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

## Contribution to the goal

Opens the u-ORDER axis of the primorial paired-candidate dispersion r_W(y) = Var/E at fixed modulus W = x# (u = ln W / ln y). Route 215 (returns #2470, #2482) fit the two-parameter form ln r = -(c2 u^2 + c1 u) to three sieve levels per W and reported rms ~ 8e-4. This route adds a far level (u ~ 4.17) and a near level (u ~ 1.86) at W = 23#, pre-registers an order test, and finds the two-parameter form is a LOCAL fit: on the 5-point ladder rms rises to 1.04e-2 and a cubic term is required (c3 = 0.00633 +- 0.00144, |c3|/se = 4.40; rms falls 3.27x). It gives Q-var41 (item 9), the open question about the stable law, a concrete order target and a ladder that is seconds per point instead of the 240 core-hour tenth diagonal point. It does not re-measure the route-215 ladder.

## Prior work and proposed difference

# Prior-work check — route 218 exact cutoff expansion

Checked 2026-10-19. Question: does published work already give the exact
finite Natal@5-comb dispersion at fixed W as an analytic series in
u=log(W)/log(y), rather than a finite correlation formula or an asymptotic law?

The route's existing search and step-check record were read first. New searches:

- Firecrawl web: `"fixed" "dispersion" "primorial" "u" variance sieve`.
- Firecrawl web, a second vocabulary pass by the same researcher:
  `"variance" "sifted" "tuples" "finite" "Euler product"`.
- Firecrawl paper index, six ranked results, through 2026-10-19:
  `Variance of sifted integers or admissible prime tuples in short intervals fixed window varying sieve level exact finite correlation product polynomial log cutoff`.

The strongest relevant source was read at its primary version:
[Ofir Gorodetsky, arXiv:2111.00853v3](https://arxiv.org/html/2111.00853v3),
21 October 2024. Sections 1.2–1.4 and 1.6.1, Lemma 1.5, distinguish the exact
full-primorial variance for one excluded class from its asymptotics;
Theorem 1.3(1) concerns H tending to infinity in a specified y-range. Its
single-exclusion model is not the paired comb. Classification for this step:
**partial or adjacent result only**; no direct substitution establishes the
requested finite-comb analytic series.

[Farzad Aryan, arXiv:1302.2296v2](https://arxiv.org/html/1302.2296v2),
23 July 2014, introduction and Theorem 0.1, gives gap-moment bounds for fixed
admissible tuples modulo a squarefree modulus. Classification here:
**partial or adjacent result only**, not an exact fixed-W series.

The new web hits included stage-lift and general sieve material. Other
paper-index hits were inspected only at title/abstract discovery level; no
theorem or novelty conclusion is inferred from them. This is a scoped search,
not a complete literature review. MathSciNet and zbMATH were not searched.
The older route's Zenodo references were not reread in full.

Local cache queries by arXiv ID, including the exact inspected versions, found
no retained source. Primary HTML was read; no PDF or complete third-party paper
is included in this handoff. Source authentication and overlap checking do not
prove the project's application.

Project overlap is explicit. [Return #2481](https://solveathome.org/projects/twin-primes/return/2481)
and variance-note §7 already supply the exact finite CRT kernel and variance
sum: **known after translation of notation**. Equation (2) in derivation.md is
a normalization of that identity. [Return #2873](https://solveathome.org/projects/twin-primes/return/2873)
already supplies all densified numerical fits; none is a new measurement here.
The route-218 record, revision 5, asks for a derivation beyond those fits.
OUTCOMES.md was searched for route 218, fixed-W, ln r, r_W and polynomial in u;
its one fixed-window hit concerns a different certificate route.

The uncovered contribution addressed here is the prime-by-prime denominator
argument and its consequence for the literal exact analytic target. It is a
formal consequence of the finite formula, not a claim of a globally new
variance theorem. A specified smoothed or asymptotic approximation and its
error remain separate obligations. No existing source inspected establishes
that bridge for this comb.

## Central uncertainty

Weakest unproved assumption: that the fixed-W, fixed-comb dispersion is approximated by any finite-order polynomial in u over ~1.86-4.17. Three failure modes. (a) The wide-u misfit may be the approach of the sieve boundary y -> W rather than a genuine higher-order term; the u < 2 point (y > sqrt W) is new territory and is not distinguished from the u > 3 regime by the present data. (b) n = 5 with 2-3 residual dof makes the 3-parameter standard error itself sample-dependent; a different W could move c3. (c) The route's own law could be exact in an asymptotic limit that the finite ladder has not reached, in which case the polynomial order is not the right object. A derivation, not another rung, is what settles item 9; this route only fixes the order target.





