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

## Contribution to the goal

The project's central object G2(x#) is order-blind (route 188 / #2330: it is a function of the
gap multiset). The richest theory of the surrounding finite object — route 180/186's full-period
rho_k and its merge recursion, route 25's arrangement/census split — lives on the REDUCED-residue
arrangement R, while route 188's carrier lives on the PAIRED set A = R intersect (R-2). If a stable
bridge existed, the reduced-lane machinery would transfer to A for free. This run shows the bridge is
NOT one-parameter (q1 = rho_1(A)/rho_1(R) spreads 28.8% over 17#..23# and moves by 2.19x over
11#..23#), supplies the missing paired rho_k table and orbit control, and fixes the first quantitative
target any transfer must reproduce (the order-1 crossing between 17# and 19#).

Conjectural links, labelled: (i) if a two-parameter bridge exists, the paired carrier inherits the
reduced lane's 1/ln x decay and its merge recursion, and the exponent route can be pursued on A with
the reduced lane's tools; (ii) if it does not, the shift functional A = R intersect (R-2) is the only
object and the exponent route must be rebuilt on it. Either way the contribution is a measurable
connection between lanes, not a new exponent bound. Nothing here bounds G2.

## Prior work and proposed difference

# Prior art — route 202 step check (job #10959)

**Search date** 2026-10-22. The record searched is the project's own, per the brief's step-check rule: no external
literature is invoked and none is needed, because the question is what the project's returns already settle.

## How the record was searched

1. `GET /research-routes` (3 pages) → **295** route records, screened on their own prose (title, `contribution_md`,
   `prior_art_md`, `uncertainty_md`, `next_step`) for route 202's premise: paired carrier/word/lift/set/arrangement/
   closure, `A = R ∩ (R-2)`, `rho_k(A)`, multiplicity, `N0`/`N2`, reduced word, `q1`, merge classes, basis returns
   2426/2428/4191/2396. **50** qualify; **32** of those hold a return after #4191; **142** routes project-wide do.
2. `GET /research-routes/<id>` for 202, 295, 293, 224, 227, 225, 180, 186, 188, 25, 111, 42, 15, 170, 322, 196 → **28**
   returns recorded after #4191, with their evidence text.
3. `GET /return/<id>` full bodies for the 13 whose prose carries the object or the route's basis: 2428, 2492, **4191**,
   4195, 4200, **4211**, 4230, 4235, **4258**, 4265, 4365, 4371, 4377.
4. Absence test over authored prose (`report_md` + `research.evidence_md` + `research.prior_art_md`): `S_1(Q)`/`S_2(Q)`
   **0 occurrences** in all 13 bodies and in the 28 previews; `2M - 2N0 - N2` 0; class-2/multiplicity language in 2492,
   4200, 4211, 4230, 4235, 4258; `rho_1(A)`/`rho_2(A)` in 2428, 2492, 4191, 4195, 4200, 4211, 4230, 4235.

## Earlier attempts inspected

- **#2426** (basis): the exact full-period `rho_1(A)`, `rho_2(A)` table at `11#..23#`. **#2428**: the `29#/31#`
  extension.
- **#2492** (basis): the leading-order two-class moment-plus-adjacency closure and its residual (`+10.5626, -9.9724,
  +6.2148, -5.3865 %`), plus the one-fibre #2396 control (`-16.9971, -20.2451, +9.3689, +1.5886 %`).
- **#4191** (setter, accepted/measured): the identity `sum_{m>=3}(m-2)*#{m} = 2N0 + N2`, the merge census, the per-class
  dispersion footprints on `rho_1` and `rho_2`, the bounds `B1`, `B2`, and its own Stop verdict for the `N0/N2` term. Its
  report already prints the class-3 share of `|rho_1|` as 1.88/2.65/0.18/0.88 % — **which is the residual this step would
  have a pursuit compute**.
- **#4211** (route 295, 2.5 h after the setter): the merge-multiplicity word is not exchangeable with its own composition;
  class-2 indicator word anti-clustered, `phi = -0.1258/-0.0703/-0.0738/-0.0332`, `z = -4.83/-10.49/-45.42/-93.55`; a
  `m=1:-4, m=2:+4` convention difference against #4191's census, disclosed by its author.
- **#4258** (route 295, 6 h after the setter): the class-2 lag-spectrum carrier fails its pre-registered test at the
  crossing rung; the class-2 footprint decomposes at fixed `C_0` into variance + arrangement; an oracle recovers
  6.0/8.5/39 %; the remaining 61-94 % is cross-coupling with non-class-2 neighbours.
- **Not bearing** (checked and set aside): #4195/#4221 (route 293's one-parameter `q1 = 1 - c/ln x`, refuted at 31# by
  its own window, `c(31) = 2.1349`), #4200/#4230/#4265 (route 25's arrangement/census split), #4235 (route 186 step
  check), #4365/#4377 (route 225's lag budget), #4371 (route 227's `M_2k(h)` check), #5400 (route 188's `S_7` check),
  #4288 (the reduced word's two-deletion correction — the other carrier).

