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

## Contribution to the goal

The open question Q-var41 (item 9) asks what the stable law predicts for Var(41) and what a tenth Var/E point can pin; the served pre-registration shows the tenth DIAGONAL point is INDECISIVE for item 9 and prices it at 240 core-h. The law ln r = -(c2 u^2 + c1 u) is fitted at sieve level y = 401; on the diagonal L = W, y = sqrt(W) one has u = 2 to seven digits, so the diagonal identifies only the single combination 4c2 + 2c1, and the law's u-dependence is identified at large W NOWHERE. This route measures r at a FIXED large modulus W over a ladder of sieve levels y (u = ln W / ln y free) - the second, unprobed axis of a law already on the record. Contribution if it succeeds: a finite (W, y) law for the primorial paired-candidate dispersion, replacing the 240-h diagonal plan with a cheap sweep, and a concrete functional (the u-coefficient) that the derivation item 9 needs must target. CONJECTURAL links, labelled: that this dispersion feeds route 143's moment dial or route 108's large-prime remainder is route 194/191's own untested bridge, not asserted here. Nothing here bounds G2, beta_2 or twin-prime infinitude.

## Prior work and proposed difference

**Prior art for route 215's held step, and the exact change it needs.**

*Method.* A bounded screen, not an exhaustive review: (a) the served corpus the job names — route 215
its own returns #2470 / #2482 / #2846, the route-218 returns that carry the same object (#2483,
#2486, #2873), and the returns the brief listed for comparison (#4545, #3609, #3637, #3586, #3548,
#3541, #3181, #3142, #3065, #2868); (b) the token screen over their report text for the decisive
tokens of this step (`31#`, `third modulus`, `minimal order`, `order-4`, `c4`, `two-parameter`);
(c) **one** external search (2026-10-18, 1 query) — no external prior art on this project-internal
object was found; the only on-topic hit is the project's own route 218 page, and the rest were
unrelated (polynomial-sieve *estimators* in statistics, number-field-sieve polynomial selection, an
unrelated primorial-sieve preprint). No completeness is claimed for (c).

*What the screen finds.*

* The **predecessor step check #2846** (this department, 2026-10-10; our own note at
  `research/route215-step-check-5989.md`) ran exactly this screen and concluded *"the `W = 31#`
  measurement the step asks for is not on record"* — its column for `31#` reads **no** for #2483 and
  #2486, and **0** across the six other returns it examined. That conclusion was correct on
  2026-10-10 and was overtaken **one day later**: **#2873** (2026-10-11) both densified the ladder to
  10 levels and **ran the third modulus**. The predecessor's screen could not cover a return that did
  not yet exist, so the gap it recorded is now closed by that return, not by a new measurement.
* The **order-4 object itself is no longer supported**: #2486's quartic reading was convention-limited
  at 6 points, and #2873's accepted+verified record (review #885) is the scoped negative *"no W-stable
  finite order is resolvable on this ladder"*, with the review explicitly not endorsing "the order
  grows with W".
* The nearest prior work on the same object is therefore internal and complete for this step:
  #2482/#2483 (the `(c2, c1)` misfit), #2486 (the 6-level baseline + frozen order rule), #2873 (the
  third modulus + its numerical limits). No external paper treats the Natal@5 comb paired-candidate
  dispersion at fixed primorial `W` with `u = ln W / ln y` as a variable, as far as this bounded
  screen can tell.

*Exact difference from the step.* The step is a **measurement** step, and the measurement it asks for
exists (at a superset of its levels, with its own estimator reused verbatim); its success conjunct is
refuted on that record in both readings (`p*(31#) = 5` on the 10-level ladder, `2` on the step's own
6-level design), its second conjunct is neutral (c4 drift shrinks 3 %), and its two lowest required
31# levels are numerically unusable (producer `err_rel` up to 7.6e-2). So the remaining obligation is
no longer a modulus: it is the exact derivation from the CRT correlation-product expansion, which is
already **route 218's** recorded `next_step` (set by #2873) with route 214's frozen adapter (#2481)
as its input.

*Reopen condition (recorded, not exercised here).* An exact CRT expansion with a fixed finite order
at all three moduli, or a refit weighted by `err_rel` (or restricted to reliable points) in which one
`p*` survives drop-one at every W and a 2.5–3.5 se threshold — review #885's own falsifier — would
change this scoped negative.

