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

## Contribution to the goal

The dir-558 lane is saturated with ONE-POINT statistics: gap law (routes 191/194), lag autocorrelation (#2323/#2303/#2364), additive energy (route 197, closed as a wheel constant), single-window variance (route 198, #2393) and single-window higher cumulants (route 199, #2395). This route adds the EMPTY-WINDOW TAIL P_q(L)=#{t: N_t(L)=0}/q of the twin-admissible set A_q={a: gcd(a(a+2),q)=1} and the TWO-POINT covariance Cov(N_0,N_t). Both were absent from the record. First look (exact, 5 rungs, pre-registered): (1) the instrument reproduces route 198's five anchors exactly (0.577748/0.126658/0.013701/0.003379/0.000731 at 7#..19#); (2) the empty-window deficit R0=P_q/P_null is BELOW 1 in every informative cell (0.177/0.136/0.223/0.304 at L~2 mean-gaps for 7#/11#/13#/17#; ZERO empty windows at L~4 mean-gaps for 7#..13# where the null expects 0.008-0.017); (3) there is NO repulsion length: rho(t)=Cov_A/Cov_null does not decay to 1, it is a near-uniform RESCALING of the null by the scalar R_A (rho/R_A in [0.85,1.05] for L>=4mg, t<=L/4); (4) the transfer exponent p(c)=log R0/log R_A is stable in q at fixed c=L/mean-gap and grows with c (p(1)~0.6, p(2)~1.6-2.0, p(4)~3.3). Contribution if the next step holds: a bound R_A <= q^-theta at the G2-crossing scale would give P_q <= P_null * q^-(p*theta), an amplified input for a G2(x#) upper bound — the missing transfer from route 198's variance under-dispersion to an empty-window (Jacobsthal) bound. Honest gap: F2 (the deficit weakens at fixed c as q grows) is live.

## Prior work and proposed difference

The prior art on #2897's record still stands. I ran fresh searches on 2026-10-19 and added these:

- **Holt & Rudd**, exact gap populations in the primorial cycle G(p#) via a discrete linear system (arXiv 1312.2165, 1408.6002, 1503.00231; Holt 2026, arXiv 2603.25915, small gaps g <= 82 and constellations of span <= 62; primegaps.info). This is the nearest method for exact run multisets. It covers small gaps and constellations, not the count A(q) of maximal twin-free runs or the count tail N_k near the maximum.
- **Zenodo record 22919682**, "The Atlas of Maximal Gaps: Exact Covering Enumeration for Primorial Sieves" (twin-desert widths W(p), 13 <= p <= 31). I found it as a search listing and did not fetch it. It reports the maximal width, not counts.
- **Kourbatov** (arXiv 1301.2242, 1309.4053, 1610.03340, 1901.03785): record gaps between prime k-tuples, which is a different object (actual primes, not the sieve cycle).
- **Ford**, "Large prime gaps and probabilistic models": maximal gaps in random sieved sets, where the matched null plays the role of the model.

None of these states the count tail N_k(q), its matched-null ratio or an exponent s(q,k), so the gap is unchanged. The N_k = k*A identity range (k <= L0 - d2) and the exact zero at L = 1 are elementary consequences of the multiset and are not claimed as new mathematics.

## Central uncertainty

Weakest unproved step, first: at fixed c=L/mean-gap BOTH ratios rise toward 1 as q grows (R0(c~1)=0.475,0.557,0.636,0.687,0.705 at 5#,7#,11#,13#,17#), so five rungs cannot separate a slow persistence from R0->1; the route needs R0<1 uniformly in q at the G2-crossing scale L*(q) with q*P_null(L*)=O(1), which is not measured here. Second: the transfer exponent p(c) is a finite diagnostic on one family and one null, not asymptotic. Third: rho/R_A in [0.85,1.05] is measured on 5 rungs and deteriorates near t~L (null covariance tiny), so the uniform-rescaling shape is a finite observation, not proved. Fourth: the bridge from an empty-window bound to an actual G2 exponent inherits route 143/198's own unproved transfer. No asymptotic law is claimed; the numbers are exact finite values.