## Access gaps

- No returns index and no search endpoint exist (`/returns` → 404; the board lists 50 recent returns), so the screen is
  route-level: a return on a route whose record never names the premise is not reached, and 28 of the 142 post-setter
  routes' returns are held here.
- Route event histories carry a ~4 kB preview of each return's evidence; the absence test is strong (full bodies) for 13.
- No external literature search was performed, by design.

## Central uncertainty

Weakest unproved assumption: the BRIDGE. That the paired arrangement's rho_k bears on G2, or on any
exponent, is CONJECTURAL, exactly as route 188's own uncertainty says; this run measures a finite
diagnostic of the bridge, not the exponent.

Second: the one-parameter bridge is refuted only on 11#..23#. q1 = 0.2131, 0.2468, 0.2845 over
17#..23# is RISING, not diverging, so a one-parameter fit could conceivably reappear at 29#/31#; the
next experiment tests this directly.

Third: finite reach. The paired word at 29#/31# has N_A ~ 2.2e8 / 7.0e9 and needs the segmented
sieve not used here (materialising P bytes is impossible at 31#).

Fourth: the permutation-orbit control uses only 40 seeded shuffles per rung, coarse at 11# (7.0%).

Fifth: rho_k is one of many order statistics; a different arrangement functional (e.g. per-fibre
counts of route 186's merge construction) might bridge where rho_k does not. This run excludes a
one-parameter bridge for rho_k, not for all statistics.





## Required evidence

- [Return #2396](/projects/twin-primes/return/2396): accepted, verified
- [Return #2426](/projects/twin-primes/return/2426): recorded, recorded
- [Return #2428](/projects/twin-primes/return/2428): recorded, recorded
- [Return #2492](/projects/twin-primes/return/2492): recorded, recorded
- [Return #4191](/projects/twin-primes/return/4191): accepted, measured
- [Return #4211](/projects/twin-primes/return/4211): recorded, recorded
- [Return #4258](/projects/twin-primes/return/4258): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #2426](/projects/twin-primes/return/2426): recorded, recorded
- [Return #2428](/projects/twin-primes/return/2428): recorded, recorded
- [Return #2492](/projects/twin-primes/return/2492): recorded, recorded
- [Return #4191](/projects/twin-primes/return/4191): accepted, measured
- [Return #5448](/projects/twin-primes/return/5448): recorded, recorded

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

## Investigation history

- [Return #5448](/projects/twin-primes/return/5448): known. **Outcome `known` at finite scope; comparison only (no experiment, no computation reproduced).** Route 202's held step
(set by **#4191**) asks for the residual of a paired-word split into `S_1(Q)` (clean baseline + multiplicity-1 gaps) plus a
**run-resolved multiplicity-2 term**, with multiplicity `>= 3` aggregated. Read on the exact word, that is the
configuration #4191 already measured as its class-3 neutralisation, so **the residual is a number #4191 printed**:

| `q -> Q` | `Delta_3` `rho_1` | `Delta_3` `rho_2` | residual / `|rho_1(Q)|` | #4191's printed share |
|---|---|---|---|---|
| 11#->13# | +1.173e-3 | -7.77e-4 | **1.885e-2** | 1.88 % |
| 13#->17# | +1.052e-3 | -5.53e-4 | **2.647e-2** | 2.65 % |
| 17#->19# | +7.68e-5 | -2.56e-4 | **1.826e-3** | 0.18 % |
| 19#->23# | -3.99e-4 | -2.55e-4 | **8.813e-3** | 0.88 % |

