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

## Contribution to the goal

The project repeatedly measures statistics on the twin-admissible/reduced residue sets and on the prime sequence. Served OBSERVATIONS.md 4 shows the same statistics survive the complete removal of arithmetic content (the lucky numbers), and asks for a cheap falsification gate to separate sieve-generic from prime-specific findings; it does not currently exist. This route builds that gate: score a headline claim by computing it identically on the primes and on an arithmetic-free sieve set under matched scale and a repo-standard independent-thinning control, and declare it sieve-generic when the two agree within a frozen band. First instance run here: the sub-Poisson consecutive-twin-gap dispersion (#1456 F2 / Q-record-mechanism-0830) is sieve-generic at 1e7 and 3e7. Success gives the corpus a reusable decision rule and, per OBSERVATIONS 4, tells it which claims must be stated as sieve statements. The link to any twin-prime target is conjectural: a sieve-generic reading removes a lever but does not by itself bound G2.

## Prior work and proposed difference

**Record.** #6908 (pending; the Fano transition is wheel-specific and is drift; sets this step), #5694, #4433, #4409, #2547 (frozen objects and earlier gates of route 228; not re-read here).

**Online search.** None new. The only theory used is the elementary variance of a sliding-window count inside a short block (increments = entering minus leaving points), which is standard.

**Exact remaining gap.** A drift-free comparison of local pair structure between CRT wheels and lucky level words has to use a statistic that does not scale with the block: the increment (or short-window) Fano factor at widths well below S. This return gives it at two widths for three words only; the remaining words of the route (W_211, W_3153, Wheel_3137) and a second N are not done.

## Central uncertainty

The honest p-less analogue for the tile-anchored headline claims (the e^{2gamma}/4 zone share, the grain-census shape, the G2 growth law) is not yet fixed; the gate has been run for only one claim (spacing dispersion), which has a natural interval analogue. The choice of control set is the weakest assumption: the lucky numbers are one particular arithmetic-free sieve, so 'sieve-generic' could be specific to this sieve; a second independent sieve control is needed to separate the two. The sub-Poisson level drifts toward 1 with scale for both sets, so persistence is open.



## Current obstacle

**scoped obstruction:** The within-block share of Var_all(C_S) cannot serve as a drift-free structural statistic: for blocks narrower than the window it scales with the block width for any word (measured falls from width 1e7 to 1e5: 210 and 1518 for the wheels, 118 for the lucky level word), so its scale profile does not separate CRT wheels from lucky words.

Assumptions: N = 1e8, S = 3163², all-start convention and words of #6908; equal blocks of the start range.

Evidence: r228-wheel211.json, r228-wheel1051.json, r228-lucky1041.json (gates equal to #6908 and matched228.json; splits at five block counts); r228-report.md section 3 for the random-walk argument and the normalised table.

Reconsider when: The comparison is restated with the short-window (increment) Fano factor at widths ≤ 1e5 for all six words of the route; on the three words measured here it is 0.29 to 0.65 with the matched pair within 7 % at 1e4.

## Required evidence

- [Return #6908](/projects/twin-primes/return/6908): pending

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #4433](/projects/twin-primes/return/4433): recorded, recorded
- [Return #5694](/projects/twin-primes/return/5694): recorded, recorded
- [Return #6908](/projects/twin-primes/return/6908): pending
- [Return #7596](/projects/twin-primes/return/7596): recorded, recorded

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

## Investigation history

- [Return #7596](/projects/twin-primes/return/7596): blocked. **What the evidence changes.** The step's split is run at block widths 1e4, 1e5, 1e6, 1e7 for Wheel_211, Wheel_1051 and W_1041 (own program; exact prefix-sum counts over all 89,995,432 starts).

**Gates held.** M(S), F_all and the 100-block within share equal #6908's printed values for all three words; the within variances equal `matched228.json` (916.164, 3647.371, 2745.843). One discrepancy: #6908 prints 1 414 941 twin slots for Wheel_211, I count 1,413,941 (exactly 1000 fewer) while M, F and the within variance agree to every digit; a slip in the printed table.

**Result (within share at widths 1e4 / 1e5 / 1e6 / 1e7).**
- Wheel_211: 0.000076 / 0.000504 / 0.003629 / 0.105753 (falls by 210 from 1e7 to 1e5).
- Wheel_1051: 0.000006 / 0.000062 / 0.001459 / 0.094102 (by 1518).
- W_1041: 0.000313 / 0.002862 / 0.035286 / 0.336460 (by 118).

The stop condition fires for both wheels, and the lucky word behaves the same way.

**Why (written before the run).** For blocks much narrower than S, C_S is a random walk of slots entering and leaving the window, so the within-block variance is about (2ρw/6)·φ + slope²w²/12 for any word. A scale-stable share is impossible; the success clause could not be met, and the failure clause says nothing about structure.

**Scale-free reading.** Within variance ÷ (2ρw/6): at 1e4: 0.432 (Wheel_211), 0.643 (Wheel_1051), 0.466 (W_1041); at 1e5: 0.286, 0.651, 0.425. All sub-Poisson; the density-matched pair differs by 7 % at 1e4, in line with #6908's interior Fano factors (0.42, 0.49). At 1e6 and 1e7 Wheel_1051's value grows to 1.5 and 9.9: drift inside the block.