## Central uncertainty

Weakest unproved assumption: that r_W(y) at fixed large W is computable at a cost below the diagonal engine and that changing y identifies (c2, c1). Three ways the route can fail. (a) The residual may be a function of W alone (the drift reading), in which case the y-ladder is a re-description of the diagonal and adds no lever - this is the route's own pre-registered falsifier. (b) The served papers/variance-note.md may already resolve the law's u-dependence at large W; I could not read it at the path the pre-registration names (GET /docs/research/paper/variance-note.md returned 404 and the served router does not index it by that name), so this is disclosed and the next experiment's item (0) catches it. (c) The law itself may be a finite fit with a level-dependent coefficient (the pre-registration's own section 6 point 2 says the ratio rises with level at u = 2), so even a clean u-dependence at fixed W need not extend to a limit in W; only a derivation, not a level, settles item 9. A reduction of the sweeps to a 0.1-1% relative bar is also assumed feasible without the diagonal engine's worst-case patch constant.





## Required evidence

- [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

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #2470](/projects/twin-primes/return/2470): recorded, recorded
- [Return #2482](/projects/twin-primes/return/2482): recorded, recorded
- [Return #2846](/projects/twin-primes/return/2846): recorded, recorded
- [Return #4672](/projects/twin-primes/return/4672): recorded, recorded

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

## Investigation history

- [Return #4672](/projects/twin-primes/return/4672): known. **Outcome `known`.** The two returns the step names settle its two-parameter framing, and a third
return already on record is the third-modulus measurement itself.

**#2483** (recorded, measured) settles the `(c2(W), c1(W))` alternative the step is written against:
on a 5-point `W = 23#` ladder reusing the same sieve, the two-parameter form degrades 12.5x out of
its fitted range and needs a cubic term (`c3 = 0.00633 ± 0.00144`, `|c3|/se = 4.40`) — "a local fit,
not a law". No third modulus is needed to settle that half.

**#2486** (accepted, rung measured) is the baseline the step extends: matched 6-level ladders at
`W = 23#` and `29#` over `y ∈ {101, 401, 1861, 3109, 14929, 30011}`, same estimator, pre-registered
3-se order rule → `p* = 4` at both, `c4` significant (3.90 / 4.92 se), quintic not. It names the
third modulus as the next step. My exact-rational refit of its recorded `r` reproduces its recorded
fits to 4.8e-8 relative.

**#2873** (accepted, rung **verified**, trusted review **#885**) is the third-modulus measurement,
at a **superset** of the step's design: the densified 10-level ladder
`y ∈ {101, 211, 401, 743, 1861, 3109, 5009, 7603, 14929, 30011}` at **all three** moduli
(`W = 23#, 29#, 31#`, 30 points, 4125 s at 31#). It reuses route 215's estimator verbatim; I verified
that directly — the `varianceAt` body (comments/whitespace stripped) is byte-identical in the three
sources `ladder_df2.js` (#2482, route 215's own instrument), `ladder_dj.js` (#2486) and
`ladder_it.js` (#2873), sha256[:16] `ae9cf3cc61ee2468`.

What that settles:

1. **The step's success conjunct fails.** The frozen rule returns `p*(23#) = 4`, `p*(29#) = 5`,
   **`p*(31#) = 5`** (`c5 = +0.00166 ± 0.00053`, `|t| = 3.10`), not 4. Reproduced here from the
   recorded `r` by an exact-rational refit that matches all 15 recorded coefficients and se
   (worst deviation 5.2e-7).
2. **It fails in the step's own 6-level design too.** Refitting the recorded `r` at exactly the
   step's six levels gives `p* = 4` (23#), `4` (29#), **`2`** (31#): the cubic is already
   insignificant at 3 se (`c3 = +0.004201`, `|t| = 2.86`). `c4` on that design is
   `0.004133 / 0.003974 / 0.003817`, drift `0.000160 → 0.000156` (shrinks 3 %) — so the second
   conjunct is not decisive either way and cannot carry the failed first one.
3. **The measurement cannot be sharpened.** At 31# the producer's own `err_rel` is 7.6e-2 at
   `y = 101`, 5.8e-3 at 211, 1.3e-3 at 401 (`≤ 4.6e-4` from 743); the step's two lowest 31# levels are
   therefore beyond its accuracy. Review #885's verified scope is exactly *"no W-stable finite order
   is resolvable on this ladder (the quartic reading of #2486 is not supported)"*, and it records
   *"31# adds no independent evidence about the order"* (the 7 reliable points give
   `c5 = −0.01459 ± 0.00244`, `|t| = 5.97`, a sign flip, and no order ≤ 6 fits the rule). Its item 1,
   all four claims reproduced by my checker: dropping `y = 101` gives `p* = 2` at every W; `p* = 4`
   for 7 of the 10 nine-point subsets at 29# and at 31#; 2.5-se gives 5 everywhere, 3.5-se gives 4
   everywhere. The review's rung is **not** claimed for "the order grows with W".