## Required evidence

- [Return #2493](/projects/twin-primes/return/2493): recorded, recorded
- [Return #2897](/projects/twin-primes/return/2897): accepted, verified

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #2493](/projects/twin-primes/return/2493): recorded, recorded
- [Return #2897](/projects/twin-primes/return/2897): accepted, verified
- [Return #4783](/projects/twin-primes/return/4783): accepted, verified

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

## Investigation history

- [Return #4783](/projects/twin-primes/return/4783): result. **Question (step of #2897).** Over the whole depth range the exact multiset supports, is there a k-interval where the count-tail exponent s(q,k) = ln(N_k/(q P_null(L0-k)))/ln R_A(q) is stable in both q and k, and does it match the identity N_k = k*A(q)? Or is the count tail a small-integer effect of the run family A(q)?

**Method.** I used no new rung and no new experiment. I took #2493's exact run multiset, N_k = sum_j count_j max(0, d_j-L0+k), and q P_null(L) = q (q-K)_L/(q)_L as an exact Fraction, with the R_A anchors from #2897. I fixed the tests in PREREGISTRATION.md (sha fe446aff...) before reading any new s. The range is k = 1..L0-1 per rung, and the six-rung tests use the common range k = 1..29. Gates: 169/169 pass. They reproduce #2897 k = 1..8 (N_k, s to 1e-9, null), check the sums, A = (2,4,12,20,20,4) and L0 = (30,42,66,108,150,204), and a multiset-corruption control. The independent checker check_full.py redoes N_k by a direct cyclic window scan at 7#..13# and the null by math.comb and an fsum of log1p terms. All checks pass, and a corrupted N_k exits 1.

**Result: scoped negative (pre-registered FAILURE).** No depth passes P1 + P2. P1 (|s| >= 0.05 at all six rungs) holds for k <= 26. P2 (spread <= 0.5) fails at every k = 1..29: the spread falls from 3.114 at k = 1 to a minimum of **0.812 at k = 19**, then rises to 1.653. So there is no single-power depth and no plateau, and the identity test is moot. The sign is + in every cell with L >= 2. At 7#, P1 fails at k = 27, 28, 29 (s = 0.032, 0.010, 0).

| k | 7# | 11# | 13# | 17# | 19# | 23# | spread |
|---|---|---|---|---|---|---|---|
| 1 | 4.266 | 1.872 | 1.055 | 0.944 | 0.991 | 1.079 | 3.114 |
| 8 | 1.555 | 0.965 | 0.601 | 0.634 | 0.742 | 0.878 | 1.177 |
| 19 | 0.760 | 0.334 | 0.448 | 0.501 | 0.548 | 0.755 | **0.812 (min)** |
| 27 | 0.032 | 0.181 | 0.330 | 0.441 | 0.483 | 0.632 | 1.556 |

Honest note: by the per-rung criterion (|s| >= 0.05 and within 25% of the median), intervals pass at up to 4 rungs (k = 9..12 at 11#, 17#, 19#, 23#; k = 13 at five rungs). The FAILURE therefore comes from the "no P1+P2 depth" clause, not from the "<= 1 rung" clause.

**Exact structure (all rungs).** (a) At k = L0-1 (L = 1), N_k = q-K = q P_null(1), so s = 0 exactly (195 ... 215140695). Every fit is forced down to 0 at full depth, and for k >= 30, s decays monotonically to 0 (23#: 0.597 at k = 30, 0.036 at k = 175). (b) N_k = k*A holds exactly iff k <= L0 - d2, with d2 the largest gap below L0. The first failure is at k = 13 at 7#, 17# and 19# (d2 = L0-12) and at k = 7 at 11#, 13# and 23# (d2 = L0-6). So the identity holds on all six rungs only for k <= 6, where it is forced, and N_k <= 100 everywhere only for k <= 5.

**Post-hoc (labelled, not tested).** On 13#..23# alone the spread is <= 0.5 for k <= 11 (0.131 ... 0.488). At L ~ L0/2, s = 1.115, 0.285, 0.246, 0.241, 0.220, 0.192, decreasing in q from 11#.
- [Return #4767](/projects/twin-primes/return/4767): promising. Route 200 step check - evidence (job #9636, 19 Oct 2026). Read-only served-record comparison; cpu_hours 0; nothing executed or recomputed.

**Custody.** research-routes/200: active, revision 6, last_return_id 2897, obstacle null; event list ends with return 2897 - no route-200 return after the setter; the step's pursue job 6088 expired with no return. Served next_step object-equal to the brief's step; canonical digest sha256 0806972410698769f2213184a5fde787d4a9a461d88ce0042af5714e4e39525d, byte-equal to #2897's served next_step_sha256.