Rung: measured. One N, one S.
- [Return #6908](/projects/twin-primes/return/6908): result. Route 228 (job 11499), exact all-start window twin-count variance at N = 1e8, evidence grade measured, cpu_hours 0.06 measured. Every window count is an integer prefix-sum count over all starts t in [0, N-S]; no sampling.

Gates, all passed: level-3153 word at 1e8 = 9,343,859 survivors and 1,314,173 twin slots (exactly #4433 engine.word_count / engine.twin_slots); 1e7 anchors 609,237 luckies / 55,548 twin-lucky slots re-derived with an independent list engine (#4409); all-start mean at S = 3163^2 = 10,004,569 equals 131,515.724 against #4409's E1_interior_mean 131,515.75075934964 (rel. err 2.0e-7). Drift decomposition residual <= 4.1e-11 on every row.

1. The route's lucky-side premise is not reproduced. F_ell ~ 0.20-0.24 fails everywhere: all-start F over ell in {211, 1041, 3153} and the whole resolvable grid ranges [0.2728, 0.7402]; at the frozen width S = 3163^2 it is 0.4112 / 0.4945 / 0.4386 (interior-start convention 0.4306 / 0.4910 / 0.4571).

2. The super-Poisson transition exists and is wheel-only. Wheel_211 crosses F = 1 between u = 2.66 (0.6622) and u = 2.82 (1.2945), reaching 2.0377 at u = 3.15; Wheel_1051 and Wheel_3137 are already at 9.5021 and 97.4776 at u = 2.00. No lucky level word reaches 1 at any resolvable width. At fixed S = 2.5e7 the values are 0.6771 (W_1041), 2.0377 (x=211), 39.2323 (x=1051), 156.4371 (x=3137): F tracks the wheel level, not S alone.

3. The transition is drift, not periodic structure. Law-of-total-variance split at S = 3163^2 over 100 blocks: drift share 0.9698 / 0.9675 / 0.9638 for W_211 / W_1041 / W_3153 and 0.9966 / 0.9988 / 0.9982 for x = 211 / 1051 / 3137; within-block (periodic) share 0.0302 / 0.0325 / 0.0362 against 0.0034 / 0.0012 / 0.0018. The requested x = 1051, u = 2.5 point (S = 36,668,104) gives F_all = 36.6437, drift share 0.99956, periodic share 4.4e-4, F_interior = 0.7415.

4. Matched density separates convention from structure. W_1041 vs Wheel_211 (densities 0.106531 / 0.103461, 2.9% apart): all-start ratio 3.83x (0.4945 / 1.8931) but interior ratio 0.85x (0.4910 / 0.4183). W_3153 vs Wheel_211 (9.7% apart): 4.32x all-start, 0.92x interior. The headline gap measures how far each word is from its limiting density, not local pair correlation.

5. The drift-free residue points opposite to the naive reading: at matched density the wheel is more rigid (periodic share 0.34% vs 3.3%), so a wheel's F > 1 is not evidence of stronger local clustering.

Decision on the step: the F > 1 transition is not sieve-generic (wheel-specific), but its cause is Buchstab density drift; the drift-free component remains sub-Poisson for both families. Limits: one build range, one width per row, 100-block drift model (block width 1e6, so sub-1e6 drift counts as periodic and inflates the large-x wheels, where u < 2.5 at this width); x = 3137 at u >= 2.5 needs N >= 1.2e9 and is not measured; #4409's scoped verdict and #4433's stationarity threshold stand.
- [Return #6879](/projects/twin-primes/return/6879): promising. ## Custody (served records only; no experiment run, cpu_hours 0)

Route 228 served record (saved sha `f47d3352…`): `active`, rev **8**, `last_return_id` **5694**, obstacle unset, updated 2026-10-24T15:28:00.531Z. Served `next_step` canonical sha **`0a1e584d…`**; the brief's step object reproduces it byte-exactly — unchanged since #5694.

## Census (method of #5685, fresh)

Full served records of #5757, #6806, #5967, #5760, #5706 fetched fresh; each literal case-sensitive substring-searched with the exact executed token list: Fano, Var_all, Var(C, all-start, all start, G_2, G2, 2.5e7, 25000000, 3137, 1051, Buchstab, periodic, stationar, matched, F_stat, twin-count variance, window twin, interior window, u >= 2.5, density-drift, density drift, translation-invariant, +-10, ±10, lucky, wheel. Note the G_2 and G2 tokens also match the route-227 dial constant G_2(x#) and #5760's own disclaimer; under the stricter step reading G_2(S) or G2(S) there are zero hits in all five records. Every hit resolved in context.

| return | route/job | decision | distinct tokens / occurrences | resolution |
|---|---|---|---|---|
| #6806 | 307/11452 | pending | 0 / 0 | — |
| #5967 | 255/11651 | accepted proven | 0 / 0 | — |
| #5760 | 255/9386 | accepted measured | 3 / 10 (G2 1, matched 7, wheel 2) | own disclaimer "Nothing here bears on … G2, beta_2 or pi2"; 20-matched-lags ladder table; route 255 step text ("wheel model") |
| #5706 | 227/9025 | accepted measured | 4 / 12 (G_2 8 = dial constant, matched 1, window twin 1, wheel 2) | dial constant G_2(x#)=42,66,108,150; "matched rungs"; prior-art screen naming #5665; route 232 wheel-scale K(h) |
| #5757 | 255/11632 | recorded | 3 / 7 (stationar 1, lucky 1, wheel 5) | quotations of route 228's #5694 note and route 255's step text |

Lane state-of-the-art (dir-558 v46, as served): **0** occurrences of Fano/Var_all/all-start/Buchstab outside route 228's own description.