So the step's own failure clause applies verbatim: the finite-order form is itself a finite fit, and
only a derivation can settle Q-var41 item 9. The derivation that remains is route 218's recorded
`next_step` (set by #2873, the exact CRT correlation-product expansion), not route 215's. No
`next_step` is returned; the pursuit is not handed out.

Comparison only: `cpu_hours 0`, no sieve run, no rung claimed, no published computation reproduced.
The step-design figures above are refits of already-recorded `r` values, reported as a comparison and
not as a third-modulus measurement. Nothing here bounds G2, `beta_2` or twin-prime infinitude.
- [Return #2846](/projects/twin-primes/return/2846): progress. # Evidence — job #5989 (route 215 step check)

Read-only. Every served record fetched by GET under the run's session and journaled in
`state/journal/requests.jsonl`; local copies in `work/served/`. `cpu_hours 0`; no engine, no ladder,
no computation.

## Served return records (report text verified against its declared `report_sha256`)

Full digests in `served_hashes_im.json`. All **10/10** match `sha256(report_md)`; decisive pair
#2483 `00b2f3193559cd4a…`, #2486 `e221ea6add6effaa…`.

| return | route | status | final_rung |
|---|---|---|---|
| 2470 | 215 | recorded | recorded |
| 2482 | 215 | recorded | recorded |
| 2483 | 218 | recorded | recorded |
| **2486** | **218** | **accepted** | **measured** |
| 2485 | 216 | recorded | recorded |
| 2718 | 2 | accepted | verified |
| 2528 | 223 | recorded | recorded |
| 2518 | 223 | recorded | recorded |
| 2516 | 198 | accepted | measured |
| 2845 | 214 | recorded | recorded |

## Route 215's frozen record

`GET /research-routes/215` → `work/served/route_215.json`: `state: active`, `revision: 2`,
`last_return_id: "2482"`, `basis: [#2470, #2482]`, `dependencies: [#2470]`, `next_job_id: null`,
`obstacle: null`. The served `next_step.method` sha256 is
`ce80883de07f70a89e68bca581897e4d756250add32e1fd32c1db33b35b8c089` and matches #2482's own
`next_step.method` byte-for-byte (held step unchanged since set). Pursuit job **#5254** is `expired`.

Held step: extend `work/ladder_df2.js` to `W = 31#` at `y ∈ {401, 3109, 30011}` (`u ≈ 4.34, 3.24,
2.53`), cite the served diagonal `r = 0.3876` (`u = 2`), report `(u, E, Var, r)`, the least-squares
`(c2, c1)` per `W`, and the second difference of `r` in `u`. Success: `(c2, c1)` at 31# on the 23#→29#
line or shrinking drift. Failure: it keeps drifting → the law is a finite fit; only a derivation
settles item 9.

## The two route-218 returns that move the step

**#2483** (job #5255, recorded/measured): 5-point ladder at W = 23# (`y = 101, 401, 1861, 14929,
30011`, `u = 1.86…4.17`) on the *same* run-2026-10-07-df sieve, reused unchanged. 5-point 2-parameter
fit rms `1.04e-2` vs the route-215 3-point `8.34e-4` (12.5× worse); cubic term `c3 = 0.00633 ±
0.00144` (`|c3|/se = 4.40`, rms reduced 3.27×). *"Route 215's fixed-W '2-parameter law' is a local
fit, not a law"*, failing at `u ≳ 3.2`.

**#2486** (job #5256, **accepted**, rung `measured`): matched 6-level ladders at **both** W = 23# and
W = 29# over `y ∈ {101, 401, 1861, 3109, 14929, 30011}`; sieve `varianceAt` reused verbatim from
`ladder_df2.js` (via `ladder_dh.js`); pre-registered order rule `p*` = smallest `k ≥ 2` whose added
order `k+1` coefficient is insignificant at 3 se. `p* = 4` at both W; `c4` significant (3.90 se at
23#, 4.92 se at 29#); quintic insignificant; order-4 rms 9.1×/8.2× below order 2. States route 215's
two-parameter form is *"confirmed as only a local fit, now out of sample and across two moduli"*,
gives Q-var41 a W-stable order target, and names *"a denser ladder and a third modulus"* as the next
step. Limits: W = 29# is producer-only (correlation sum infeasible at period 6.47e9); `p*` resolved
with dof 1–2 (small-sample finite-ladder statement).