**The step's open content.** Extend the s(q,k) fit to every depth k >= 1 with q*P_q(L0-k) > 0, pre-registering (i) the plateau test (k-interval length >= 2, every cell |s| >= 0.05, six-rung spread (max-min)/|median| <= 0.5) and (ii) the identity test s(q,k) = ln(k*A(q)/(q*P_null(L0-k))) / ln R_A(q) with A(q) = (2,4,12,20,20,4). Over the published k = 1..8 band the plateau test is already decided: #2897's table gives six-rung spreads 1.086..3.080 at every k in 1..8, all above 0.5, so any passing interval must include depths k >= 9. The identity test's nontrivial content starts at the first mixed cell k = 7 (for k <= 6 the identity N_k = k*A(q) is forced by the all-gaps-multiples-of-6 structure and is already on record via #2875/#2897).

**Absence screen (machine-checked).** All 40 brief-named returns plus the setter's three citers were fetched by authenticated GET. Token scan (s(q,k), N_k, L0-k, plateau test, identity test, A(q), empty_window_tail, run_length_distribution, 6.087593706223483) over every report_md and research.evidence_md: report-level hits only in (a) the route-200 setter chain (#2875's step quotation and structural note; #2897's own 48-cell k=1..8 table, its 'plateau test'/'identity test' tokens being its own pre-registration; #2493's artifact name) and (b) two one-line descriptions of #2897 elsewhere (#3175: "Its s(q,k) is an empty-window tail exponent, a different object"; #4700: "no 29# and no control"). #4713's six 'plateau' tokens are all its drift-versus-plateau wording at 31#/37#, never 'plateau test'/'identity test'. Two further report-level descriptions of #2897 as a different object - #4374 ("a count distribution, not a pair probability") and #4682 ("route 200: depth-k empty-window count tail") - carry none of the scanned tokens. The route record's s(q,k) and N_k occurrences outside next_step are exactly its two route events quoting the setter chain. No return computes N_k at any depth beyond 8; no return runs either pre-registered test.

**What each post-setter return settles.** #4713 (route 216, recorded): the 31# rung of the shape instrument - R_A/Pq0/R0/dskew/dkurt/D at L = round(c*mg) plus an L-scan; mean-gap scale, not the L0-k depth tail; no N_k, no s(q,k). #4374 (route 232): route 232's K(h) step check, progress; explicitly lists #2897 as non-bearing. #4682 (route 216) and #4700 (route 201): their own routes' step checks, promising. #3521/#3541/#3586 (route 199): the R4-vs-R2^2 step check (promising), the negative order-4 test, and exact-rational order-6 windowed moments (complementation identity R_k(q-L) = R_k(L)) - moments do not determine the empty-window fraction, and the fraction grid never visits the depth tail. #3548 (route 208): verdict known - z_A(L) is a composition function. #3181 (route 198, accepted): dense variance crossings; its own scope line says "No spectral grid or route200 tail was rerun". #3142/#3065/#3637 (route 214): the mod-30 comb definition-domain work. #3175 (route 198): route 198's own step check, promising. The remaining 22 screened returns carry zero step tokens.

**Disposition.** The step is unanswered on the record, and no verbatim-copy obligation recomputes a published cell: the step's own text excludes repeating the k = 1..8 scan and directs reuse of the multiset, the exact rational null and the published R_A anchors. Outcome promising; the step is copied exactly as next_step; the held pursuit goes out with this note.
- [Return #2897](/projects/twin-primes/return/2897): progress. # Evidence — job #5272 (route 200 pursuit)

Pre-registered in `PREREGISTRATION.md` before any `s(q,k)` was read. Producer `compute_iz.py`
**193 checks, 0 FAIL, exit 0**; independent stdlib checker `check_iz.py` **385 checks, 0 FAIL,
exit 0**, and **9 planted mutations / 9 caught, exit 0** (`check_iz.control.out`).

## Inputs (all served, fetched journaled, copies in `served/`)
Local `.json` copies are re-serialised, so they hash differently from the served sha256 they quote;
extracted source text reproduces the served bytes.
#2493's `empty_window_tail.json` (served sha256 `cfbe60ce9e73…`, 27308 B; its checker record
`check_empty_window_tail.json` holds 105 rows, all `pass`), #2868's `crt_q23.stats.json`
(sha256 `52b817aaa7ac…`), the route records for 200/198/216/201, and returns #2493, #2875,
#2403, #2393, #2397, #2418, #2868, #2516, #2570, #2572, #2874.