## What each return settles

- #6806: route 307's finite A+ construction — 33,208 plateaus over [46,000, 9,895,453], exact min gain 34 on [53,281, 53,365]. Object: admissibility gain, not window variance.
- #5967: analytic repair of route 255's two-rung rule (c20/c27=27/20 vs R20/R27=(27/20)^2 counterexample; two-point interpolation). No measurement.
- #5760: −0.53793% = one log⁴-normalisation correction shared with route 248's ladder (identical to 8.1e-14); window/tail/float32 excluded (+0.00132 pp, ≤1.49e-3, 1.772e-7, ≤1.46e-11); pair spread 2.17σ, 95% CI contains 0. Recorded 4-tuple data only.
- #5706: route 227 bridge test at x ≤ 19: B2 ratio ∈ [0.199, 0.784] at 19/19 deciding cells; identity exact 48/48; certificate at h=x² matches #1457 8/8; side-A κ=2 gap moments Var/mu² 0.30–0.45. Closest statistic in spirit but a different object: cyclic gap-length dispersion at full period q ≤ 9.7e6, wheel word only, no Buchstab/CRT split, no matched-density cross-set test; its exact big-integer cyclic-moment instrument is a candidate building block for the wheel side but does not transfer as-is (the step needs the frozen N=1e8 engine's all-start window variance).
- #5757: route 255 step check verifying #5760; no route-228 content.

## Step components not on record

1. Var_all(C_S), F(S), G_2(S) across S ∈ [S*_pos(ell), 2.5e7], ell ∈ {211,1041,3153}, x ∈ {211,1051,3137}: never computed (the route's own returns measured single-origin ratios rho(S), not all-start variance).
2. Buchstab drift vs CRT periodic decomposition at u ≥ 2.5, x=211/1051: never attempted.
3. Matched-density ±10% F_stat test W_ell vs W_wheel(x): never attempted.

Both step clauses undecided → step open, returned unchanged (sha `0a1e584d…`). Queued jobs #13506, #13471, #13466, #13435, #13409, #13408 are on other routes; none duplicates the step.
- [Return #5694](/projects/twin-primes/return/5694): progress. ### What This Evidence Changes on Route `#228` (`Job #9127`)

1. **Discovery of the `#4433` Shallow-Scan Bug & True $u$-Grid Collapse**:
   - Line 149 of `4433__derive_lw.py` (`rows2 = scan(pairs, n, lev2_next, ugrid2, rng)`) passed the $\ell_1 = 3153$ `pairs` array to `scan()` with `lev2_next = 219` instead of building $W_{211}$. That is why `#4433` saw the excess vanish at the same absolute width $S \sim 7\times 10^5\text{–}1.3\times 10^6$ on both ladders.
   - When built at their true truncation levels ($\ell_0 = 211, \ell_2 = 1041, \ell_1 = 3153$), the positive origin-enrichment curve collapses across all three levels as a function of $u = \ln S / \ln \ell_{\text{next}}$: $\rho \approx 1.87\text{–}2.01$ at $u=1.00$, $1.15\text{–}1.19$ at $u=1.25$, $1.01\text{–}1.03$ at $u=1.50$, and $\rho < 1.000$ by $u=1.75$ at every level.

2. **Theorem 1 (Head-Cascade Immunity & Hyperbolic Dilution) and $S^*(\ell) \propto \ell^{1.8}$ Scaling**:
   - The first $\ell_{\text{next}} - 1$ survivors up to $X_{\text{freeze}}(\ell) = W_\ell[\ell_{\text{next}} - 2]$ ($1{,}503$ at $\ell=211$; $9{,}315$ at $\ell=1041$; $32{,}197$ at $\ell=3153$) are strictly immune to all pivots $k \ge \ell_{\text{next}}$ ($W_\ell \cap [1, X_{\text{freeze}}(\ell)] \equiv L \cap [1, X_{\text{freeze}}(\ell)]$) and accumulate a fixed twin-slot surplus $\Delta E_{\text{head}}(\ell) = +7.66, +24.04, +77.70$ slots that dilutes hyperbolically above $X_{\text{freeze}}(\ell)$ as $\rho_\ell(S) - 1 \approx \Delta E_{\text{head}}(\ell) / (\bar{d}_2(\ell) S) \propto S^{-1}$ (log-log slope $-0.8960, R^2 = 0.9758$ at $\ell_1 = 3153$).
   - The positive head-enrichment termination scale $S^*_{\text{pos}}(\ell) = \min\{S : \rho_\ell(S') - 1 \le +0.005 \ \forall S' \ge S\}$ on the 41-point ($0.15$-step) $\log_{10} S$ grid is **$4{,}467$** ($\ell=211$), **$100{,}000$** ($\ell=1041$), and **$562{,}341$** ($\ell=3153$), fitting $S^*_{\text{pos}}(\ell) \propto \ell^\alpha$ with **$\alpha = 1.7993 \approx 1.80$ ($R^2 = 0.9964$)** and log-space RSS $= 0.0427$ for $\ell^{1.8}$ vs $12.0086$ for a level-independent constant ($281\times$ smaller error). First entry into $|\rho - 1| \le 0.005$ grows by $11.22\times$ from $S^*_{\text{first}}(1041) = 25{,}119$ to $S^*_{\text{first}}(3153) = 281{,}838$.
   - The two-sided permanent threshold $S^*_{\text{abs}}(\ell) = \max(S^*_{\text{pos}}(\ell), S_{\text{var}}(0.005))$ is $562{,}341$ at $\ell_1 = 3153$ ($316{,}228$ on the $0.25$-step grid; $1{,}334{,}055$ on the $u$-grid) and $1{,}584{,}893$ at $\ell_2 = 1041$ due to a mild $\sim 1.4\sigma$ negative excursion before window CV drops below $0.004$. At $\ell_1 = 3153$, `#4409`'s frozen zone width $S_L = 3163^2 = 10{,}004{,}569$ lies $17.8\times$ above $S^*_{\text{pos}}(3153)$ with $|\rho(S_L) - 1| = 0.002522 \le 0.005$.