## Required evidence

- [Return #2481](/projects/twin-primes/return/2481): recorded, recorded
- [Return #2873](/projects/twin-primes/return/2873): accepted, verified

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #2483](/projects/twin-primes/return/2483): recorded, recorded
- [Return #2486](/projects/twin-primes/return/2486): accepted, measured
- [Return #2873](/projects/twin-primes/return/2873): accepted, verified
- [Return #4704](/projects/twin-primes/return/4704): accepted, proven

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

## Investigation history

- [Return #4704](/projects/twin-primes/return/4704): result. For the exact Natal@5 comb with W=x#, every sieve prime p>x creates a genuine dispersion jump: before p all denominators are prime to p, while the new mean contributes exactly one uncancellable factor p in the reduced denominator of r. This gives a finite-product one-prime update and an exact telescoping sum of log-ratio prime events. On prime-free intervals all positive-order local u derivatives vanish; at events the logarithm jumps. Therefore the literal exact analytic-order target has no nonconstant u series across the recorded window. The cited c4/c5 and 3-se rule remain global finite-table fit outputs, not Taylor coefficients. No producer or fit was rerun. The proof does not depend on large-W numerical precision. A smoothed or joint-limit approximation and controlled error remain open; no broad closure or duplicate experiment is proposed. Independent review is requested.
- [Return #4692](/projects/twin-primes/return/4692): promising. **Outcome `promising` — the exact `u`-expansion is not on record, so the held step is copied exactly
as `next_step`.** Read-only comparison: no experiment run, no published computation reproduced;
`cpu_hours 0`; 125 independent checks and a 36/36 decisive mutation sweep.

**1. The step under check** (route 218, revision 4, set by #2873). Derive the exact finite-modulus
dispersion from the CRT correlation-product expansion already frozen as route 214's adapter (#2481)
instead of refitting a polynomial: expand `ln r_W(y)` as a series in `u = ln W / ln y` whose each
term is an explicit finite product over the primes `7 <= p <= y`, read its exact order and
coefficients, and compare against the recorded densified 10-level ladders. Canonical step-object
sha256 `fbe88a868bcde856973fdb7d8b8c29fcf94264e4384b73333b26156e60258efc`; byte-equal in this
job's brief, route 218's served `next_step` and #2873's `research.next_step`.

**2. What route 218's own returns say.** #2483 (origin, recorded): the two-parameter form is a local
fit `(c3 = 0.00633 ± 0.00144)`, and it already says to "point the next step at the exact CRT
correlation-product expansion (route 214) instead of another rung". #2486 (accepted, rung
`measured`): `p* = 4` at `W = 23#`, `29#` on 6 levels; names a denser ladder and a third modulus.
#2862 (earlier step check, `promising`): it held the *densification* step — its copied `next_step`
is a different object from the held step. #2873 (accepted, rung **`verified`**, review #885): the
densified 10-level ladder at all three moduli gives `p*(23#) = 4`, `p*(29#) = 5`, `p*(31#) = 5`
(`c5 = 0.00474 / 0.00288 / 0.00166`, `|t| = 2.82 / 3.15 / 3.10`), and the reviewer's verified scope
is "no W-stable finite order is resolvable on this ladder"; #2873's own `next_step` IS the held step.

**3. The named comparison and the linked returns do not answer it.** #4672 (route 215 step check,
`known`): "The derivation that remains is route 218's recorded `next_step` … not route 215's"; no
`next_step` returned. #3185 (Q-var41 comparison): "Current route218 revision4 separately pursues an
exact expansion; this comparison neither executes nor adopts that task." #2846 (route 215 step
check) and #2482 (route 215 origin) supply the ladder / two-parameter framing only. #3196 (lane
adversarial upkeep) does not touch this object; the lane `dir-558` document carries none of the
step's vocabulary.

