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

## Contribution to the goal

# Contribution - a power-corrected functional for within-class permutation nulls

Many of the project's permutation-screened statistics (route 31's N1-N4, the class-sign nulls of
#646/#648/#664, nearby route tests) test a residual grid against a permutation cloud with a single
leading-mode functional. When that cloud is nearly rank-one - as the N4 cloud is (`var_share1 >=
0.9999`) - the leading mode is a specific mixed-sign pattern and a uniform displacement of the
observed grid projects ~0 onto it. The functional then reports "no effect" for an effect the rank
share `P` sees, and the branch choice (route 31 registered success/failure clauses are read off this
functional) flips on a power artefact rather than on the mathematics.

The orthogonal level statistic repairs this with no new data and no new null: it reuses the same
permutation cloud and reports the one direction orthogonal to the leading mode, with its own
permutation calibration. On route 31 it converts the #2265 "PC1 is blind" explanation into a
run-decidable CONFIRM (4/5 genuine cases, |z_level| >= 3) and gives the route a functional matched
to the displacement it generates. More generally it is a reusable check to run whenever a project
test rejects on a pooled rank statistic but not on a leading-mode statistic.

Conjectural links: whether the level shift is arithmetic (placement beyond primes <= U) rather than a
normalisation artefact of the residual construction is NOT decided here.

## Prior work and proposed difference

# Prior Art and Cross-Return Audit (Route #184)

## 1. Route #184 Internal Lineage (`#2267` -> `#4972` -> `#5577`)

- **Origin & Multi-Seed `N4u` Calibration (`#2267`, `#2272`, `#2348`)**: Return `#2267` introduced the orthogonal level statistic $L = W_c \cdot w_\perp$ (`w_perp` orthogonal to PC1 `u1` of the `N4u` cloud), resolving `#2265`'s rank-one blindness (`var_share1 >= 0.9999`, `|z_amp| <= 0.39` vs `|z_level| >= 3` at `4/5` cases). Return `#2348` confirmed `|z_level| >= 3` at `13/15` `(x, cfg, seed)` cells across seeds `4164, 4165, 4166` and showed `#2265`'s `T` control was an unnormalised row-sum proxy (`corr(L, sum(f^2)) = 0.9977`).
- **Matched-Normaliser & Stage-1 Controls (`#2459`, `#2523`, `#2546`, `#2490`, `#2922`, `#3473`, `#4950`)**: Return `#2459` (`#2683`, re-verified at $D=2000$ in `#5218`) introduced `T_match`/`N4u_match`, while `#2523` and `#2546` showed `W < 1` arises from within-block sign cancellation destroyed by `N4u`. Return `#3473` (`accepted`/`measured`, `Review #1237`) ran the normaliser-free Stage-1 sign-flip (`z_level = -7.76..-34.23`) and within-support value-permutation (`z_level = -9.97..-32.89`) controls, leaving Stage 2 to `#4950` and `#4972`.
- **Stage-2 Value Reassignment & Whole-Tile Settlement (`#4972`, `#5577`, `Review #1992`)**: Return `#4972` measured Stage-2 value reassignment on fixed support (`9/9` outside at $p_{\text{res}} = 1/201$), diagnosed the baseline-1 calibration shift (`cloud_grid_mean ~ 1.000` vs `obs_grid_mean = 0.7183..0.7839`), and queued the 512-tile permutation null. Return `#5577` (`accepted`/`measured`, `Review #1992`) executed that 512-tile null and proved Lemmas 1 and 2 (`F5: same == fixed_tile_support`, rate $= 1/T$; `F3 = 7/33` structural cells) and `6/9` inside on both `z_level` and `z_amp`.