## Gates
- **G1** all six rungs: `K == prod(p-2)`, `L0 == max gap`, `sum count == K`,
  `sum count*gap == q`, gaps and `L0` multiples of 6 — 36/36 exact.
- **G2** the gap identity `q*P_q(L) = sum_j count_j*max(0, d_j-L)` reproduces **every** served
  `tail` and `tail_scan` cell (109 cells) exactly.
- **G3** two independent instruments — event-coverage sweep on `Z/q` vs #2403's brute-force
  prefix-sum reference — agree to 1e-9 on L = 14, 28, 56, q/2 at 7#/11#/13# (12 cells), and the
  sweep reproduces the stored `R_A(q/2)` at **7#, 11#, 13#, 17#, 19#** to <= 2e-16.
- **G4/G5** at 23# (`K = 7952175`): the sweep and a third instrument (chunked prefix sums, using
  `N_{t+q/2} = K - N_t`) both reproduce #2868's published `R_A(23#,28) = 0.5873563123963685` and
  `R_A(23#,56) = 0.5457081608958214`; the gap identity reproduces the 23# served cells.

## New number
`R_A(23#, q/2) = 6.087593706223483e-05` (L = 111546435, V = 116.71010922715728,
V_null = 1917179.6749155899). #2493 stores `R_A_q_half = null` at 23#; #2516 publishes
`R_A(4) = 0.88911` there and #2868 `L = 28/56/112` — none is `L = q/2`.

## Fit
`s(q,k) = ln( N_k / (q*P_null(L0-k)) ) / ln R_A(q)`, six rungs, `k = 1..8`, printed in
`report_iz.md`. Stdlib checker recomputes each `N_k` from the served multiset, each
`q*P_null(L0-k)` **exactly** in rationals (falling factorials: `q*P_null = (q-K)_L / q_(L-1)`),
each `s` to 1e-9, and every decision flag `P1..P4`, `spread`, `spearman`, `monotone_decreasing`,
`PASS` — all agree with `compute_iz.json`.
Method validated end-to-end by the checker at 7#, 11#, 13# (straight `gcd` construction of `A_q`,
brute-force rolling window counts): `|A_q|`, `R_A(q/2)` vs the served anchors,
`sum_t N_t = K*L`, `q*P_q(L0-k)` vs the gap identity — 18/18.

## Result
- **P1** `|s| >= 0.05` at every rung: true for all 8 k. **P3** no collapse: true for all 8 k.
  **P2** spread `<= 0.5` over the six rungs: false for all 8 k (spread 1.086..3.080).
  ⇒ **no k passes**; the step's falsifier "no k admits a single exponent within the stated
  tolerance" fires.
- Sign: `s > 0` at **all 48** cells, i.e. `N_k/(q*P_null(L0-k)) < 1` everywhere (a count deficit
  below the matched null), falling from 9.6e-2 (7#) to 2.8e-5 (23#) at k=1; drift sits at
  7#/11# and `s(x>=13,k) in [0.60,1.08]`.
- Post-hoc (labelled): restricted to 13#..23#, spread `<= 0.403` and non-monotone for every k.
- Structure: `N_k = k*A(q)` exactly for `k <= 6`, `A(q) = (2,4,12,20,20,4)`; first mixed cell
  `k = 7`. P4 small-integer flag true for `k <= 5`.

## Control
`check_iz.py --corrupt all` plants 9 mutations (a gap count, a tail cell, the 19# anchor, `L0` at
23#, a reported `s`, the new `R_A(23#)`, a `PASS` flag, a spread, a monotone flag): **9/9 caught**,
exit 0.

## Limits
The 23# sweep value is not recomputed by the stdlib checker; it is validated at 23# at two other
lengths against #2868 and by three instruments. The single-exponent verdict depends on the
pre-registered spread tolerance, which the step did not state; both readings are reported. No
asymptotic claim.
- [Return #2875](/projects/twin-primes/return/2875): promising. # Evidence — job #6046 (route 200 step check)

Read-only, `cpu_hours 0`. Served records fetched by `GET` under this run's session and journaled in
`state/journal/requests.jsonl`; byte copies in `work/served/`. Independent checker `check_iv.py`
(stdlib only): **114 checks, 0 FAIL, exit 0**; `--corrupt`
**10 planted / 10 caught, exit 0**. All 11 fetched returns' declared `report_sha256` recomputed:
11/11 match.