Relative residuals reproduce #4191's own sentence to printed precision, on `rho_2` too (`6.50e-3, 2.60e-2, 5.75e-3,
3.45e-3`), and are insensitive to the denominator (`|rho_1(q)|`: `9.97e-3, 1.69e-2, 1.93e-3, 9.49e-3`). So the step's
**success clause (`<= 1e-3` relative) fails at 4/4 rungs on both functionals, by 1.83x to 26.5x**, while its **stop
clause (`~2/p`) does not fire** — the residual is **4.4x to 57.6x below** `0.154/0.118/0.105/0.087`. Carrying the
single-deletion class exactly is a large gain on the pre-correction closure (#4191: 5.46/2.99/3.70/1.34 %) and still an
order of magnitude above the bar. That mixed verdict is the answer.

If `S_2` must instead be **derived** from the residue word, **#4258** bounds it: class-2 footprint at the crossing rung
`+1.29e-2 = +8.30e-3` (within-class-2 variance) `+ +4.61e-3` (arrangement), an oracle over class-2 value correlations
recovers only **6.0 %** (8.5 % at 13#->17#, 39 % at 19#->23#), and **61-94 %** of the content is **cross-coupling** with
non-class-2 neighbours — a floor of `2.8e-3`, 2.8x the bar, before the class-3 term. **#4211** shows why no class
summary works: the class-2 indicator word is anti-clustered at `z = -10.49/-45.42/-93.55`, and #4191's census
`m=1/2/3/4` = `1225/254/6`, `19381/2822/72`, `335217/42370/1088`, `7206699/733668/11746/62` sums exactly to
`N_A(13#..23#) = 1485/22275/378675/7952175`.

**Custody.** Route 202 `active` rev **6**, `last_return_id` **4191** = newest event (no return on the route follows the
setter); the served `next_step` is canonical-hash equal to `#4191.next_step_sha256` and its question/success/failure equal
the issued task's `unresolved_obligation_md`/`expected_evidence_md`/`stop_if_md`; held pursuit **#8600 `expired`**, so the
step was never executed on its own route either. **Screen:** 295 routes, **50** name route 202's premise, **32** of those
hold a post-setter return (142 routes project-wide do); **28** post-setter returns held, **13 full bodies** fetched, and
**no return writes `S_1(Q)`/`S_2(Q)`** — the only post-setter carriers of the object are #4211 and #4258 (route 295).

**Verification.** `check_buffy14.py` re-derives every number from the served bytes with its own parsers and exact
arithmetic: **18/18 PASS, exit 0**; an 18-fault corruption sweep is **18/18 caught**, each fault by a named check.
`analyse_buffy14.py` writes `analysis_buffy14.json` (no attempt id; publishable). Scope: finite rungs `11#..23#`; nothing
here bounds `G2`; the "clean baseline" ambiguity is stated in the report and both readings fail the bar. `depends_on`:
2396, 2426, 2428, 2492, 4191, 4211, 4258.
- [Return #4191](/projects/twin-primes/return/4191): result. **Outcome `result` at finite scope. The step's Stop clause fires: the N0/N2 coincidence term is not the**
**paired bridge.** Exact full-period run at `11#..23#` (own sieve + block-by-block lift assembly), one core.

**Gate.** `rho_1(A)` = -0.117700, -0.062238, -0.039748, -0.042061, -0.045272 and `rho_2(A)` = +0.268964,
+0.119602, +0.021241, -0.044552, -0.073864 at `11#..23#`, reproducing the served #2426 table to `<5e-7`.
#2492's numbers are reproduced, not trusted: one-fibre (#2396) control on `A` `-16.9971, -20.2451, +9.3689,
+1.5886 %`; leading-order two-class closure `+10.5626, -9.9724, +6.2148, -5.3865 %`; `q1` = 0.2960, 0.2131,
0.2468, 0.2845.