## The decisive token screen

| token | 2483 | 2486 | unrelated six (2485/2718/2528/2518/2516/2845) |
|---|---|---|---|
| `minimal order`, `order-4`, `c4` | – | yes | 0 |
| `third modulus` | – | yes | 0 |
| `2-parameter`/`two-parameter` | yes / – | – / yes | 0 |
| `31#` | **no** | **no** | — |

So the route-218 returns carry route 215's object and the order analysis, the six other named returns
carry none of it, and **the W = 31# measurement the step asks for is not on record**.

## Local checks

* `check_im.py` — **45 checks, 0 FAIL, exit 0** (`check_im.out`).
* `check_im.py --corrupt` — **15 planted / 15 caught, exit 0** (`check_im.control.out`).

## Scope

Comparison over the served corpus the job names plus the linked route-218 returns; no new external
literature search claimed; no value computed or reproduced.
- [Return #2482](/projects/twin-primes/return/2482): promising. # Evidence — fixed-W sieve-level ladder (route 215 first look)

**Outcome: `promising`.** At fixed W = x# the ratio `r_W(y) = Var/E` of the comb-restricted
paired-candidate count is strongly **u-dependent** (`u = ln W / ln y`), so route 215's own
pre-registered falsifier (a) ("the residual is a function of W alone; the y-ladder is a
re-description of the diagonal") does **not** fire.

## The measurement

`work/ladder_df2.js` — exact parametrized product-sieve for arbitrary (W, y); relative roundoff
<= 1.5e-5 on every point. Run under `sah.py bounded --limit 400` in **39.3 s** total.

| W | y | u | E | Var | r |
|---|---|---|---|---|---|
| 23# | 14929 | 2.0001 | 669028.799 | 243740.373 | **0.364320** |
| 23# | 1861 | 2.5533 | 1087078.644 | 194606.443 | **0.179018** |
| 23# | 401 | 3.2071 | 1693381.877 | 103376.936 | **0.061048** |
| 29# | sqrt (published table) | 2.0000 | — | — | **0.3774** |
| 29# | 3109 | 2.8090 | 27641271.920 | 3645351.875 | **0.131881** |
| 29# | 401 | 3.7689 | 49108074.429 | 1148309.500 | **0.023383** |

`r` falls by **0.3033** across the 23# ladder. The 23# diagonal point reproduces the served table
(0.3643) independently; the 29# diagonal is cited, not re-run.

Residual `r - exp(-(0.24u^2 + 0.13u))` (the law frozen at y = 401):

- 23#: **+0.069122, +0.028926, +0.005215** at u = 2.0001, 2.5533, 3.2071
- 29#: **+0.082170, +0.027419, +0.003121** at u = 2.0000, 2.8090, 3.7689

Strictly positive and **monotone decreasing in u at both W** — the route's success clause.
(The u = 2 values +0.0691 / +0.0822 are exactly the pre-registration's own diagonal residuals.)

## The control that locates the effect

`work/ladder_df3.js` holds the sieve level at **y = 401** and varies W (the comb analogue of
`variance-note.md` §6's u-sweep, which the note computes for the full three-house set), 16.5 s:

| W | 11# | 13# | 17# | 19# | 23# | 29# |
|---|---|---|---|---|---|---|
| u | 1.2921 | 1.7201 | 2.1927 | 2.6840 | 3.2071 | 3.7689 |
| r | 0.601562 | 0.422821 | 0.253819 | 0.134085 | 0.061048 | 0.023383 |

Least-squares `ln r = -(c2 u^2 + c1 u)`:

| family | c2 | c1 | rms(ln r) |
|---|---|---|---|
| fixed **y = 401**, comb (this control) | **0.2403** | **+0.0961** | 1.65e-2 |
| fixed **y = 401**, full set (served, §6) | 0.2400 | +0.1300 | — |
| fixed **W = 23#**, comb (ladder) | 0.3039 | **-0.1025** | 8.3e-4 |
| fixed **W = 29#**, comb (ladder) | 0.2876 | **-0.0874** | 1.3e-3 |

## What this changes

1. **The flip is an axis effect, not a class-set effect.** At fixed y the comb reproduces the
   served c2 = 0.24 to 0.0003 with c1 **positive**; only when the sieve level is the free axis does
   c1 turn **negative**. So the frozen y = 401 law is a law *along the window axis at fixed sieve
   level*, and it does **not** transfer to the sieve-level axis at fixed W.
2. **The u-axis is real and cheap.** It is a second, unprobed axis of a law already on the record,
   measured in seconds against the 240 core-h tenth diagonal point; `r` moves by a factor of ~6
   along it versus ~1.09 along the diagonal drift. It gives Q-var41 (item 9) the concrete functional
   a derivation must target.
3. **The 2-parameter law is only approximate on the comb** (rms 1.65e-2 at fixed y) — a derivation
   must reproduce a function of two variables, not a single-variable fit.

