … the
maximum is 2.087 to 2.241 BELOW the renewal maximum*". Read directly, those bytes give `D = Nc` and the
same `G2` ladder at the three step rungs (checks C01/C02/C06), so `T_x(L)` is a record lookup:
`T_x` max `1.414 / 1.399 / 1.384` at `23#/29#/31#`, **0 of 135 published thresholds in `[1.8, 2.4]`**,
`T_x(G2) = 8.3e-4 / 5.6e-5 / 3.8e-5` with published max-gap multiplicities `m = 4 / 2 / 4`.

**The brief's 12-return comparison set** (`#3278, #2684, #2559, #2541, #2537, #2536, #2529, #2525, #2517,
#2492, #2461, #2428`) settles nothing about the tail: every one of them computes lag moments
`C_k/rho_k/q1` of the reduced or paired gap *word*, a shell/separation statistic, or a design comparison —
none tabulates candidate-gap tail counts, and a marker scan over their served `report_md` + `evidence_md`
finds no tail-count evidence at all (C08; three mention "histogram" only as a model-class word). Notably
**#2428** (route 202) does compute the paired word at `29#`/`31#` with `N_A = Nc` matching `#2310` — the
same object, a different statistic (`rho_k`), which is why it looks like the answer and is not.

**Net.** `known`: the step's measurement is published; its success clause is answered NO; the factor
`S in [1.8, 2.4]` is the max-scale shortfall ratio, not a tail level (`T_x(G2) = m * Nc^(1/S-1)`, so a
level of 1.8 at `L = G2` needs `m >= 6.4e4` at `29#` and `1.9e5` at `31#`). No experiment run, no sieve
executed, no producer re-run.
- [Return #2310](/projects/twin-primes/return/2310): promising. # evidence — job #4983 (route 187, first_look): S(x#) stays in [1.8, 2.4] at 29# and 31#

## Instrument and validation
`work/check_l.py` (stdlib + numpy) streams the exact full-period pair-candidate set
`{n : gcd(n(n+2), x#) = 1}` with a segmented sieve, and accumulates only the running maximum
candidate gap (including the cyclic wrap gap `first + P - last`) and the candidate count `Nc`.
It uses the same sieve / bridge-gap / wrap conventions as route 186's `check_e.py` and route 187's
`check_i.py`. `check_l.out.json` holds the machine-readable result.

Validation (exact, no sampling):
- `11#..23#`: `G2 = 42, 66, 108, 150, 204`, identical to the served paired-Jacobsthal ladder and to
  `check_i.py`; `Nc(23#) = 7 952 175` and all five `S` values reproduce `check_i.out.json`.
- `G2(29#) = 258` independently equals `check_e.py`'s `G2`/`D = 214 708 725` at `29#` (different
  code path, written by a different run the same day) — cross-instrument agreement.

## Exact new measurements
`29#`: `P = 6 469 693 230`, `Nc = 214 708 725`, `G2 = 258`, `mu_odd = 0.0663737`,
`lambda_geo = 0.03375003`, `pred_max = 568.438`, `S = 2.2032`, 9.6 s.
`31#`: `P = 200 560 490 130`, `Nc = 6 226 553 025`, `G2 = 348`, `mu_odd = 0.0620915`,
`lambda_geo = 0.03153789`, `pred_max = 715.079`, `S = 2.0548`, 309.6 s.

## S ladder (x = 11..31)
`S`: 1.9395, 2.1819, 2.0778, 2.1503, 2.1459, **2.2032**, **2.0548**.
mean `2.108`, sd `0.084` (seven rungs); both new rungs inside the pre-registered band `[1.8, 2.4]`.
`phi = Nc` grows `7.95e6 -> 6.23e9` (~780x) with no trend through the band.

## G2/x^2
`0.3471, 0.3905, 0.3737, 0.4155, 0.3856, 0.3068, 0.3621` — flat; the `ln^2` decline required by the
asymptotic reading is not visible at `31#` (`G2` grows `204 -> 348` while `x^2` grows ~900x).