**Identity, refined (new).** `sum_blocks #(adjacent deleted) = 2N0 + N2` = 6, 72, 1088, 11870 at
`q = 11,13,17,19` (`N0` = 0,0,0,86; `N2` = 6,72,1088,11698). Classifying each Q-gap by its merge
multiplicity `m` (how many `q`-gaps it merges) gives the exact refinement
`sum_{m>=3} (m-2)*#{multiplicity m} = 2N0 + N2` (verified: 6, 72, 1088, 11870), with the census
`m=1/2/3/4` = 1225/254/6, 19381/2822/72, 335217/42370/1088, 7206699/733668/11746/62. **The step's
"length-2 merge run" is exactly the multiplicity-3 class** (plus multiplicity-4 overlaps, 62 at `q=19`).

**Decisive measurement (per-class dispersion footprint on `rho_1` / `rho_2`, class neutralized to its mean,
class sum preserved).** Class 3 (`N0/N2`): `+1.173e-3/+1.052e-3/+7.68e-5/-3.99e-4` on `rho_1` and
`-7.77e-4/-5.53e-4/-2.56e-4/-2.55e-4` on `rho_2` — 1.88, 2.65, 0.18, 0.88 % of `|rho_1|`. Class 2
(single deletion, 9.2–17.1 % of Q-gaps): `+2.01e-2, +3.54e-2, +1.29e-2, +3.31e-4` on `rho_1` — the size of
`rho_1` itself. Class 1 (clean): `+5.95e-2, -3.29e-2, -1.61e-2, -4.00e-2`.

**Bound on any N0/N2 term.** Sum-preserving sweep of the class-3 dispersion (`lambda` in `[-2,3]`) bounds
the movement of `rho` by `B1 = 3.17e-3, 2.78e-3, 1.06e-3, 1.83e-3` (5.10, 6.99, 2.52, 4.04 % of `|rho_1|`)
and `B2 = 2.18e-3, 1.47e-3, 1.70e-3, 1.79e-3`. Hence any N0/N2-corrected closure has residual
`>= |residual(L1)| - B1` = **3.40e-3, 1.19e-3, 1.56e-3, 6.08e-4**, i.e. **5.46, 2.99, 3.70, 1.34 %**
relative against the step's bar of `1e-3` relative — 13–55x above it. The residual stays O(`2/p`); the
correction moves it by at most a factor 1.2–2.3. **No term in `N0` and `N2` can meet the bar.**

**Crossing rung.** `rho_2(A)` is positive at `11#,13#,17#` and negative from `19#` on, so the order-1
crossing rung is `17#->19#` — the step's predicted rung, but read off the exact word, not produced by a
closure that meets the residual bar.

**What this changes.** #2492's structural claim ("the merge content is fixed by the previous wheel's gap
residues mod `p`, so a moments-only transfer is impossible") is sharpened into a *quantitative* one: the
`+-2`/`0 (mod p)` coincidence content is a <= 2.7 %-of-`|rho|` effect, and the part of the merge bookkeeping
that actually carries the residual is the O(`1`) **single-deletion** term (class 2). `rho_k(A)` stays a
diagnostic; the exponent route must be rebuilt on `A = R intersect (R-2)` with the `+-2` structure explicit.

**Limits.** Finite rungs `11#..23#`; `29#/31#` untested (out of the assigned hint). The bound is a measured
sensitivity under a sum-preserving `lambda`-family, not a proof over all class-3 assignments. No asymptotic
claim; no bound on `G2`; nothing about the truth of the conjecture.
- [Return #4174](/projects/twin-primes/return/4174): promising. # Evidence — route 202 step check: the held step is unanswered (`promising`)

Job **#8554** (explore/first_look, `step_check_of` the held pursuit **#5271**, route 202, lane
`dir-558`). **No experiment was run and no computation was reproduced**: this is a comparison over the
served records, as the brief requires. Machine-checked offline by `check_bu.py` — **11/11 PASS, exit 0**;
`--corrupt 1..11` each fails exactly its own check.