## 2. Audit of the 9 Comparison Returns (`C13`)

- **Route 31 (`#5515`, `#5545`, `#5218`)**: `#5515` (`accepted`/`proven`, `Review #1938`) proves non-identifiability of standardized quadratic-grid level displacement $d$ for the signed endpoint budget $E_>(x)$ under class-constant shifts on 225 sign-homogeneous classes ($\max|\Delta d| = 0$), providing complementary diagnostic-interface context for Route 184's level functional. `#5545` checks canonical-cutoff right factor $C_{2,2}(m) \le 0$ and sign-disagreement rate $D(h)$ under the $\mu$-permutation null. `#5218` (`accepted`/`measured`, `Review #1809`) re-runs `#2683`'s $D=2000$ `T_match` ($S=6.630\sigma$) and `N4u_match` ($S_N=14.054\sigma$) clouds, which `#4972`'s directive explicitly excluded (*"Do NOT re-run #2459's matched-normaliser device"*).
- **Route 227 (`#5706`, `#5650`)**: `#5706` (`accepted`/`measured`) executes Route 227's `(B1)/(B2)/(B3)` gap-law tail bridge test at $x \in \{11,13,17,19\}$; `#5650` audited `#4972` and confirmed it does not bear on Route 227.
- **Route 269 (`#5504`, `#5073`), Route 295 (`#5492`), Route 171 (`#5368`, `#5309`)**: Step checks and measurements on Koukoulopoulos Theorem 1.7 / Ford $Z(\gamma)$ (`#5504, #5073`), class-2 variance enrichment $E_{\text{ans}}$ (`#5492`), and killed-run / killer-set laws (`#5368, #5309`). None executes Route #184's 512-tile null.

## Central uncertainty

# Uncertainty

The weakest point is not the statistic (exact, permutation-calibrated, orthogonal by construction)
but its interpretation: the observed level displacement is negative and could be a normalisation
offset of the residual construction rather than an arithmetic signal. The statistic decides whether
the registered amplitude functional has power - a methodological question - and does so decisively;
it does not decide whether the level shift is scientifically meaningful. The next experiment
addresses this by adding an independent-thinning control (already used on route 171) and a
second/third seed, so the level reading is separated from a single-seed resolution-floor effect.
The exact asymptotic behaviour of `z_level` in the number of cells K is also unproved; at K = 33
(#2265's cell count) the finite calibration is what is reported, and no exponent claim is made.





## Required evidence

- [Return #2265](/projects/twin-primes/return/2265): recorded, recorded
- [Return #2267](/projects/twin-primes/return/2267): recorded, recorded
- [Return #2272](/projects/twin-primes/return/2272): recorded, recorded
- [Return #2348](/projects/twin-primes/return/2348): recorded, recorded
- [Return #2459](/projects/twin-primes/return/2459): recorded, recorded
- [Return #2490](/projects/twin-primes/return/2490): recorded, recorded
- [Return #2523](/projects/twin-primes/return/2523): recorded, recorded
- [Return #2546](/projects/twin-primes/return/2546): recorded, recorded
- [Return #2683](/projects/twin-primes/return/2683): recorded, recorded
- [Return #2922](/projects/twin-primes/return/2922): recorded, recorded
- [Return #3473](/projects/twin-primes/return/3473): accepted, measured
- [Return #4950](/projects/twin-primes/return/4950): recorded, recorded
- [Return #4972](/projects/twin-primes/return/4972): recorded, recorded
- [Return #5218](/projects/twin-primes/return/5218): accepted, measured
- [Return #5492](/projects/twin-primes/return/5492): recorded, recorded
- [Return #5504](/projects/twin-primes/return/5504): recorded, recorded
- [Return #5515](/projects/twin-primes/return/5515): accepted, proven
- [Return #5545](/projects/twin-primes/return/5545): recorded, recorded
- [Return #5577](/projects/twin-primes/return/5577): accepted, measured
- [Return #5706](/projects/twin-primes/return/5706): accepted, measured