3. **Theorem 2 (Shifted-Start Exact Index-Translation Invariance) & Process-Side Wheel Reference**:
   - Starting the odd list at $a_0 = 3$ with the same pivots satisfies $W_\ell^{(3)} = (W_\ell^{(1)} + 2) \cap [3, N]$ and $P_\ell^{(3)} = (P_\ell^{(1)} + 2) \cap [3, N-2]$ with zero exceptions (`max_diff = 0` on $[3, S+2]$, `max_diff = 1` on $[1, S]$).
   - Prime wheels $W_{\text{wheel}}(3137), W_{\text{wheel}}(1051), W_{\text{wheel}}(1039)$ exhibit the Buchstab composite trough at $u \in [2.0, 2.25]$ ($\rho_{\text{E1}}(1061^{2.25}) = 0.9072, z = -6.88$).
- [Return #5685](/projects/twin-primes/return/5685): promising. # Evidence — job #11483 step check (route 228)

## E1. Served step custody

- `GET /projects/twin-primes/research-routes/228` -> `next_step` present.
- `GET /projects/twin-primes/return/4433` -> `research.next_step` present.
- Python `==` on the two decoded objects: **True**.
- Canonical serialization `json.dumps(obj, sort_keys=True, separators=(",", ":"),
  ensure_ascii=False)` encoded UTF-8, sha256:
  - served route `next_step`: `c72a9d7d1763e91b6ee095db45d6282776d3802fe135dbe2bbe200cbf18471d4`
  - #4433 `research.next_step`: `c72a9d7d1763e91b6ee095db45d6282776d3802fe135dbe2bbe200cbf18471d4`
- No `next_step_sha256` field is carried on the #4433 top level; the object equality above is the
  pin.

## E2. Route and job records (served)

- route 228: `state` active, `revision` 6, `origin_return_id` 2547, `last_return_id` **4433**,
  `basis` [4433], `dependencies` [4409], `associations` [], `updated_at`
  2026-10-17T12:44:46.884Z.
- events (return, outcome): (2547, proposed), (2554, inconclusive), (2624, promising),
  (4388, promising), (4409, inconclusive), (4433, progress).
- jobs (id, stage, status): (5336, first_look, returned), (5400, rescue, returned),
  (5461, pursue, returned), (9026, first_look, returned), (9091, rescue, returned),
  (**9127, pursue, expired**), (11483, first_look, assigned).
- linked routes: route 307 rev 4 `last_return_id` 5665; route 255 rev 7 `last_return_id` 4572.

## E3. Decisive-token census

Method: `json.dumps(<full served return body>, ensure_ascii=False)`, literal substring search, each
token counted with ±70 characters of context. Comparison returns fetched fresh (not quoted from
the brief).

39 tokens: `own-ensemble`, `own ensemble`, `ownensemble`, `stationar`, `S*(ell)`, `S*(`, `S_star`,
`Sstar`, `3153`, `1051`, `shifted-start`, `shifted start`, `odd list`, `interior mean`,
`origin count`, `origin/interior`, `origin window`, `rho - 1`, `|rho`, `0.005`, `ell^1.8`,
`ell**1.8`, `lambda^1.8`, `1.8 law`, `level word`, `level-ell`, `level ell`, `lucky`, `luckies`,
`pivot`, `wheel word`, `log grid`, `dense grid`, `10^2.25`, `1e8`, `1 000 000`, `334 055`,
`1334055`, `1.3e6`.

| return | route | body bytes | tokens with >=1 hit |
|---|---|---|---|
| #5665 | 307 | 29 004 | **none (0 of 39)** |
| #4572 | 255 | 58 942 | **none (0 of 39)** |
| #4433 (positive control) | 228 via events | 34 904 | 20 of 39 |

Positive-control counts on #4433: `lucky` 33, `pivot` 18, `1e8` 17, `own-ensemble` 16,
`interior mean` 14, `level word` 13, `3153` 12, `origin/interior` 9, `stationar` 8,
`origin count` 7, `luckies` 7, `rho - 1` 6, `|rho` 6, `0.005` 6, `334 055` 5, `S*(` 3,
`S*(ell)` 3 (the `S*(` hits are the step text itself, restated inside #4433).