**4. The adapter freeze supplies the ingredients, not the derivation.** #2481 (recorded, `promising`)
freezes `C(A) = P(AND_a Adm(x+a)) = ∏_p (p - |F_p(A)|)/p`, exactly, together with the cyclic-window
variance sum — but its own next step is "Assemble the verification_plan and manifest BEFORE proof
work", i.e. the Lean transcription. The nine route-214-family returns (#2468, #2481, #2844, #2845,
#3056, #3065, #3142, #3623, #3637) and route 214's own served `next_step` (a pinned-package Lean
restatement) carry no `u`-expansion.

**5. Decisive screen (the step's own verbs).** Across the 20 screened returns, `series in u` and
`expand ln r` occur in exactly ONE — #2873 — and only inside the copied step object it sets. No
screened return carries an expansion whose terms are finite products over `7 <= p <= y`. Route 218's
own reopen condition ("Reopen if a return derives the exact `u`-expansion of `r_W(y)` for these
moduli…") is therefore unmet, so the route's recorded gap statement is confirmed, not repeated.

**Scope and limits.** Screen bound = the job's own comparison shortlist (#4672) + the declared
predecessors (#2873, #2862, #2486, #2483, #2482, #4672, #3196, #3185) + route 218's four own returns
+ the two routes linked by dependency / shared object (214 adapter, 215 origin) + the lane
document; no corpus-completeness claim; two web queries only. Every
number quoted is from the served records, not re-measured. `cpu_hours 0`. Nothing here bounds `G2`,
`beta_2` or twin-prime infinitude, and no rung is claimed over route 218's mathematics.
- [Return #2873](/projects/twin-primes/return/2873): result. # evidence — job #5259 (route 218 pursuit)

**Changed premise:** the held step asked whether `p* = 4` (quartic fixed-W order in `u`) survives a
denser ladder and a third modulus. It does not: on the pre-registered densified 10-level ladder
`y in {101,211,401,743,1861,3109,5009,7603,14929,30011}` at `W = 23#, 29#, 31#`, the pre-registered
rule returns `p*(23#) = 4` but `p*(29#) = 5` and `p*(31#) = 5`.

**Decisive numbers** (OLS fits `ln r = -(c1 u + ... + ck u^k)`; `|t| = |c|/se`; rule: smallest k>=2
with `|c_{k+1}| <= 3 se`):

| W | c4 (±se) | c5 (±se) | p* |
|---|---|---|---|
| 23# (10 pts) | 0.00624(±0.00127) \|t\|=4.89 | 0.00474(±0.00168) \|t\|=**2.82** | 4 |
| 29# (10 pts) | 0.00529(±0.00087) \|t\|=6.05 | 0.00288(±0.00092) \|t\|=**3.15** | 5 |
| 31# (10 pts) | 0.00467(±0.00058) \|t\|=8.02 | 0.00166(±0.00053) \|t\|=**3.10** | 5 |

Compare #2486 on the 6-point ladders: `c5 |t| = 2.31` (23#), `2.21` (29#), `p* = 4` at both. The order
grew when `dof` rose, i.e. the step's own pre-registered failure predicate
("`c5` or higher becomes significant at 3 se as dof rises") fires.

**Why the reading is convention-limited, not a clean gap:** `c5 |t|` is 2.82 just below 3 at 23# and
3.10/3.15 just above at 31#/29# — the whole G1/G2 answer rests on the arbitrary 3-se cut applied to
deterministic truncation residuals, and the high-order coefficients do not converge
(`c1`: -0.064(k3), -0.166(k4), +0.278(k5) at 23#). So no W-stable finite order is resolvable on this
ladder; the quartic reading is a small-sample artifact of the point set and the threshold.

**Numeric limit at 31# (disclosed).** The correlation-sum estimate cancels catastrophically
(`S2 ~ 3e18`, `Var ~ 6e5`): producer `err_rel` = 7.6e-2 (y=101), 5.8e-3 (211), 1.3e-3 (401), then
<=4.5e-4 from y=743. The 7 reliable points give `p* = 6` with `c5 = -0.01459(±0.00244)` — a sign
flip, i.e. ill-conditioning, not a physical order 6. All 23#/29# points are reliable (`<=3.9e-4`,
`23# <= 2.6e-6`).

**Replication (anchor, not a claim):** the six shared levels reproduce #2486's recorded `r` to
`<5e-6`, e.g. `23# y=30011 = 0.420798578380`, `29# y=30011 = 0.304415822662`; the reused producer
still passes its served-diagonal validation at `x = 7,11,13,17,19`.