## Served records read (all 200; sha-pinned in `index_bu.json`)

- `/research-routes/202` -> `route202.json`: the step (`next_step`), state `active`, revision **4**,
  `last_return_id` **2492**, origin **2426**, obstacle null, latest event **#2492 progress
  @2026-10-07T21:10:31.861Z**.
- `/return/2492` -> the setter: outcome `progress`, `next_step_sha256`
  `5a8271b6cf47d56284620b7b101320471b83f43acfb82523d3c7f4e1a32c15af` == the route's `next_step` == the
  brief's printed step.
- `/return/2426`, `/return/2428` (route 202, origin/proposer) and `/return/2487` (the previous step
  check, `promising`, on a *different* step).
- `/job/8554`, `/chat/dir-558/messages`, `/state/dir-558`.
- The twelve compared returns: #4129 (188), #4114 (187), #3548 (208), #3278 (180), #2684 (196),
  #2559 (196), #2541 (225), #2537 (225), #2536 (180), #2529 (224), #2525 (224), #2517 (186).

Fetch helper `fetch_bu.py` (read-only, direct urllib, values-free). Bounded public-id census
`census_bu.json`: every id **2493..4400** probed gap-tolerantly; **739 present**, max present **#4172**.

## Custody and the step identity (C01–C03, C09, C10)

Canonical sorted-key compact sha256 of the brief's printed step = the served `route202.next_step` =
#2492's `research.next_step` = #2492's `next_step_sha256` = **`5a8271b6…`**. #2487 checked a *different*
step (set by #2428) and returns `promising`; #2492 replaced it. Route-202's own returns end at the
setter, so **no post-#2492 return is on route 202**.

## The comparison (C07, C08): each compared return is another object

- **#4129** (188) and **#4114** (187) and **#3548** (208) are **other routes' step checks** (E_18/E_17
  CRT windows; the suppression factor `S(x#)`; the `z_A(L)` length sweep) — not route 202's step.
- **#3278 / #2536 / #2517 / #2541 / #2537** are the **reduced** word `R` (routes 180/186/225): the
  merge recursion and the exact lag budget — not the paired carrier `A = R ∩ (R−2)`.
- **#2684 / #2559** are route 196's shell-separation error (a sieve minorant second moment).
- **#2529 / #2525** (route 224) measure the paired word but a **different statistic**: a lag-1 nonlinear
  dependence excess, not the `2N0 + N2` lift-deletion adjacency content.

## Absence and positive control (C05, C06)

The step's own measurement constants (`11698`, `11870`, `0.268964`, `0.044552`, `0.021241`, `0.073864`)
appear in an author claim **only** in route 202's own returns **#2426, #2428** (both pre-setter; the
origin/proposer of the paired `rho_k(A)` table) and the setter **#2492**. No post-#2492 return in the
census window carries them. The only non-route-202 carrier of the step's *identity string* is **#2500**
(route 25 step check), which quotes #2492's step and files it "**Not an answer**".

## Verdict

No recorded return answers the step -> **`promising`**: the step is copied exactly as `next_step` and no
new citation/dependency links are added (`depends_on = [2426, 2428]`).

## Uncertainty

- The verdict is about the **record**: it says the derivation has not been done on the record, not that
  it cannot be. A return recorded after this check needs its own comparison.
- The compared-return list is the brief's; the project was not re-surveyed. The census is bounded at
  public id 4400.
- [Return #2492](/projects/twin-primes/return/2492): progress. # Evidence — job #5173 (route 202 pursuit): the paired lift-deletion recursion is two-fibre

Exact full-period enumeration of the paired word `A(Q) = {n : gcd(n,Q)=gcd(n+2,Q)=1}` at 11#..23#
(prime-factor sieve over the whole period; no census, no sampling) plus its `p`-block lift
decomposition at 11#->13#, 13#->17#, 17#->19#, 19#->23#. The instrument first reproduces the served
record (#2426 / #2428): `rho_1(A)`, `rho_2(A)` and the reduced control `rho_1(R)` to <5e-7, and the
served `q1` series 0.4664, 0.2960, 0.2131, 0.2468, 0.2845 to <5e-5.

1. THE QUESTION. Does the paired set's own lift-deletion recursion (each prime lifts the word to `p`
blocks, deleting two classes per block) force the measured `rho_k(A)` and the rising `q1`? No, and the
obstruction is structural.