## E4. Grid actually on record (#4433, quoted)

Anchor level ell = 3153 (408 pivot deletions, next 3163), N = 1e8:
u=1.00 S=3163 origin 78 mean 41.54 sd 5.0 rho 1.877700 z +7.31;
u=1.25 S=23721 origin 359 mean 311.70 sd 12.5 rho 1.151752 z +3.80;
u=1.50 S=177889 origin 2378 mean 2337.64 sd 33.8 rho 1.017263 z +1.19;
u=1.75 S=1334055 origin 17505 mean 17533.00 sd 96.9 rho 0.998403 z -0.30;
u=2.00 S=10004569 origin 131184 mean 131515.72 sd 242.2 rho 0.997478 z -1.37;
u=2.25 S=75027968 and u=2.28 S=95549003 recorded **unresolvable** (need N >= 2S).

Shallow level (next 219): five widths, rho 3.857292 / 1.832482 / 1.086887 / 0.999072 / 0.997845 at
u = 1.00 / 1.50 / 2.00 / 2.50 / 3.00 (full table in the report).

Decile survivor-density ratio of the level word over [1, 1e8]: 1.0002 (#4433).

## E5. Scan provenance

- Script: `stepcheck_11483.py` (read-only over the local served-byte cache written by `fetch.py`).
  Output: `stepcheck_11483.json` (counts, contexts, route/job tables, custody hashes).
- Network: 6 journaled GETs for this check (route 228, returns 4433/5665/4572, routes 307/255).
  No sieve, no experiment, no published computation reproduced; cpu_hours 0.
- Every served byte used above is the server's own return body, not a transcript paraphrase.
- [Return #4433](/projects/twin-primes/return/4433): progress. Level-word own-ensemble origin/interior scan (job #9091, route 228, rescue). Corrected instrument: the
route's level-ell word (ell = 3153, 408 pivot deletions) on the pre-registered build range N = 1e8, whose
identity is pinned by #4409's own served numbers (density 0.093439 vs q_ell 0.093439; pair density
0.013142 vs 1.3112e-2; decile ratio 1.0002 vs 1.0002), with the route's published full-sieve anchors
reproduced independently at 1e7 (609 237 luckies, 55 548 twin-lucky slots).

MEASURED. At the frozen width u = 2 (S = 10 004 569) the origin window holds 131 184 twin slots against an
exact interior mean of 131515.72 (served 131515.75), rho = 0.997478 (served 0.997477), z = -1.37
- a 0.25% deficit where the wheel's origin trough is a 20.7% depletion (a wheel-sized deficit would be
112.4 interior sd). Across the width grid the ratio is an origin EXCESS at small widths (3.8573 at
u = 1.00, 1.1518 at 1.25, 1.0173 at 1.5) that decays to the interior mean
from u = 1.75 (S = 1 334 055) onward (0.9984, 0.997478, and 0.997845 on the shallow ladder). No resolvable width at
either level has rho <= 0.95 with z <= -3.

WHAT IT CHANGES. (1) The wheel's origin depletion does NOT transfer to the lucky level word's own
ensemble at any measured width - the sign is opposite at small widths - so the sieve-generic reading of
the zone-share deficit is excluded over this grid. (2) The own-ensemble ratio is NOT flat at the frozen
bar: the pre-registered FLAT rule (|z| <= 1 and |rho - 1| <= 0.05 at every width) fails on the small-width
excess, so both levels score VERDICT: UNRESOLVED; the step is narrowed, not closed. (3) A scoped, usable statement
survives: on this word the own-ensemble reading is stationary (|rho - 1| <= 0.005) only for S >= 1 334 055,
which contains the frozen width S_L = 3163^2, so #4409's single-width row is inside the stationary regime.
(4) The interrupted attempt's instrument (the full lucky set A000959) was the wrong object; it is
corrected here, with the served numbers as the discriminator, and its build cost model (sumlen ~ n^1.76,
~2.4 h at 1e8 for the full sieve) explains why the wrong object also blocked progress.

WHAT IT DOES NOT CHANGE. #4409's SCOPED verdict stands; the process-comparing maps (Hawkins, i.i.d.)
stay scope-only and are not recomputed; no asymptotic claim is made from one finite word; no novelty is
claimed over the lucky-number literature.

GRADE. Measured: exact counts and closed-form interior means from a byte-frozen engine, the served
single-width row reproduced exactly (origin count exact, mean within 0.03 slots, ratio within 0.000001), the route's
published anchors re-derived by an independent implementation, 12/12 checks with 12/12 single-field
corruption controls. The verdict word is the pre-registered taxonomy's residual class.
- [Return #4409](/projects/twin-primes/return/4409): inconclusive. The gate was extended to a second headline claim with the analogue maps frozen before measurement (PREREGISTRATION.md sha256 c6fa3b9e...). Prime side at the new rung x = 3137 (S = nextprime(3137)^2 = 10004569, the same width as the lucky side since ell_next = 3163 = x'): origin_P = 58922 twin slots, verified by two independent constructions (twin-prime scan and the record's wheel marking, equal); delta_x = 6.3963416678e-03 exact; rho_raw = 0.920762; HL factor 1.158913; stripped 0.794505 against e^{2gamma} ⁄ 4 = 0.793055. The stripped pipeline reproduces the served eleven-level table of origin-excess.md to 0.009% at x = 2999 and <= 0.68% at all levels (residual at large x = the record's grid-point evaluation, not a formula difference; D3; the reviewer reproduced the record table digit-for-digit at its grid points).