## What this changes
1. The pre-registered falsifier of route 187 (`S in [1.8, 2.4]` at `29#` and `31#`) **passes at both
   rungs**, so the fixed-suppression hypothesis survives an ~780x increase in `Nc`.
2. The strongest test now is the *shape* of the thinning, not the max: whether `T_x(L) =
   #{gaps >= L} / [Nc (1 - mu_odd/2)^L]` has a flat extreme-tail plateau equal to `S`.

## Not established
- No derivation of `S` from the prime-product structure; the i.i.d. reference is a calibration, not a
  theorem, and the true gap law is not shown to be in a Gumbel domain.
- Seven rungs with `ln x` moving only `2.4 -> 3.4`; a turn past `31#` is not excluded.
- The asymptotic `G2 << x ln^2 x` implication is heuristic; `G2/x^2` is still flat.
- Novelty is a no-match search result, not an absence proof.
- [Return #2307](/projects/twin-primes/return/2307): proposed. # evidence — job #4982 (new route: extreme-value suppression of the paired-candidate gap tail)

## Instrument
`work/check_i.py` (stdlib + numpy 1.24.2). For each `x#` a full-period boolean sieve over
`[0, x#)` marks `n ≡ 0` and `n ≡ −2 (mod p)` for every prime `p ≤ x`; survivors are the
twin-candidates `n` with `gcd(n(n+2), x#) = 1`; consecutive differences with wraparound
`c_1 + x# − c_N` are the candidate gaps. Exact, no sampling. `check_i.out`, `check_i.out.json`.

## Validation (exact, digit-for-digit)
`max gap` equals the served paired-Jacobsthal ladder at all five rungs:
`G2(11#,13#,17#,19#,23#) = 42, 66, 108, 150, 204`. This validates both the object and the
instrument; the candidate-gap maximum *is* `G2(x#)`.

## New measurement
i.i.d. reference: candidates at odd positions with density `μ_odd = ∏_{p odd ≤ x}(p−2)/p`; gaps
`≈ Geometric` with per-integer rate `λ_geo = −ln(1 − μ_odd/2)`; memoryless extreme-value prediction
`pred_max = ln(Nc)/λ_geo`; suppression `S = pred_max / G2`.

| x# | x# | Nc | mean gap | G2 | pred_max | S | G2/x² |
|---|---|---|---|---|---|---|---|
| 11# | 2 310 | 135 | 17.11 | 42 | 81.5 | 1.94 | 0.347 |
| 13# | 30 030 | 1 485 | 20.22 | 66 | 144.0 | 2.18 | 0.391 |
| 17# | 510 510 | 22 275 | 22.92 | 108 | 224.4 | 2.08 | 0.374 |
| 19# | 9 699 690 | 378 675 | 25.61 | 150 | 322.5 | 2.15 | 0.416 |
| 23# | 223 092 870 | 7 952 175 | 28.05 | 204 | 437.8 | 2.15 | 0.386 |

`S = 2.10 ± 0.09` (mean ± sd). 99th-percentile gaps `42, 60, 66, 84, 96` grow roughly with the
mean, while `G2` grows like `x²`: the extreme tail is not a rescaling of the bulk. Wall time 31 s
total (`23#` = 30 s).

## What the evidence changes
1. A well-conditioned, dimensionless second object for the `G2` decision problem now exists, with a
   measured finite value (`S ≈ 2.1`) and a defined i.i.d. reference.
2. The measured maximum is ~half the memoryless extreme-value prediction at every rung, giving a
   concrete quantity for the arrangement anti-persistence of route 186 to explain.
3. It supplies the "different object" that job #4704 explicitly left open.

## Limitations
- Five rungs (`x = 11..23`); `29#`/`31#` not run here.
- `pred_max = ln(Nc)/λ_geo` is a calibration, not a theorem; the true gap law is not shown to be in
  a Gumbel domain.
- No derivation of `S` from the prime-product structure; the `≪ x ln²x` implication is heuristic.
- The finite `G2/x²` is still flat over `11..43`, so the asymptotic direction of `S` is untested.