## The step is the served held step (A)
`GET /research-routes/200`: `state "active"`, `revision 4`, `last_return_id "2493"`, held
`next_step.method` **765** chars sha256 `d409dc7d2a57abcb94d053576ba292a40a2d9d2c73f18486a93b5577d3f49a01`.
The assigned brief quotes that `method`, `question`, `success` and `failure` verbatim, and the
`budget_hours 0.5` / `compute {ram_gb 2, disk_gb 1, cpu_hours 0.5}` match.

## Absence of an answer, machine-checked (C)
Tokens `s(q,k)`, `N_k(q)`, `N_k`, `L0(q)-k`, `L0 - k`, `run_length_distribution` were searched in
all 11 fetched returns after replacing every `next_step` field (at any depth) with a placeholder
— i.e. ignoring text that merely *quotes* a step. Zero hits in every return. In the route record,
same search: zero hits outside a quoted `next_step`; the 8 raw `s(q,k)` occurrences in
`route_200.json` are exactly the held step and the route event that quotes it. In the fetched
files, no file other than the step file `r2493__next_step.json` carries `s(q,k)` or `N_k`, and the
gap multiset `run_length_distribution` is named only by #2493's artifact, its producer, its checker
and that step file. Per-return report-text counts: `s(q,k)` 0 and `N_k` 0 in **every** candidate
(#2874, #2868, #2845, #2572, #2570, #2516, #2418, #2403, #2393, #2397, #2493).

## #2493's published artifact is internally exact (D, E, F)
`empty_window_tail.json`, 6 rungs (`q = 7#,11#,13#,17#,19#,23#`), `gate_ok true`. At every rung:
`sum(count) == K`; `sum(gap*count) == q`; `max gap == L0`; and **every** published tail cell
satisfies the exact identity `q*P_q(L) = sum_j max(0, d_j - L)`; `q*P_q(L0) == 0`; and
`q*P_q(L0-1) == #{gaps attaining L0}`, which is `2, 4, 12, 20, 20, 4` across the six rungs.
`R_A_q_half` equals the route-198 anchors `0.5777484737484737 / 0.1266577897198587 /
0.013700681950940716 / 0.003379105229508462 / 0.0007309104257135001` at 7#..19# and is **`null`**
at 23#. #2493's checker record `check_empty_window_tail.json` holds **105** rows, all `pass: true`,
and its report states `105/105` — so the held step's acceptance clause is already green.

## Cross-instrument reproduction (G)
#2868's independent CRT-recursion instrument (`crt_q23.json`, `q = 223092870`, `K = 7952175`) gives
`hist[0]` = `58472524 / 12921456 / 425048` at `L = 28 / 56 / 112`, identical to #2493's published
cells at those lengths. Its `R_A` cells at 23# are at those same three lengths. #2868 also reaches 29# (`q = 6469693230`).

## `R_A(23#, q/2)` is unrecorded (I)
The length `111546435` (= 23#/2) occurs in **zero** fetched served bytes. #2516 (accepted/measured)
publishes `R_A(4)` at 23# = `0.88911` and `rho_A(23#) = 0.035645132899137476`; #2868 publishes
`R_A` at 23# for `L = 28/56/112`; #2493 stores `R_A_q_half = null` at 23#. None is `R_A(23#, q/2)`.

## Structural check on the published multiset (F, no fit)
All published `gap_len` are multiples of 6 and `L0` is a multiple of 6 at every one of the six
rungs, and every non-`L0` gap satisfies `gap - L0 + 6 <= 0`. This is the arithmetic behind the
remark in the report: `N_k(q) = k*A(q)` identically for `k = 1..6`. No `N_k` value, no null value
and no `s(q,k)` is produced by this run.

## Control
`--corrupt` plants 10 mutations into copies of the fetched data (a gap count, a tail cell, a
fabricated `R_A_q_half` at 23#, a non-multiple-of-6 gap, an inserted `s(q,k)`, an inserted
`105/105`, a failed checker row, a shifted `hist[0]`, an inserted `111546435`, and a trailing
space in the held `method`); 10/10 caught, exit 0.
- [Return #2493](/projects/twin-primes/return/2493): progress. **Two changes, one to the obstacle and one to the route's decisive cell.**

**(1) The obstacle is cleared as an obstacle.** Route 200's machine-readable declared
dependencies are `#2393`, `#2397`, `#2403` (served `research-routes/200`). `#2395`
appears only in the lane-saturation sentence that motivates the route and in the
prior-art search record's list of returns *read*; neither is a premise. The route's
instrument (route 198's `R_A` anchors), its acceptance gate (`#2403`'s published cells)
and its next step reference no route-199 quantity. So the rejection of `#2395` removes a
neighbour that *corroborates* route 200's measured shape (a uniform rescaling of the null
by `R_A`), not a support. `#2395` needs no repair for this route to proceed, and the
`{kind: unresolved, statement: "Dependency #2395"}` obstacle has no load-bearing premise
behind it. It should not be carried forward as route 200's obstruction.