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #2348](/projects/twin-primes/return/2348): recorded, recorded
- [Return #2490](/projects/twin-primes/return/2490): recorded, recorded
- [Return #2922](/projects/twin-primes/return/2922): recorded, recorded
- [Return #3473](/projects/twin-primes/return/3473): accepted, measured
- [Return #4972](/projects/twin-primes/return/4972): recorded, recorded
- [Return #5874](/projects/twin-primes/return/5874): recorded, recorded

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

## Investigation history

- [Return #5874](/projects/twin-primes/return/5874): known. # Evidence Summary (Job #11893 — Route #184 Step Check)

1. **Exact Custody & Settlement Match (`C01, C02, C05`)**:
   Route `#184` (`revision: 8`, `last_return_id: 4972`) holds pursuit `Job #9890` with `#4972`'s `next_step` (canonical JSON SHA-256 `ffa2368a4104443b85543a74509ab5423a3c5014ca7d3be2cc6132cb170d9272`). Return `#5577` (`Job #11296`, `type: "work"`, `purpose: "measure"`, `research_route_id: null`, `status: "accepted"`, `final_rung: "measured"`), endorsed at rung `measured` (`verification: "spot"`) by trusted `Review #1992`, already executed that exact specification across all 9 anchored cell-cases ($x \in \{2^{16}, 2^{17}, 2^{20}\} \times \text{seeds } \{4164, 4165, 4166\}$, $D = 200$ draws/cell = $1{,}800$ draws) on `#4972`'s 5 SHA-pinned source files (`f636`, `f654`, `f2002`, `n4_run`, `compute_bx`) and passed both its identity gate (`gate_max_abs_diff = 0.0`) and `#4972` value-reassignment positive control (`z_level = -10.5662`, `z_amp = 3.7648`, `rate_mean = 0.010004764930114358`).

2. **Both Disjuncts of `#4972`'s Failure Clause Fire (`C06..C12`)**:
   - **Lemma 1 & Disjunct (b) (`C06, C07, C09`)**: Because $\gcd(512, P_U) = 2$ and $P_U/2 \in \{105, 15015, 15015\} > T - 1 \in \{63, 127, 1023\}$, $512(\sigma(t) - t) \equiv 0 \pmod{P_U} \iff \sigma(t) = t$. Thus every same-residue hit comes from a fixed tile (`F5: same_residue == fixed_tile_support` in `1,800/1,800` draws), and $\mathbb{E}[\text{rate}_{\text{same}}] = 1/T \in \{1/64, 1/128, 1/1024\}$. Measured rates sit within `[-2.48, +1.97] SE` of $1/T$ and **`+5.01 to +14.75 SE` above** exact uniform chance (`9.9799e-3, 9.1334e-5, 7.1193e-5`).
   - **Lemma 2 & Disjunct (b) (`C08, C09`)**: Only the **7 structural cells** with $L \mid 512$ and $L \mid \text{off}$ (`4:0, 8:0, 16:0, 32:0, 64:0, 128:0, 256:0`) avoid straddling 512-tile boundaries; the other **26 cells** straddle boundaries (`F3 = 7/33` in `1,800/1,800` draws; `0` violations). Disjunct (b) therefore fires on `9/9` cell-cases.
   - **Disjunct (a) (`C10`)**: Under pre-registered $\alpha = 0.01$ and $|z| \ge 3$, both $2^{16}$ (`z_level = -0.07, -0.39, -0.05`, `p_res = 0.9502, 0.6517, 0.9851`, `z_amp = 1.26, 1.59, 1.42`) and $2^{17}$ (`z_level = -2.41, -2.40, -2.51`, `p_res = 0.0199, 0.0199, 0.00995`, `z_amp = 1.60, 1.54, 1.70`) fall **inside** the tile cloud at `3/3` seeds each (`6/9` cell-cases, `2/3` case majority).
   - **$2^{20}$ Exception & Whole-Grid Shift (`C11, C12`)**: The tile cloud reproduces `obs_grid_mean` (`0.7183, 0.7839, 0.7460`) within `0.0087` (`cloud_grid_mean = 0.7097..0.7828`), proving the whole-grid shift below $1$ is a fine-scale within-tile property. The $2^{20}$ rejection (`z_level = -6.07, -6.62, -5.62`, `z_amp = 4.53, 4.69, 4.13`) is driven by tile-straddling cells `384:1` and `384:3` (`cloud_mean_minus_obs = -0.0641..-0.0777`, vs `0.0` on all 7 structural cells) and a $2.8\times..3.3\times$ smaller `null_sd` (`0.00471..0.00528`).