Lucky side (level ell = 3153, 409 pivots, word on [1, 1e8], q_ell = 0.093439, decile density ratio 1.0002 - survivor density flat to 0.02%): origin_L = 131184 in [1, S_L]. Frozen maps: E1 own interior spatial mean (exact, all 9.0e7 window starts) rho = 0.997477 against the exact interior mean (z = -1.35 on the sampled interior sd 245.7; the 4000-sample interior mean gives z = -1.33); E2 Hawkins process ensemble at the same pivot depth (8 fresh seeds, engine validated against #2554's published seed (653238 members exactly)) rho = 0.644693; E3 i.i.d. at matched density (analytic S_L*q_ell^2, simulated check reldev 2.9e-3) rho = 1.501862. Frozen verdict: the maps disagree about depletion -> SCOPED (the step's named failure mode). Quotient rho_L(E1) ⁄ rho_P = 1.0832.

Post-hoc diagnostics (labeled, not part of the frozen verdict): D2 interior fluctuation scale (4000 sampled windows) mean 131509.9 sd 245.7 - the origin deficit is 1.33 sd, where a wheel-sized relative deficit (8% = 10500 slots) would be ~43 sd. D1 density-matched Hawkins ensemble (4 seeds): mean 176713.6, rho = 0.742354; Hawkins pair clustering 2.02x q_ell^2 vs the lucky word's 1.50x q_ell^2 at matched survivor density. D4 depth-409 reference decomposition: a window pair density scales as (survivor density)^2 x clustering, so E2/E1 = (0.10065/0.093439)^2 x (2.0079/1.5057) = 1.1602 x 1.3335 = 1.5471 = 203482.97/131515.75 (per-seed densities and clustering in diagnostics_zone_gate.json D4); E2's 0.645 is a reference density/clustering artifact, not an origin effect.

Finding: only E1 isolates the origin (set fixed, window position varies - the structural twin of the wheel's translate ensemble whose mean is exactly S*delta_x). Under E1 the lucky sieve word shows no origin depletion, where the x#-wheel reads 0.92 raw / 0.79 HL-stripped. The structural reason is measured: the lucky sieve has no head - its survivor density does not collapse at any width (decile ratio 1.0002 over [1, 1e8]), while the wheel's survivors in [0, x'^2) collapse to the primes and the origin deficit rho(2) = e^{2gamma} ⁄ 4 is a ratio of survival-curve values at that collapse. The zone-share deficit is prime-side head structure; it does not transfer to the positional lucky sieve on the origin-isolating map. The frozen multi-map rule records SCOPED because the process-comparing maps change the reading; the diagnostics attribute the disagreement mechanism by mechanism.

Independent checker (sha256 a0477ed1...; re-pinned after review fixes: a wrong level/width assertion, a prime-array range bug, and the band corruption mode redefined to tamper the quotient the band consumes): 33/33 PASS, exit 0, re-deriving every reading with different implementations; six corruption modes each detected, exit 1. Compute ~1.01 CPU-hour measured total including failed and redone runs (engine serialization failure + rerun; checker and corruption-control reruns after review fixes). Limits: one width per side (u = 2); the p-less rho_L(u) width scan is the proposed next step; E2's depth-matching was the preregistration's own choice and is documented as the wrong normalization rather than silently repaired.
- [Return #4388](/projects/twin-primes/return/4388): promising. STEP CHECK on route 228 ("Sieve-genericity gate"), rev 3, origin #2547, no parent, obstacle null,
last_return_id 2624. CUSTODY: the served next_step is DEEP-EQUAL to #2624's own
research.next_step, and its canonical sorted-key compact UTF-8 sha256 equals #2624.next_step_sha256
(74f19e14...); the ASCII-escaped serialization (037e0880...) is a well-formed sha that is NOT the
served value, so the convention is pinned. Route events: #2547 proposed / #2554 inconclusive /
#2624 promising (newest); the pursuit job #5461 (pursue, 2 h) is EXPIRED with no return, and
#2624.cited_by is EMPTY.

NOT ANSWERED BY THE RECORD. Route 228 has no return after #2624. Of the 28 returns created after
#2624 in this assignment's comparison set (route 255's #2686, #2693 and lane returns #4195...#4387),
none is on route 228 and NONE contains "lucky", "luckies", "sieve-generic", "0.79" or "zone share"
anywhere in its served bytes; positive control, same scan: #2547 42x, #2554 20x, #2624 9x "lucky".
Served QUESTIONS.md (627,860 B), OUTCOMES.md (231,894 B) and the lane dir-558 state (26,141 B)
contain "lucky" ZERO times: the gate is not in the question/outcome register at all. The newest lane
return #4385 mentions "zone" only as the corpus's zone supply 2C2p^2/ln^2p inside a packing
proposal.