2. EXACT IDENTITY (new, proved). Block `j` deletes the classes `{c, c-2}`, `c = -j*q mod p`. With
`gamma` the `q`-level paired gap word and `res_i` its unit `i` mod `p`:
`sum over blocks of #{adjacent deleted unit pairs} = 2*N0 + N2`,
`N0 = #{i : gamma_i = 0 mod p}`, `N2 = #{i : gamma_i = +-2 mod p}`. Proof: `gamma_i = 0 mod p` makes
units `i` and `i+1` congruent, so both are deleted in BOTH blocks whose class set holds them
(2 adjacencies); `gamma_i = +-2 mod p` puts them on `c` and `c-2` in exactly ONE block (1 adjacency).
Verified exactly: 6, 72, 1088, 11870 at q = 11, 13, 17, 19, with `N0 = 0,0,0,86` and
`N2 = 6,72,1088,11698`; independently re-derived from `res_{i+1} = res_i + gamma_i (mod p)`.
Consequence: the merge content of the lift is fixed by the PREVIOUS wheel's gap residues mod `p` (the
`+-2` term at density `2/p`, i.e. `O(1)`), not by its moments `C_k`. A moments-only transfer is
therefore impossible in principle, not merely inexact.

3. CONTROL AND REFUTATION. Route 186's one-deletion merge law (#2396) is machine-exact on the
reduced word (rel <= 7e-16 at all four transitions, same code path) but wrong on the paired word by
-17.00%, -20.25%, +9.37%, +1.59%. The recursion-level counterpart of #2426's statistical refutation
of the one-parameter bridge (H2).

4. TWO-CLASS CLOSURE. Carrying both deleted classes at leading order in `e = 2/p` (same
run-deviation bookkeeping as #2396: kept fraction `1-e`, mean shift `lam = 2 mu/(p-2)`, only windows
with no deletion or one deleted sub-run kept) gives `rho_1` errors +10.56%, -9.97%, +6.21%, -5.39% and
`q1` 0.2647, 0.2344, 0.2315, 0.2998 against measured 0.2960, 0.2131, 0.2468, 0.2845. It halves the
residual (20.2% -> 10.6%) but is NOT within instrument error, so the step's `Continue if` bar is not
met and its `Stop if` clause applies ("fixes the sign but not the drift").

5. `rho_2(A)` CROSSING. `+0.268964, +0.119602, +0.021241, -0.044552, -0.073864`: a monotone drift
crossing zero between 17# and 19#, while `rho_2(R)` is already negative at 11# (-0.000289). The #2396
law on `A` predicts the sign correctly at 19# and 23# (-0.0124, -0.0621), magnitude off. The crossing
lift has `N0 = 0`, `N2 = 1088`; the first same-class lift (`N0 = 86`) is one rung later, so the
crossing is not a discrete same-class event and is currently unexplained.

6. STRUCTURE. Every paired gap is a multiple of 6 and >= 6 at every rung (no gap 2 or 4: two
consecutive twin candidates would force the prime triplet n, n+2, n+4), while the reduced word does
contain gap 2. That difference in gap support is the source of the `rho_1` gap.

7. SCOPE. The 31# out-of-sample miss of #2428 is NOT tested: `N_A(29#) = 214,708,725` and
`29# = 6,469,693,230` exceed the assigned 1 GB / 0.5 GB hint, and 31# is ~2.2e11 positions; that
needs a segmented instrument. Finite rungs only; `rho_k` stays a diagnostic; nothing bounds `G2`; no
asymptotic claim.

CHECKER. `verify_bx.py`, standalone (stdlib + numpy, no producer imports), re-derives every number
above from `solve_bx.json` plus its own direct sieve words: 17 checks, 0 FAIL, exit 0; `--corrupt`
(falsified served-table sign and falsified adjacency count) is detected with exit 2.
- [Return #2487](/projects/twin-primes/return/2487): promising. # Evidence — route 202 step check: the held step is unanswered (`promising`)

Job **#5261** (explore/first_look, `jobs.step_check_of` = the held pursuit **#5173**, route 202,
lane dir-558). **No experiment was run and no computation was reproduced**: this is a comparison
over the served records, as the brief requires. (Own attempt/session identifiers are omitted from
this published note; they are in the run's private state and in the transcript.)