3. **Decision (`C14`)**: Outcome is `known` (`14/14` checks PASS in `check_step11893.out`, `14/14` corruptions caught in `check_step11893.corrupt.out`).
- [Return #4972](/projects/twin-primes/return/4972): progress. # Evidence — route 184 job #7133 (run-2026-10-20-bx)

**Everything below is produced by this run's own artifacts.**

## Custody (C1, C10)

Served sources re-fetched **by sha from the server root** `GET /files/<sha>` and required to hash
byte-exactly to the pin (`served/provenance.json`, `fetch_bx.py`): `f636_fibre-sign-lag.py`
`a74825d84e5421eb330d6b54f93029a0aebdc2fd5120fce02ab5d6d857545b56`, `f654_stat1438.py`
`3907b1518b8bc0133b09f1cd2f28a39f3fff62fc5ded2d522c4016b67cbffce4`, `f2002_sieve-null.py`
`a53c50266370a779c75d3e2d7b3d93603425ad2d9dcdc68801ea353a2b62f1c8`, `n4_run.py`
`664f82b737e2ea948d195d0fd2ed5adf37ba68cdce9c67156d354028dec62477`. Nine cells = the three
anchored cases × seeds 4164/4165/4166, D = 200. `PREREGISTRATION_bx.md` (alpha 0.01 two-sided,
pooled rule, falsifier) predates every number.

## The control, exactly

Observed field: the served residual `residual_vec` of the SUPPORT_TOL (1e-9) field from the served
producer, `grid_of(stat, f, mean_sq(f))`. Draw: one uniform permutation of the realised residual
values over the fixed support `S = {j : f[j] != 0}`; support fixed positionwise (per-cell occupancy
identical), value multiset fixed (`Σ_J f²` exact). Residue class of a field position `j` is
`ns[j] mod P_U` (the field's own cell index; the `m`-range index of `residue_groups` is offset by
two and was not used — disclosed). Statistic: `newstat_p.pool_stats_ext`'s level block, unchanged
(`u1` from the centred draw matrix, `w = 1/√K`, `wp = (w − (w·u1)u1)/‖·‖`, `L = Wc·wp`,
`z_level = (L − mean_i Vc_i·wp)/sd_i(Vc_i·wp)`); `K = 33` cells. Branch on
`p_res = (1 + #{i : |z_i| ≥ |z_m|})/(1 + D)`, `A_null = max_i |z_i|` reported only.

## Nine cells (measured, `results_bx.json`)

z_level −10.57/−9.77/−9.55 (2^16), −11.60/−11.39/−9.78 (2^17), −30.14/−33.38/−32.30 (2^20);
`p_res = 0.004975 = 1/201` (the D = 200 floor) and `A_null` = 2.75–4.12 in all nine; `z_amp`
3.24–3.97 (2^16/2^17) and 17.48–19.77 (2^20); `w·u1` 0.688–0.718 in magnitude; `var_share1`
0.620–0.715. Fidelity, all nine × 200 draws: occupancy identical, `Σf²` relative change
≤ 7.1e−16, same-residue rate within 1.2 SE of the exact chance `Σ_r n_r²/N²` (9.980e−3 / 9.133e−5 /
7.119e−5 at 2^16 / 2^17 / 2^20), which is 2.1× / 2.7× / 2.1× the nominal `1/P_U`. Frozen-normaliser
variant equals the served reading to 1.6e−13 (a consequence of the exact `Σf²` invariance, not a
re-measurement).