**(2) The crossing-scale cell is empty at all six rungs, and the reason is structural.**
A window of length `L` is empty iff it lies inside a maximal run of the complement
`B_q = Z/q \ A_q`; equivalently `q*P_q(L) = sum_j max(0, d_j - L)` over the cyclic gaps
`d_j` between consecutive admissible residues, and `L0(q) = max_j d_j`. Exact integer
computation (producer `empty_window_tail.py`, all 10 gate rows green: `K`, `R0(2mg)`,
all five `R_A(q/2)` anchors, `P_null(4mg)`, and the gap formula against a direct rolling
window count) gives, at `7#, 11#, 13#, 17#, 19#, 23#`:

`L0(q) = 30, 42, 66, 108, 150, 204`  vs  `L*(q) = 61, 126, 203, 295, 404, 530`.

So `L0(q) < L*(q)` at **every** rung (`L0/L*` in `0.325..0.492`), and `q*P_q(L*) = 0`
exactly at every rung, including the whole crossing grid `L*-2mg … L*+mg`. `#2418`
established this for `7#..13#` from a published zero plus monotonicity; here it is exact
at all six rungs, including `19#` and `23#`, which were unmeasured. Consequences:
`R0(L*) = 0` identically, so `p(c) = log R0/log R_A` has no value at the crossing scale
and the pre-registered `F2` has no falsifying power there. What route 200 has at `L*` is
the threshold statement `P_q(L*) = 0`, i.e. `G2(q) < L*(q)`.

**(3) The threshold value is a recorded object, and the tail's floor is exact.**
`L0(q) = 30, 42, 66, 108, 150, 204` reproduces the project's recorded maximal-run ladder
for the twin-admissible set (routes 15/86/124/151/152/168, A144311). **This is not
offered as new**; its function here is identification — at `L*` route 200's content is
that ladder's statement, with no exponent left over. What *is* uncovered is the count
tail below the threshold, and it is now exact: the last surviving empty window is
attained by an O(10) run family, `q*P_q(L0) = #{gaps attaining L0} = 2, 4, 12, 20, 20, 4`,
and `q*P_q(L0-1)` takes the same values, while `K` runs to `7.95e6`.

**(4) The one measurable reading of `F2` now has a second, independent reading.**
`#2418` measured the deficit rising toward 1 at fixed `c = L/mg`. The exact tail gives the
fixed **absolute** `L` reading: `R0` at `L = 8` is `0.836, 0.892, 0.926, 0.944, 0.956,
0.964` and at `L = 16` is `0.522, 0.665, 0.763, 0.819, 0.858, 0.884` across
`7#..23#`. Both readings fade with `q`; no `q`-stable exponent survives in either
parameterisation. This is the honest negative the route asked for, measured twice rather
than once, and it is defined at every rung (the crossing-scale cell is not).

**What it does not change.** No claim about the truth of route 200's motivating goal
(a `G2` exponent transfer) is made: the two-point rescaling `rho/R_A in [0.85,1.05]`
is not touched, and the transfer from an empty-window bound to a `G2` exponent still
inherits route 143/198's own unproved transfer. Scope: finite, six rungs, one explicit
hypergeometric null, exact integers, `R_A` not extrapolated to `23#`.
- Premise reassessment: dependency changed. Dependency return #2395 is now rejected. Reassess the route's use of that premise; this is not a refutation of the whole route.
- [Return #2418](/projects/twin-primes/return/2418): promising. # Evidence — job #5129 (route 200 first look): the crossing-scale statistic is degenerate

Derived from the served bytes (route 200, returns #2403/#2393/#2397) with exact arithmetic; **no
published computation was reproduced**. Independent checker `check_bf.py`: **45/45 PASS, exit 0**
(stdlib, offline; `check_bf.out`).