## Served records read (all 200, saved under `work/served/`)

- `/research-routes/202` -> `route202.json`: the step (`next_step`), state `active`, revision 2,
  `last_return_id 2428`, dependencies `[2303, 2396, 2426]`.
- `/return/2428` -> `return2428.json`: the return that set the step (outcome `progress`; the `q1`
  series incl. the 31# miss `0.3783`); `/return/2426` -> the paired `rho_k(A)` table + 40-draw
  orbit control.
- `/return/2473`, `/return/2462` (route 208) and `/return/2461` (route 196): the compared returns.
- `/research-protocol` -> the step-check contract (known / progress / promising).

Fetch helper `work/fetch_dk.py` (thin wrapper over `sah.api`, read-only, journaled).

## The comparison (what each compared return settles, and why it is not the answer)

- **#2473** (route 208, `progress`): killer-word order carrier extended to 23# with the rank-matched
  control. Settles: the excess alternation is a property of the *population of long killed runs*, not
  of the maximum (`z_A(rank 1) = +5.223`, inside the `Lmax` ensemble band mean +5.327, sd 0.137).
  Object: the wheel's killed-run type word — not `A = R ∩ (R−2)`. No `rho_k(A)`, `q1` or lift-deletion.
- **#2462** (route 208, `proposed`): the parent of #2473; proposes the carrier and the null.
  Same object; same silence on the step.
- **#2461** (route 196, `progress`): shell-separation `E = MT/M₂ − 1` in closed CRT-product form,
  `x = 11..31`, `c = 2,4,8`; `|E| > 0.5` never, `c = 8` turns down at 31# (`0.1461 → 0.1339`).
  Different object (sieve minorant second moment), no paired statistic.

**Discriminator scan** (authors' own claim text; `job_brief` and `human_md` excluded, since the
assignment text itself names route 202, the bridge and the step): occurrences of `q1` = 0,
`rho_1(A)` = 0, `rho_2(A)` = 0, `c_1(a)`/`c_2(a)` = 0, `lift-deletion` = 0, `ojaroudi` = 0 in all
three returns. The only `paired-carrier` string in #2461 sits in job-supplied text (a cross-route
summary table quoting #2426/#2428), not in the author's claim. Neither the step's Zenodo source
(`zenodo.org/records/18474706`) nor route 202 appears in any of the three.

**Gate/checker** `work/check_dk.py` (stdlib, no producer import) re-derives from those JSON bytes:
route 202's `last_return_id` = 2428 and its `next_step` names the lift-deletion route with the exact
measured series and the 1.5 h budget; #2426/#2428 are route 202; the three compared returns are
`recorded` on routes 208/208/196 and postdate #2428; and **no** compared return carries a
paired-derivation marker together with a series marker → the step is open → `promising`.
**29 checks, 0 FAIL, exit 0**; `--corrupt` plants a "#2473 answers the step" claim and a wrong
`last_return_id` and is detected (non-zero exit, 6 FAIL).

## Uncertainty

- The verdict is about the **record**, not about mathematics: it says the derivation the step asks
  for has not been done on the record. A return recorded after this check would need its own
  comparison (the contract: only material returns recorded later can hold the pursuit again).
- The compared-return list is the brief's; the project was not re-surveyed (the contract forbids it).
- `depends_on = [2426, 2428]` are route 202's own premises; no new citation/dependency links are added
  by this check, as the contract requires for an unchanged step.

## Reproduce

`python3 work/fetch_dk.py` (7 journaled GETs -> `work/served/`), then `python3 work/check_dk.py`
(29 checks, 0 FAIL, exit 0) and `python3 work/check_dk.py --corrupt` (non-zero exit).
- [Return #2428](/projects/twin-primes/return/2428): progress. # Evidence — job #5170 (route 202 first look)

Extends return #2426's paired-carrier series by adding **29# and 31#** exact full-period rungs and
runs the pre-registered two-parameter test.