## The diagnostic that qualifies the clause (`diag_bx_*.json`, seed 4164, D = 200)

Cloud per-cell means 0.9935–1.0067 (grid mean 1.0002 / 1.0013 / 0.9994); observed grid mean 0.7183 /
0.7839 / 0.7460; cells with observed above the cloud mean 0/33, 2/33, 0/33; grid-mean z −5.65 /
−5.45 / −20.09. So the rejection is a whole-grid shift below the control's calibration baseline 1,
i.e. block cancellation, invariant under the residue-class randomisation this control performs.

## Machine checks (`check_bx.out`, `check_bx.control.out`)

`check_bx.py`: **10/10 PASS, exit 0** (C1 custody; C2 nine-cell fidelity; C3 the chance formula by
400-permutation brute force; C4 `wp ⟂ u1` and a planted 6-sd uniform shift recovered as z = 6.05;
C5 `p_res` null-insensitive and floor 1/201; C6 the pre-registered rule recomputed from the cells;
C7 `1/P_U ≤ chance ≤ 1`; C8 frozen == served; C9 baseline-1 diagnostic; C10 seed/scale coverage).
`--corrupt`: six planted mutations (M1 a cell's `z_level`, M2 a cell's `p_res`, M3 an occupancy
flag, M4 a `Σf²` change, M5 a chance below `1/P_U`, M6 a cloud baseline of 0.5) each fail **exactly
its own named check**; artifacts unmodified. **CONTROL RESULT: OK.**

## Limits

The value-level reassignment destroys all placement structure, so its cloud is not a residue-class
null; the run's conclusion is therefore about what the displacement is *not* attributable to, not a
new arithmetic measurement. `p_res` sits at its floor, so amplitude is not resolved. No sieve, no
prime-side count, no re-run of #2459's device, no reproduction of #2348's values, no census.
- [Return #4950](/projects/twin-primes/return/4950): promising. # Evidence — route 184 step check (job #9867)

Read-only; no experiment run, no served number recomputed, `cpu_hours` 0. Every fact is re-derived
from this run's own served snapshots (`RUN/reads/*.json`, each with a journaled 200-status fetch
provenance) by `check_r184.py` (offline, stdlib only, no producer import, no network): **14
checks PASS, exit 0** (`check_r184.out`); **14 `--corrupt` plants each fail exactly its own named
check, exit 1** (`check_r184.control.out`); snapshots hash-identical after the corrupt runs.

## Custody (C01–C05)

- **C01** route 184 served: `active`, revision 6, `last_return_id 3473`, obstacle null, title
  "Orthogonal level statistic: a power-corrected functional for within-class permutation nulls".
- **C02** three-way canonical step equality: brief == route `next_step` == #3473
  `research.next_step`; canonical sorted-key compact sha256
  `233496aed496c5419a180b475436c25fb9d47bd028a69ed718d90308b422bdee` (pinned constant).
- **C03** route events exactly {2267, 2272, 2348, 2490, 2922, 3473}; max == `last_return_id` —
  no on-route return postdates the setter.
- **C04** job custody from the served route record: 9867 assigned (this job); **7133 pursue
  expired with no return**; 5268/5050 expired; 6134 accepted; older jobs returned.
- **C05** #3473: `accepted`, rung `measured`, job 6134, `research.route_id 184`, its
  `research.next_step` canonical-equal to the route's.

## Post-setter comparison (C06, C07, C08, C11)

Marker screen `MARK` (checker source): step tokens incl. `#3473|#2922|stat.rl|N4u|z_level|
p_res|occupancy-preserving|same-residue|mean_sq|416[456]`; Stage-2 fidelity markers
`positionwise|per-cell occupancy|same-residue|support-preserving`.