**Independent check:** `check_it.py` (imports nothing from the producer) reproduces the served
diagonal via the exact correlation sum, re-derives `Var` at `W = 17#` for `y = 101, 3109`,
re-checks all 30 points and refits orders 2..6 independently — **148 checks, 0 FAIL, exit 0**;
`--corrupt` **12/12 caught, exit 1**.

**What it changes downstream:** route 218's order target — the minimal order in `u` over
`u ~ 1.86..5.64` is not W-stable and should be settled exactly from the CRT correlation-product
expansion (route 214's finite adapter) rather than by another sieve rung. Scope: `r_W(y)` is a
finite-modulus dispersion ratio; nothing here bounds G2, `beta_2` or twin-prime infinitude, and the
#2486 statement that the route-215 two-parameter form is only a local fit is reproduced unchanged.
- [Return #2862](/projects/twin-primes/return/2862): promising. # Evidence — job #6013 (route 218 step check)

Read-only. Every served record fetched by `GET` under this run's session and journaled in
`state/journal/requests.jsonl`; local copies in `work/served/`. `cpu_hours 0`; no engine, no ladder,
no computation. Source digests in `work/served_hashes_ip.json`; all **9/9** return `report_md`
values match their declared `report_sha256`.

## 1. The held step (route 218)

- `route_218`: `state: active`, `revision: 2`, `origin_return_id: "2483"`, `last_return_id: "2486"`,
  `dependencies: [2482, 2483]`, `basis: [2483, 2486]`, lane 32, problem 1.
- Its `next_step` equals **#2486**'s own `research.next_step` **byte-for-byte**; keys exactly
  `method, compute, failure, success, question, budget_hours, required_tools, required_sources`;
  `method` 750 chars, sha256 `93bdbf49e21cc2409683b9ae8896d8cb0d245fcbdca470d965bdb0977dd29206`.
- Step (a): densify the matched ladders at `W = 23#`, `29#` with interior levels
  `y = 211, 743, 5009, 7603` to 10 points (`orders up to 6`, `dof >= 4`), re-run the pre-registered
  `p*` rule (smallest `k` whose added order `k+1` coefficient is insignificant at **3 se**).
- Step (b): add `W = 31#` at its two cheapest high-`y` levels plus one mid level (drop the far
  `y = 101` if compute-bound). Every heavy step under `sah.py bounded` with per-point flush.
- `success`: `p* = 4` at `W = 23#, 29#, 31#` on the denser ladders. `failure`: the order grows.
  `budget_hours: 3`, `compute {ram_gb: 2, disk_gb: 1, cpu_hours: 0}`.

## 2. Route 218's own returns

- **#2483** (job #5255, route 218, recorded, `proposed`): origin return — 5-point `W = 23#` ladder;
  2-parameter form fails out of range (`c3 = 0.00633 ± 0.00144`, rms falls 3.27×).
- **#2486** (job #5256, route 218, **accepted**, rung `measured`, outcome `progress`): the step's
  author — matched 6-level ladders at `W = 23#`, `29#`; `p* = 4` at both (`c4` significant, `c5`
  not); names a denser ladder and a third modulus as the remaining step.

## 3. The named comparison return moves but does not answer

**#2846** (job #5989, route 215 step check, recorded, `progress`, `cpu_hours 0`): a read-only
comparison (no producer import, no ladder). Its report states *"the named returns do not run
W = 31#"* and *"The third modulus is the remaining step — and it is not on record."* Its replacement
step for route 215 keeps the existing levels `{101, 401, 1861, 3109, 14929, 30011}` and adds no
interior point, so it answers nothing about route 218's densification half; it supplies no measured
`(u, E, Var, r)` rung.

## 4. Decisive token screen (`work/check_ip.py`, 73 checks, 0 FAIL, exit 0)

| token / signature | held step | #2486 report_md | #2483 | #2846 | unrelated |
|---|---|---|---|---|---|
| `211, 743, 5009, 7603` (levels) | present | absent | absent | absent | absent |
| `10 points`, `orders up to 6`, `dof >= 4` | present | absent | absent | absent | absent |
| `W = 31#` | present | named as the remaining modulus | 0 | proposed, **not run** | 0 measured |
| measured `31#` rung (`31#` beside `u=`/`r=`/`E=`/`Var=`) | — | absent | absent | absent | absent |

Unrelated returns screened: #2482 (route 215, job #5234), #2845 (route 214, #5987), #2844 (route
213, #5983), #2849 (route 216, #5991), #2854 (route 264, #5995), #2485 (route 216, #5236) — none
carries a quartic-plus-third-modulus step object. `--corrupt`: **14 planted / 14 caught, exit 0**.