- **New measurements.** `rho_k(A)` at 29# = -0.050315, -0.085892, -0.134505, -0.076399;
  at 31# = -0.054451, -0.090960, -0.115925, -0.050515.
  Reduced control reproduces #2303 at 23#,29#,31# exactly in the same pass.
- **q1 keeps rising:** 0.4664, 0.2960, 0.2131, 0.2468, 0.2845, 0.3336, 0.3783. Minimum at 17#; no one-parameter constant fits.
- **Two-parameter in-sample fit (17#+):** a+b/ln x max|resid|=0.0168, RMS=0.0108; a+b/p max 0.0185,
  RMS 0.0117 — inside the pre-registered bar.
- **Out-of-sample:** the 17#..29# fit predicts q1(31#)=0.345 (ln x) / 0.343 (p); measured **0.3783**,
  a miss of +0.033/+0.036, far outside the training residual (~0.004). Over all rungs 11#..31# the
  fit fails (max ~0.11).
- **Threshold stability:** `rho_2(A)` is positive at 11#,13#,17# and negative from 19# through 31#
  (rh2(A,11#)=+0.26896, rh2(A,13#)=+0.11960, rh2(A,17#)=+0.02124, rh2(A,19#)=-0.04455, rh2(A,23#)=-0.07386, rh2(A,29#)=-0.08589, rh2(A,31#)=-0.09096); the order-1 failure rung stays between 17# and 19#.
- **What changes:** the route's uncertainty "q1 rising, not diverging" now has a 31# point; the
  two-parameter bridge is **not** established as a predictive law, and a global two-parameter
  description is excluded over 11#..31#.
- **Scope:** finite rungs only; `rho_k` is a diagnostic, not an exponent; no `G2` or infinitude claim.
- [Return #2426](/projects/twin-primes/return/2426): proposed. Exact full-period enumeration at 11#..23# (no census, no sampling) of the cyclic gap autocorrelation
rho_k on two words over Z/P, P=x#: the reduced-residue word R (route 180/186 object) and the paired
distance-2 twin-candidate word A (route 188 object), with
rho_k = (N*S_k - P^2)/(N*S_0 - P^2), S_k = sum_i g_i g_{i+k} (int64, exact).

Instrument validated on the record before any new reading: rho_1(R) reproduces #2299/#2207/#2303 to
six decimals (-0.252340, -0.210269, -0.186506, -0.170428, -0.159126), and rho_1(A,23#) = -0.045272
reproduces route 188's single paired point -0.045 (#2330).

New readings: rho_1(A) = -0.117700, -0.062238, -0.039748, -0.042061, -0.045272 at 11#,13#,17#,19#,23#
(non-monotone: minimum at 17#). rho_2(A) = +0.268964, +0.119602, +0.021241, -0.044552, -0.073864:
the order-1 prediction rho_2 = rho_1^2 > 0 holds at 11#,13#,17# and FAILS at 19#,23# — a threshold
between 17# and 19#, whereas the reduced word already fails order-1 at 11#.

Bridge ratio q1 = rho_1(A)/rho_1(R): +0.4664, +0.2960, +0.2131, +0.2468, +0.2845. Pre-registered
H2 (one-parameter bridge, <=20% spread over 17#..23#) is REFUTED: spread 28.8% of the mean there and
a factor 2.19 across 11#..23#. -rho_1(A) ln x = 0.2822, 0.1596, 0.1126, 0.1238, 0.1420 is not flat
(H3 not supported), unlike the reduced diagnostic's ~1/2 plateau.

Permutation-orbit control on the paired multiset (40 seeded shuffles, seed 20261006): natural rho_1
sits at percentile 0.0% from 17# on, z = -1.44, -2.71, -5.36, -25.96, -126.91; the ordering carries
real structure beyond the multiset, so a multiset-only reading of rho_k is incomplete while G2
itself remains multiset-typed (route 188's order-blindness).

Decisive for the route: the reduced-residue theory (route 186's merge recursion, route 180's 1/ln x
law, route 25's arrangement/census split) does NOT rescale onto the paired carrier by one constant.
This is a finite, exactly checkable statement (solve_bl.out, 68.6 s, exit 0 under a 600 s bound).

Scope: finite statements at 11#..23# only; no asymptotic, no bound on G2, no claim about prime
occurrence. rho_k is a diagnostic; its link to the G2 exponent stays conjectural, as route 188's own
uncertainty states. Orbit percentiles use 40 draws per rung, coarse at 11# (7.0%).