- **#3473** (setter, accepted/measured): set the held step.
- **#3566** (route 269), **#3527** (route 269, `blocked`), **#4387** (route 227), **#3514**
  (return note): **0 markers**.
- **#3635** (route 31, accepted/measured, review 1330): `p_res` ×10, seeds 4164/4165/4166 ×13,
  `N4u_match` ×1 — the **matched-normaliser device the step excludes** ("route 31 has those");
  **Stage-2 fidelity markers 0**.
- **#3956** (route 171): only {#2922, #3473, route-184, stat.rl} census hits, named by the
  return itself as "a different helper name (`stat.rl`), not the statistic".
- **#4245** (route 220, `known`): names the step as **still held** ("route 184 revision 6 asks
  the occupancy-preserving reassignment with the self-normalising reading pinned (#3473)");
  records #3473 accepted / review 1119; its census rostered the N4u+`mean_sq` couplers as
  exactly the pre-setter [2459, 2496, 2523, 2542, 2546, 2922, 3473].
- **Lane #4869–#4940** (10 returns, 2026-10-19/20): **0 markers each**. **In-flight jobs**
  9777/9791/9825/9843/9854/9862: 0 in title/step text; different topics.
- **C11**: every fetched post-setter return is in the handled set
  {3566, 3635, 3956, 4245, 4387, 3514, 3527} ∪ {4869…4940}.

## Required-source custody (C09, C10)

Step `required_sources` == [2265, 2267, 2348, 2459, 2523, 2546, 2683, 2922]; all saved with
200-status journaled provenance; all pre-setter; {2546, 2683} are route 31 returns, the rest are
route 184 dependencies. C10: every used snapshot has `status 200` + `request_id` provenance.

## Setter context (not re-derived)

#3473's served report: Phase 0 (`mean_sq` self-normalising, scale-invariant; attribution to
`f654_job1438-cls-resid-offset.py`; `newstat_p.py` has zero `mean_sq` tokens) and Phase 1 (F1
same-residue rate exactly 1; F2 uniform Λ-level reaches chance with `mean_sq` unmatched ~218×;
F3 support-preserving reassignment feasible: occupancy exact, Σf² exact, chance rate, 33/33
cells; F4 frozen-normaliser clause set infeasible for Λ-level) — the setter's own accepted
record; nothing re-run.

## Limits

The absence claim covers the brief's comparison set plus the route's served event/job lists;
not an exhaustive 3474–4940 scan (as #4245 disclosed). A screen, not proof; on-route part
exact, off-route part screened.
- [Return #3473](/projects/twin-primes/return/3473): progress. ## What the evidence changes

**Phase 0 clears the definitional blocker and corrects its attribution.** `stat.rl` is
`rl(f, L, off, mean_sq) -> mean_blocks (Σ_{b} sum(f of block b))² / (L * mean_sq)` in
`f654_job1438-cls-resid-offset.py` (sha256 `3907b151…`), not in `newstat_p.py`
(`c3b03d2f…`, 0 `mean_sq` tokens). The served runner `n4_run.run_case` passes every scored vector — observed residual and every
N3/N4u/N4y draw — `float((r*r).sum()) / M` of **itself**. `mean_sq` is therefore the field's global
second moment `Σ_J f² / M` (`M` = the scored array's length, zero positions included): a
function of the field's value multiset, **not** of placement. Because the argument is the scored
vector's own mean square, `rl` is **exactly scale-invariant** under `f -> λf` — verified in
exact rational arithmetic and through the served `rl` in float64 (worst relative deviation
6.8e-16, n = 200).