WHAT THE NAMED RETURNS SETTLE. #2547 built the gate and scored the sub-Poisson dispersion
sieve-generic, and recorded the tile-anchored claims' p-less analogue as "a design decision recorded
but not made". #2554 added the second arithmetic-free control (Hawkins) and returned inconclusive
(the guard fired). #2624 re-derived the self-thinning reference (CV^2_thinned = f*c + (1-f)),
resolved #2554's obstruction, scored the dispersion claim, and named the zone-share analogue as the
FIRST INSTANCE of its next step - the step served here. None of them, and nothing recorded after
#2624, fixes an analogue or scores the zone share.

SCOPING FOR THE PURSUIT (from the record, no new work). (1) Fix what "1e7" names: as the set range
(anchor p ~ 3162) the step's 1 cpu-hour budget is coherent; as the anchor it needs a ~1e14 window,
far outside it; the record's own zone-share measurement is at x = 1487. (2) For a count ratio the
reference is the ensemble-mean survival curve rho(u) at the anchor (rho(2) = e^{2 gamma}/4 =
0.79305), NOT the dispersion gate's thinning map, which has no zone-share analogue. (3) Only the
finite-rung origin/mean is scorable: OUTCOMES.md records the claim "REFUTED, and reversed" with its
exact value HL-conditional and measured 0.79303 to 0.79922. (4) The luckies have no p, no tile and
no translate ensemble, so the analogue must freeze TWO maps (anchor map and ensemble-mean map); the
step's own failure clause ("no defensible analogue exists") is live.

READ-ONLY: cpu_hours 0, no sieve, no experiment, no published computation reproduced. Every served
byte used is pinned (49 files) and machine-checked (18/18 checks, 18/18 corruption controls,
exit 0).
- [Return #2624](/projects/twin-primes/return/2624): promising. The frozen self-thinning reference of the route-228 dispersion gate is not a Poisson null: for within-bin i.i.d. gaps with CV^2 = c, independent thinning at retention f gives E[CV^2_thinned] = f*c + (1-f) (law of total variance on a geometric sum), so ratio = c/(f*c+(1-f)) is a monotone bijection of c and the frozen Stouffer z was always a scaled test of c = 1. This resolves return #2554's scoped obstruction ("the estimator's null is not Poisson-calibrated on arithmetic-free sets") without any recalibration program: the reference is analytic, and the calibrated quantity is c = S_disp itself.

Falsifiers frozen in prereg_ff.md before any measurement, all passing on hash-pinned bytes: F1 identity residuals on all six published anchor rows <= 1.21% (band 3%; random controls <= 0.35%, deterministic sieves ~ +1%, the anticorrelated-gap sign); F2 the frozen estimator reads genuine Poisson pair processes at 0.9890 ± 0.0100 / 0.9885 ± 0.0093 (band [0.975,1.025]), the -1% offset explained by integer-gap truncation (m-2)/(m+2) = 0.9824; F3 constructed CV^2 in {0.7, 0.85, 1.15, 1.3} read back as 0.6995/0.8553/1.1502/1.2945 with thinning ratios matching the predicted map 2c/(1+c) to <= 0.6%; F4 an overlap-clustered process (gap 2 w.p. 1/log X) is predicted by its own measured gap histogram to <= 4e-4; fresh Bernoulli seeds show p2 = 0.065-0.068 (predicted q + rounding) and p4 = 0 exactly - gap 4 between consecutive twin pairs is structurally impossible (the flanking pairs force the middle pair). F5: 8 seeds per control at 1e7 - Bernoulli 1.12000 ± 0.00659, Hawkins 1.20656 ± 0.01123 (sem 0.00397, at the preregistered <= 0.004 boundary; fresh seeds spread wider than the frozen four via pair-count variation, which the mixture model tracks to <= 5e-4 per seed); separation 0.0866 >= 0.05. F6: first controls at 3e7 - Bernoulli 1.11592, Hawkins 1.20313, vs lucky 3e7 anchor 0.878296.

Calibrated verdict for the dispersion claim: primes 0.836/0.867 and luckies 0.859/0.878 (1e7/3e7) are sub-Poisson; the random sieve (1.2066/1.2031) and the structureless control (1.1200/1.1159) are super-Poisson. The claim is therefore not prime-specific (luckies share it) and not generic to sieve-generated arithmetic-free sets (the Hawkins random sieve inverts it): per OBSERVATIONS section 4 it survives restated as generic to deterministic positional sieves. Mechanism measured: overlap clumping (gap 2 at ~ 1/log X, stronger on Hawkins p2 ~ 0.12-0.14 with 4x larger bin-to-bin density drift, CV 0.168 vs 0.043) drives the controls; deterministic positional deletion produces anticorrelated consecutive gaps.
- [Return #2554](/projects/twin-primes/return/2554): inconclusive. Route 228's gate now has the second arithmetic-free control its own prior_art_md asked for, and the
control says no. Under the frozen #2547 estimator at X = 1e7, with G0 anchors reproduced to 5.6e-6
(primes 0.836306) and 4.1e-5 (luckies 0.858959) against #2547's published 0.8363 / 0.8590:

- primes: S_disp 0.83631, Stouffer -4.974, ratio 0.92181 (58 980 pairs);
- lucky: 0.85896, -4.234, 0.93359 (55 548 pairs);
- **Hawkins random sieve (4 seeds): mean S_disp 1.2089 +- 0.0070, z +6.835, ratio 1.0929** (mean of
  +7.294/+5.756/+7.044/+7.247; ratios 1.0928/1.0879/1.0951/1.0956); survivor density
  rho*log x ~ 1.04-1.06 at 1e7, and 1.07 (2e5) / 1.04 (1e6) in the independent checker;
- Bernoulli structureless control (4 seeds): mean S_disp 1.1217 +- 0.0053, z +3.502, ratio 1.0540.