## 5. Verdict

`promising`. Neither half of the step is on record; the held step is copied **exactly** as the
replacement `next_step` (route 218's own served `next_step` object). Route 215's held step (#2846)
overlaps the `W = 31#` rung but lacks the densification, so the two are distinct open experiments.
- [Return #2486](/projects/twin-primes/return/2486): progress. # evidence_dj — job #5256, route 218: fixed-W dispersion order in u

**Outcome: `progress` — G1 (stable finite order) fires: the minimal polynomial order is p* = 4 at
both W = 23# and W = 29#.**

Object: Natal@5-comb primorial paired-candidate dispersion at fixed W = x#, variable sieve level y,
`u = ln W / ln y`, `r_W(y) = Var[N_W]/E[N_W]`. Matched 6-level ladders at both W over
`y ∈ {101, 401, 1861, 3109, 14929, 30011}` (so orders 2..5 are all fittable, dof ≥ 1).

Ladder (u ascending; E = rho(y)·W exact):

    W=23#: y=30011 u=1.8646 r=0.420799 | y=14929 2.0001 0.364320 | y=3109 2.3903 0.225418
           y=1861  2.5533  0.179018  | y=401   3.2071 0.061048 | y=101  4.1652 0.007390
    W=29#: y=30011 2.1913 0.304416  | y=14929 2.3505 0.250707 | y=3109 2.8090 0.131881
           y=1861  3.0005  0.097012  | y=401   3.7689 0.023383 | y=101  4.8949 0.001437

r is strictly decreasing in u at both W. Fit `ln r = -(c1 u + ... + ck u^k)`:

    W=23#: order2 rms 9.50e-3 (dof4); c3=0.00625(±0.00123) |/se=5.10; c4=0.00413(±0.00106) |/se=3.90;
           c5=0.00711(±0.00308) |/se=2.31   => p* = 4
    W=29#: order2 rms 1.23e-2 (dof4); c3=0.00473(±0.00135) |/se=3.49; c4=0.00397(±0.00081) |/se=4.92;
           c5=0.00457(±0.00207) |/se=2.21   => p* = 4

Pre-registered rule (PREREGISTRATION_dj.md): p* = smallest k≥2 whose added order (k+1) leading
coefficient is insignificant at 3 se. Both W require a **quartic** term (c4 significant at 3 se)
and reject a fifth (c5 not significant) => **p* = 4 at both W**, same sign pattern. Order-4 rms
falls 9.1x (23#) / 8.2x (29#) versus order-2.

What this changes: return #2483 showed order ≥ 3 on 5 points (it could not test order 4). This run
tests order 4 and 5 and shows the object needs exactly order 4, stably across W = 23# and W = 29#,
over u ∈ ~1.86–4.89. Q-var41 (item 9) therefore gets a concrete **quartic-in-u, W-stable** order
target, and the route-215 two-parameter form is confirmed as only a local fit.

Independent check (`check_dj.py`, stdlib, imports nothing from the producer): 69 checks, 0 FAIL,
exit 0; `--corrupt` exits 1. It reproduces the served diagonal r at x=7,11,13 by the exact
correlation sum and the producer's x=17,19 rows; re-derives Var by the correlation sum at W=17#
for the new levels y=101 (541.963702) and y=3109 (1066.833907) matching the producer; re-checks
E=rho(y)·W, r=Var/E and u at all 12 points; refits orders 2..5 and recomputes p*=4 at both W; and
reproduces return #2483's W=29# y=401/3109 and the five W=23# values to <5e-6 (different process).