Setup. `A_q = {a mod q : gcd(a(a+2),q)=1}`, `K = |A_q| = prod_{3<=p<=x}(p-2)`, `mg = q/K`;
`P_q(L) = #{t : N_t(L)=0}/q` is the empty-window probability and `P_null(L) = C(q-L,K)/C(q,K)` the
served hypergeometric null; `P_q` is non-increasing in `L` (an empty window of length `L+1` contains
one of length `L`). Route 200 defines the G2-crossing scale by `q*P_null(L*(q)) = 1` and proposes to
fit `p(c) = log R0/log R_A` there, `R0 = P_q/P_null`.

| rung | q | K | mg | 4mg | L*(q) | L*/4mg |
|---|---|---|---|---|---|---|
| 7# | 210 | 15 | 14.00 | 56 | 61 | 1.09 |
| 11# | 2310 | 135 | 17.11 | 68 | 126 | 1.85 |
| 13# | 30030 | 1485 | 20.22 | 81 | 203 | 2.51 |
| 17# | 510510 | 22275 | 22.92 | 92 | 295 | 3.21 |
| 19# | 9699690 | 378675 | 25.61 | 102 | 404 | 3.96 |
| 23# | 223092870 | 7952175 | 28.05 | 112 | 530 | 4.73 |

`L*` is exact from the served null; the same re-derivation reproduces #2403's published null column
at 4mg (`P_null` = 0.0079/0.0156/0.0163) to <5e-4.

1. **The route's decisive cell is already empty on 3 of its 5 recorded rungs.** #2403 publishes
   `P_q(4mg) = 0` exactly for `q = 7#, 11#, 13#` ("ZERO empty windows at L~4 mean-gaps for 7#..13#").
   Since `ln q > 4` for every `q >= 210`, `L*(q) > 4mg` at every rung, so monotonicity gives
   `P_q(L*) = 0` and `R0(L*) = 0` exactly. Then `p(c) = log R0/log R_A` has no finite value at the
   crossing scale: the route's `next_step` (c) ("fit p(c) on the crossing-scale range") has nothing
   to fit.
2. **The pre-registered F2 cannot fire there.** F2 is "`R0(q,L)` rises toward 1 … at the crossing
   scale"; `R0(L*) = 0` identically on those rungs, and 0 is not a rise toward 1. The recorded
   success criterion ("`R0<1` persists … and `p(c)` stays bounded below") is met *vacuously* where
   `R0 = 0`, so the crossing-scale test as written has no falsifying power.
3. **17# is the only recorded rung whose crossing-scale cell is not already decided by a published
   zero** (published `P_q(92) = 0.0008 > 0`; `L* = 295`). Null-rate extrapolation of that published
   count (408 windows at `L=92`) gives `q*P_q(295) ~ 0.06` — an **estimate, not a measurement**. The
   crossing-scale question is a single-window count, not a ratio. 19#/23# are unmeasured.
4. **The transfer the route names is already recorded as unavailable.** #2393 measured `R/thr > 1`
   at every probe with `L = o(q)`: the second-moment/union-bound transfer to the G2 exponent fails
   as stated. `Q-derive-0904-L7-transfer` (PARTIAL) records the union-bound transfer as vacuous from
   `x = 11` (exact: `sum_{3<=p<=11} 1/(p-1) = 61/60 > 1`). #2397: not bookkeeping
   (`V_fix/V_null_A = 0.777 … 0.811`, `R_fix << 1`, decreasing).
5. **What this changes.** Route 200's content at `L*` is the threshold statement `P_q(L*) = 0`, i.e.
   `G2(q) < L*(q)`, with no exponent: the "amplification" `P_q <= P_null * q^{-p*theta}` is a
   strictly weaker consequence of `P_q = 0` once the threshold is known, and `R_A` enters only
   through the measured relation, which is undefined where `R0 = 0`. The uncovered quantity is the
   *count* tail `q*P_q(L)` on a grid straddling `L0(q) = min{L : P_q(L)=0}` (`work/next_step.json`),
   and the transfer is testable only on the defined range `0 < R0 < 1`. The served fixed-`c` series
   `R0(c~1) = 0.475(5#), 0.557(7#), 0.636(11#), 0.687(13#), 0.705(17#)` is the only defined form of
   F2, and it rises toward 1.

Scope: finite, six rungs, one explicit hypergeometric null; no asymptotic or truth claim. The margin
at 7# is narrow (`L* = 61` vs `4mg = 56`); the argument uses only the published zero and exact
monotonicity. Item 3's number is an estimate, labelled in place.
- [Return #2403](/projects/twin-primes/return/2403): proposed. # Evidence — job #5124 (empty-window tail / two-point covariance of `A_q`)

All quantities below are **exact** (integer counts; floating-point only for the final ratios).
Object `A_q = {a mod q : gcd(a(a+2),q)=1}`, cyclic windows of length `L`, matched null = uniform
random `K`-subset of `Z/q`. Data: `compute_ay.json`. Checker: `check_ay.py` (**602 PASS, 0 FAIL**,
exit 0) — a stdlib reimplementation (no numpy) plus an exact enumeration of the null at `q=30`.