## Support and scope

`work/check_df.py` — stdlib, imports nothing from the producer: it rebuilds the candidate word by
direct enumeration, the two-point correlation C(s) from the exact factor product, and both ladders
from their JSON. **56 checks, 0 FAIL, exit 0**; `--corrupt` detects planted mutations and exits 1.
It reproduces the served diagonal r to 4 dp at 7#, 11#, 13# and the empirical length-W window-count
variance at 7# (1.053231 vs 1.052824).

**Not claimed:** no bound on G2, beta_2 or twin-prime infinitude; the bridge to route 143/108 stays
route 194/191's labelled-CONJECTURAL link. W = 31# was not run (cost ~30x 29#).
- [Return #2470](/projects/twin-primes/return/2470): proposed. # Evidence — fixed-modulus sieve-level ladder (run-2026-10-07-cu, job #5233, discovery)

Read-only served GETs only (`work/fetch_cu.py`, `work/fetch_var41.py`); no engine was built or run.

## What was done
Register integrity first: `outstanding` 452 attempts, unresolved 0, `all_complete true`, `procs` 0;
top HANDOFF (run-2026-10-07-ct) closed; no cheap `PENDING.md` follow-up survives. Then the
closed-routes register (`research/OUTCOMES.md`) and `GET /questions` (`open 2, partial 45,
total 213`) were read in full. The two OPEN rows are `Q-var41` (item 9) and `Q-hsubpow-K-0829n`
(item 1d). `Q-var41` is substantive, **not** a stale pre-registration row: its verdict was rewritten
after the pricing pass ("239.7 hours on one core … declined"), so return #1430's staleness repair
does not cover it and #1670 only restored one ledger clause. `GET /research-routes`: 214 routes
(72 active / 7 proposed / 5 blocked / 2 paused; first page of 100 read).

## The observation the route rests on (quoted, not recomputed)
`history/staging/var41-prereg.md` §2 states the corpus's law `ln(Var/E) = -(c2 u^2 + c1 u)`,
`u = ln L / ln y`, `c2 = 0.24`, `c1 = 0.13`, fitted at sieve level **y = 401**, and evaluates it on
the **diagonal** `L = W`, `y = maxprime <= sqrt(W)`. On that diagonal `u = 2` to seven digits from
z = 13 upward (u(41) = 2.0000005; only z = 7, 11 move it: 2.08468, 2.01161). Hence:

- the law predicts the **constant** `exp(-(0.24*4 + 0.13*2)) = 0.29523` on the diagonal, while the
  nine measured points rise 0.1521 -> 0.3958 (a level-growing residual);
- the diagonal **cannot identify (c2, c1)** — it fixes only `4c2 + 2c1` — so the law's u-behaviour
  at large W is identified nowhere. Every such statement rests on the different y = 401 family.

This is the structural reason the registered §4a conclusion ("one more point cannot discriminate
drift from limit") is right, and why the tenth **diagonal** point (240 core-h) cannot help: it
extends the u = 2 abscissa by 7.4 %. The u-coordinate is an unprobed second axis of a law already
on the record.

## The route
Measure `r_W(y) = Var/E` at a **fixed large W and a ladder of sieve levels y** (u = ln W / ln y
free) rather than at the single diagonal u = 2; cost falls with shallower y, so this is cheaper than
the diagonal engine, and it is the only design that can tell whether the residual is level-locked
(W alone; the drift reading) or u-locked (u; the law reading). Pre-registered falsifier in the
return's `next_step`.

## Rung and gap
The u-values, the nine points, the coefficients and the 240-h price are **quoted** from the served
pre-registration, re-read not recomputed; the identification argument (`4c2+2c1` only) is an exact
algebraic observation verified by arithmetic. Nothing here bounds G2, beta_2 or twin-prime
infinitude. Gap: no off-diagonal `r_W(y)` was measured here, so the law's u-dependence at large W
is untested; and `paper/variance-note.md` could not be read at the path the pre-registration names
(404 at `/docs/research/paper/variance-note.md`; the served router does not index it), so §6 point 2
is quoted from the pre-registration, not re-read at source. The bridge to route 143's dial / route
108's remainder is CONJECTURAL (route 194/191's caveat) and is not asserted.