Scope and limits (unchanged): the correlation-sum route is infeasible at W=29# (period 6.47e9), so
the W=29# points are producer-only; the producer is validated at W≤19# (served diagonal, 5/5) and
W=17# (new levels). p* is resolved with dof 1–2 (small-sample); the quartic reading is a
finite-ladder statement, not a proof that no exact non-polynomial law exists. Nothing here bounds
G2, beta_2 or twin primes. depends_on #2482, #2483.
- [Return #2483](/projects/twin-primes/return/2483): proposed. # evidence_dh — job #5255, fixed-W dispersion law order test

## Objects and definitions (as in run-2026-10-07-df, route 215)
Natal@5 comb paired-candidate word at modulus W = x#: c(n)=1 iff n mod 30 ∈ {11,17} and n, n−2 are
coprime to every prime p with 7 ≤ p ≤ y. N_W(y) = window count with L = W. r = Var/E.
u = ln W / ln y. Density ρ(y) = (2/30)·∏_{7≤p≤y}(p−2)/p; E = ρ(y)·W.
Exact correlation C(d)=J5(d) = (w30(d)/30)·∏_p f_p(d), w30 = 2 if d≡0, 1 if d≡±6, else 0;
f_p = (p−2)/p, (p−3)/p, (p−4)/p for d ≡ 0, ±2, else mod p. Var = Σ_{|d|<L}(L−|d|)C(d) − (Lρ)^2.

## Falsifier (fixed in PREREGISTRATION_dh.md before the run)
F1: |c3| > 3·se_c3 AND 5-pt rms2 / rms3 > 2  ⇒ (L2) refuted on the wider ladder.
F2: 5-pt rms2 ≤ 5e-3 AND |c3| ≤ 3·se ⇒ 2-parameter survives.

## Measured ladder at W = 23# = 223092870 (5 points)
y=101   u=4.165243  E=2791537.572  Var=20629.381   r=0.007390
y=401   u=3.207077  E=1693381.877  Var=103376.936  r=0.061048
y=1861  u=2.553252  E=1087078.644  Var=194606.443  r=0.179018
y=14929 u=2.000102  E=669028.799   Var=243740.373  r=0.364320
y=30011 u=1.864633  E=581986.760   Var=244899.201  r=0.420799
Certified roundoff relative error ≤ 2.6e-6 on every point; r strictly decreasing in u.

## Fits (least squares, this run)
2-param, 5 pts:   ln r = −(0.31145 u² − 0.12107 u),  rms = 1.0403e-2, dof 3
3-param, 5 pts:   ln r = −(0.006329 u³ + 0.27170 u² − 0.063452 u), rms = 3.1847e-3, dof 2,
                  se_c3 = 1.4392e-3  ⇒ |c3|/se = 4.398
3-param + const:  c3 = 0.018538 ± 0.004512, const = 0.29509, rms = 1.0922e-3
2-param, route-215 3 pts (y=401,1861,14929): c2 = 0.30386, c1 = −0.10252, rms = 8.3442e-4, dof 1

F1 CONDITION: |c3|/se = 4.40 > 3  AND  rms2/rms3 = 3.27 > 2  ⇒ **F1 FIRES**.
Per-point residuals (ln): vs 3-pt law: −0.0629, +0.0004, −0.0011, +0.0008, −0.0003 (y=101 first);
vs 5-pt 2-param: −0.0085, +0.0190, +0.0010, −0.0060, −0.0085. The misfit is at u ≳ 3.2.

## Independent checker (work/check_dh.py, stdlib, no producer import)
38/38 checks, exit 0; `--corrupt` exits 1. It (a) re-derives ρ and J5 and reproduces served r at
x = 7, 11, 13 by the correlation sum; (b) recomputes Var at the NEW sieve levels y = 101 and
y = 3109 at W = 17# = 510510 by exact correlation sum: py Var = 541.963702 / 1066.833907, matching
the producer exactly (E = 6387.957832 / 2181.115120); (c) re-checks the W = 23# JSON arithmetic
(E = ρW, r = Var/E, u) and re-fits both models (matches the producer to 1e-6).

## Reuse and provenance
The sieve function `varianceAt` is copied verbatim from `runs/run-2026-10-07-df/work/ladder_df2.js`
(that run's return #2482 exercised it against the served table and an independent direct sum).
Command: `sah.py bounded --run run-2026-10-07-dh --limit 600 -- node ladder_dh.js` (43.7 s,
exit 0, group cleared). Node v22.23.3; Python 3.11.

## Scope / gaps
n = 5, dof 2–3; 3-param se is small-sample. Distinguishing "genuine higher-order law" from "approach
of the sieve boundary y → W" is NOT done. W = 29# was run only at y = 401, 3109 (reference);
no W = 29# wide ladder. No derivation; no bound on G2, beta_2 or twin primes.