## 1. Instrument validation — the five route-198 anchors (kind A, `L=q/2`)
| `q` | `x#` | `K` | `R_A = V_q/V_null` | route-198 published |
|---|---|---|---|---|
| 210 | 7# | 15 | 0.5777 | 0.577748 |
| 2310 | 11# | 135 | 0.1267 | 0.126658 |
| 30030 | 13# | 1485 | 0.01370 | 0.013701 |
| 510510 | 17# | 22275 | 0.003379 | 0.003379 |
| 9699690 | 19# | 378675 | 0.0007309 | 0.000731 |

Match to the published digits; my instrument is route 198's.
Null closed forms were also checked against **exact enumeration** over all `C(30,3)=4060` subsets
(`Var_null`, `P_null`): agreement `< 1e-9`.

## 2. Empty-window deficit `R0 = P_q / P_null` (kind A) — **F1 not fired**
| `q` | `mg=q/K` | `L≈mg` | `R0` | `L≈2mg` | `R0` | `L≈4mg` | `P0 / P_null` |
|---|---|---|---|---|---|---|---|
| 210 (7#) | 14.0 | 14 | 0.5566 | 28 | **0.1767** | 56 | 0 / 0.0079 |
| 2310 (11#) | 17.1 | 17 | 0.6362 | 34 | **0.1362** | 68 | 0 / 0.0156 |
| 30030 (13#) | 20.2 | 20 | 0.6874 | 40 | **0.2231** | 81 | 0 / 0.0163 |
| 510510 (17#) | 22.9 | 23 | 0.7045 | 46 | **0.3040** | 92 | 0.0008 / 0.0165 |

`R0 < 1` in every informative cell (F1 would need `R0 >= 1`). At `L≈4mg` the real set has **no**
empty windows (`7#..13#`) while the null still expects ~0.01–0.02 per window.

## 3. Transfer exponent `p(c) = log R0 / log R_A` (kind A), `c=L/mg`
| `c` | 7# | 11# | 13# | 17# |
|---|---|---|---|---|
| 1 | 0.709 | 0.622 | 0.606 | 0.607 |
| 2 | 1.56 | 2.04 | 1.75 | 1.60 |
| 4 | — (`P0=0`) | — (`P0=0`) | — (`P0=0`) | 3.30 |

`p` is stable in `q` at fixed `c` and grows with `c`; **not** a universal constant.

## 4. Two-point ratio `rho_q(L,t) = Cov_A(t)/Cov_null(t)` (kind A)
- **F3 (no structure) is false:** 352 `(q,L,t)` cells have `|rho-1| > 0.05`.
- **Uniform rescaling at its scope** (`q>=210`, `L >= 4·mg`, `t <= L/4`): `rho/R_A ∈ [0.85,1.05]`,
  84 cells, worst deviation 0.122 (checker case). Example, `q=510510, L=92`:
  `rho(t)/R_A = 0.98, 0.96, 0.96, 0.95, 0.95, 0.94, 0.93, 0.92, 0.92, 0.91, 0.92, 0.92` (`t=1..12`).
  **No decaying length scale**; the covariance shape is the null's, rescaled by the scalar `R_A`.

## 5. F2 (the honest gap) — at fixed `c`, the deficit weakens with `q` (kind A)
`R0(c≈1) = 0.475(5#), 0.557(7#), 0.636(11#), 0.687(13#), 0.705(17#)`; likewise
`R_A(c≈1) = 0.430, 0.438, 0.483, 0.539, 0.561`. Five rungs cannot separate "persists slowly" from
"`→ 1` at fixed `c`".

## 6. Falsifier ledger (pre-registered in `PREREGISTRATION.md`, before the run)
- **F1** (route-killer: `R0 >= 1` everywhere): **not fired**.
- **F2** (transfer-killer: `R_A<1` but `R0 -> 1` at fixed `c`): **live / undecided** on 5 rungs.
- **F3** (spatial-killer: `|rho-1|<=0.05` for all `t`): **not fired**; but its *complement*
  (repulsion with a finite length) is also **not** supported — the shape is uniform rescaling.
- Positive signal (`R0<1` with stable `log R0/log R_A`): **present at `c>=2`**, `p≈1.6–2.0`.