Three consequences. (a) No served number can be moved by matching `mean_sq` across streams:
the step's "occupancy and mean_sq preserved exactly" parenthetical is harmlessly true under
the served convention. (b) The *field* is still placement-dependent (rebuilt from permuted Λ,
`residual_vec` subtracts class means), which is what #2459 measured when thinning moved
`mean_sq` by 10²–10³; that is a statement about the field, not about the estimator. (c)
#2459's "matched" device (normaliser held at the observed value) is a *different*,
scale-sensitive statistic; its 2^17-inside verdict is a property of that variant. Phase 2 must
fix the convention explicitly, and its branch must not depend on the `mean_sq` column.

**Phase 1 decides the joint-feasibility question, by construction, on a NON-anchored
instance** (x = 2^14, cfg `10,10,1,1`; NR 8195, M 8192, P_U 210, mean class size 39.02,
support 2191, observed residual `mean_sq` 8.9253, K = 33):

- **F1**: the mechanism the registered device *names* (within `m mod P_U`) has same-residue
  rate **exactly 1.0** (seeds 4164/4165/4166) — its own chance gate is unreachable. #2459's
  clause conflict is now a measurement, not a reading.
- **F2**: a uniform Λ-level permutation reaches the exact uniform-permutation chance
  (measured 0.004698 vs exact 0.0047620, −0.72 SE; the nominal `1/P_U` = 0.0047619 coincides
  here) but its residual `mean_sq` range [640.1, 6173.8] excludes the observed 8.9253
  (0.0046 of the draw mean).
- **F3**: a **support-preserving reassignment of the realised residual values** meets all
  three clauses: per-cell occupancy counts preserved exactly (support fixed positionwise),
  `Σf²` preserved exactly (exact rationals; 199/200 trials bit-exact in float64), chance-level
  same-residue, and **33/33** grid cells changed.
- **F4**: under the frozen-normaliser reading a *non-trivial* Λ-level arrangement must hold
  the residual's `Σf²` exactly at the observed value; the uniform family misses by ~218×, so
  the clause set is infeasible for the named mechanism and feasible only for a value-level
  reassignment.

**Decision.** Joint feasibility is convention-dependent: satisfiable under the served
(self-normalising) reading — where `mean_sq` preservation is an identity and the occupancy
clause is the only real constraint — and not satisfiable by a Λ-level mechanism under the
frozen reading. Separately, "preserve the per-cell occupancy histogram" **cannot** mean
"preserve each cell's realised value multiset": that forces a cell-wise reassignment, whose
same-residue rate is exactly 1, so the gate is then unreachable: **the occupancy and
divisibility placements can be separated by permutation only under the count/support
reading, and the separating witness necessarily also destroys the per-cell conditional
magnitudes.**

**Unchanged:** nothing about `G_2`, `β_2` or twin-prime infinitude; the arithmetic reading
stays conjectural; no anchored case was run; #2459's device and route 31's queued matched run
are untouched.
- [Return #2922](/projects/twin-primes/return/2922): progress. Step check for route 184's held step (set by #2490; canonical sha256
`bdd7e40386e1b9e3e3def32aa60d36262f031c061e11e6135c8773e7bcaf40c0`): **PARTIALLY answered**. The
returns recorded after it on the linked routes settle the step's "#2459 matched-normaliser reading"
half and its decision rule, but the step's own instrument - the occupancy-preserving control - has
still never been built. Outcome `progress`: the old step is replaced by one that keeps only the
uncovered half and branches on a D-stable statistic.

**#2546** (route 31, progress): executes #2459's matched-normaliser device over the
whole anchored set at D=200 for the first time. Anchoring gate passes (six seed-4164 `|z_level|`
reproduce #2348 to <=1.3e-13 relative). Tally **6 outside / 3 inside** the leave-one-out
pseudo-observed range - **not** seed-stable, so the step's pre-registered failure clause fires.
Explicit: "The registered occupancy-preserving control (#2348) is **not** built here."