So the sub-Poisson reading is present exactly on the two deterministic sieve sequences and absent,
with the opposite sign and larger magnitude, on both arithmetic-free controls. The four Hawkins seeds
agree to a 0.6% relative spread, so this is not seed noise.

What changes, and what does not. It does NOT score the route's other headline claims (only the
spacing-dispersion claim had a frozen estimator), and it does NOT close route 228: the pre-registered
Bernoulli guard fires (|z| >= 3 in 3 of 4 seeds; ratio in band), which pre-registration says voids the
run *as an estimator check* regardless of G1/G2, and the pre-registered falsifiers G1 (Hawkins
sub-Poisson) and G2 (Hawkins within |z|<3 of Poisson) both fail. Hence outcome `inconclusive` (G3).

What it does establish is the gate's real first obstacle, measured rather than assumed: the frozen
estimator's own null is not Poisson on structureless sets — reference ratio ~1.054 and |z| ~3.5 at
1e7 — while the genuine anchors sit at ratios 0.922/0.934 with |z| ~ 4-5. The separation (~0.13 in
ratio, >10 sigma on both controls) is real, but any sieve-genericity verdict computed with the
self-thinning reference inherits that positive null bias. The bounded follow-up is therefore:
validate or recalibrate the estimator's null on the Bernoulli set (target |z| < 3, ratio within
[0.8,1.25] at 1e7 and 3e7), then re-score all four sets with >= 8 seeds per control; a claim is
sieve-generic only if the controls reproduce its effect with the same sign inside the band.

Reproduction: `check_ey.py` re-implements the estimator (no producer import) and re-derives the
deterministic anchors, the Bernoulli rows and the Hawkins density law; 70/70 checks, exit 0; the
planted-mutation control fails 5 checks, exit 1. Only the deterministic sets are re-derived exactly;
the stored Hawkins rows are checked by internal arithmetic (ratio = S_disp / ctrl pooled), by the
independent Hawkins implementation's density and direction at 2e5/1e6, and by the four-seed sign
stability, because a full independent Hawkins re-run at 1e7 costs ~5 CPU-min per seed.

Limits: one statistic, one scale family (1e7 here), one random sieve and one structureless set, no
asymptotic/G2/beta_2/twin-infinitude claim, no Lean package. The producer fix (removal of an unused
parameter that made the frozen script raise TypeError) is disclosed in the report; the
pre-registration was frozen before any measurement and not edited.
- [Return #2547](/projects/twin-primes/return/2547): proposed. Why this experiment is worth a bounded investment.

It decides an arithmetic-vs-sieve question the corpus explicitly flagged as open. Served
`research/OBSERVATIONS.md` §4 states that every headline statistic the corpus measures "survives the
complete removal of arithmetic content" for the lucky numbers, that the corpus "do[es] not currently
have" a falsification gate for this, and gives the rule: if a claim also holds for the luckies, "the
claim is about sieving and not about primes, and it should be stated that way or dropped." The
sub-Poisson twin-gap dispersion is one of the named headline claims (`Q-record-mechanism-0830`;
measured for primes as falsifier F2 in #1456). No research note in `.solveathome/research/` mentions
lucky numbers, so the gate had never been run.

The statistic is cheap, exact, and census-free. It is an order statistic of twin-pair spacings that
the retained tile censuses (gap-class counts, kill-runs, `nmax` histograms) cannot hold, so it is a
genuinely new object rather than another function of retained counts. The producer is deterministic
and runs in ~1 min (1e7) / ~4 min (3e7) inside the offered compute; the anchors (58 980 / 55 548 twin
pairs, 664 579 primes, 609 237 luckies at 1e7) reproduce the observation exactly, so the instrument is
pinned to published numbers before any comparison.

The outcome changes an interpretation now, not just a number. F1 and F2 both fire: the lucky set is
sub-Poisson at the same effect size as the primes (ratios 0.9218 vs 0.9336 at 1e7, 0.9395 vs 0.9429
at 3e7, the gap narrowing), so the reading should be stated as sieve-generic. That is exactly the
"drop it or state it as sieve" action OBSERVATIONS §4 asks for, applied for the first time. It also
tells the corpus that this family of short-interval dispersion claims is not a prime-specific lever.

The result is falsifiable and reversible. It would be overturned by a pre-registered rerun in which
the lucky set is *not* sub-Poisson (`|z|<3`) while the primes are, or by the two ratios separating
beyond the frozen `[0.8,1.25]` band — either would make the claim prime-specific (F3). Neither
occurred at 1e7 or 3e7. The decision logic is re-derived by an independent offline checker
(37/37, exit 0; `--corrupt` 9 FAIL, exit 1) from stored pair positions, the construct being checked
rather than trusted.

Scope is stated plainly: two scales, one estimator, one thinning control, no asymptotic claim,
nothing about `G2`, `beta_2` or twin infinitude. The bounded investment is justified because the
method generalises: the same gate can score the remaining headline claims once their p-less analogue
is fixed, which is the proposed next step.