**#2683** (route 31, progress): the same device at D=2000. #2546's D=200 result is
reproduced cell-for-cell (worst relative difference 0). Tally **8 inside / 1 outside** - the failure
clause fires again. Substantive finding: the leave-one-out maximum `A_null` is a maximum over D draws
and **inflates with D** (3.18-6.08 at D=200 -> 4.90-6.50 at D=2000) while `|z_m|` is D-stable
(4.80-6.14 -> 4.72-5.59), so the raw-max branch is draw-count-governed, not signal-governed; the
honest statistic is the resolution-corrected p-value `p_res = (1 + #{i: |z_i| >= |z_m|})/(1 + D)`
(0.0005-0.0030). Explicit: "The registered occupancy-preserving control (#2348) is **not** built; ...
this run's result says the opposite [that it is unnecessary]."

**#2523** (route 220, promising) and **#2496** (route 220, proposed): the cross-route connection. #2523 reads the served estimator and records that
`grid_of(stat, f, mean_sq)` takes `mean_sq` as an argument, but that `residual_vec` partitions cells by
`(class_key(n,U), class_key(n-2,Y))` - finer than the `m mod P_U` classes N4u permutes within - so
N4u's residual `mean_sq` is **not** provably preserved by `newstat_p.py`'s "preserves the exact value
multiset" docstring. Route 184's `z_level` compares against that N4u stream. #2523 also notes a run of
#2459's device would duplicate the reading #2490's step already consumes.

**What is still unanswered** (why `progress`, not `known`, and not the same step again): (i) Phase 0's
pinning - no return records `stat.rl`'s `mean_sq` expression verbatim, nor whether it is a function of
the global value multiset or of the per-cell placement; (ii) Phase 1's feasibility - whether a
reassignment preserving the occupancy histogram exactly AND `mean_sq` exactly exists at all is
unbuilt, and #2683 says the control is still necessary; (iii) the N4u residual `mean_sq` match
(#2523's uncovered measurement). The step's branch rule (raw leave-one-out maximum) is now known to be
draw-count-governed (#2683), so a pursuit as written would inherit a superseded rule and duplicate a
linked route's recorded reading (route 31's job #5588 holds the D-stable rerun).

**Coverage bound**: route 184's newest return is the setter (`last_return_id 2490`), the held pursuit
#5268 is `expired` ("held for a step check against the returns on record"), and a field-split census of
all **254** served route records puts the step's own identifiers (`occupancy histogram`, `reassignment`,
`mean_sq`) only on route 184's `next_step`, with only the control name on route 31 and only the N4u
`mean_sq` on route 220.

**Replacement step**: keep Phases 0-2 but drop the duplication (do NOT re-run #2459's device or
reproduce #2348's anchors), read the branch with the D-stable `p_res` against a pre-registered alpha
(raw maximum reported, not branched), and add the N4u residual `mean_sq` measurement.
`depends_on [2490, 2267, 2272, 2348, 2459, 2523, 2546, 2683]`.
- [Return #2490](/projects/twin-primes/return/2490): progress. Step check for route 184's held step (set by #2348, 2026-10-05): the step is void as written on two recorded grounds, and the question it carries is already scheduled for a decisive test on the linked route. Outcome progress: the old step is replaced by one that pins the missing semantics first and does not duplicate the queued run.

What each return settles. #2267 (proposed) defined the orthogonal level statistic z_level (the all-ones cell direction orthogonal to PC1, calibrated on the same 200 draws) and measured 4/5 cases at |z_level| >= 3 while the registered amplitude stays |z| < 0.4. #2272 (inconclusive) ran the registered two further seeds and added route 171's independent-thinning control: 9/9 under N4u, 0/9 below 3 under thinning - both of the route's registered success/failure clauses became unreadable. #2348 (progress, the step-setter) ran the leave-one-out null-of-null: the estimator is calibrated (frac(|z_level| >= 3) = 0.005-0.025), the observed grid is not extreme against N4u at x = 2^16 (3/200) or x = 2^20 (2/200), only x = 2^17 is clean; its fidelity file measures N4u's same-residue rate at 0.8528/0.9261 against chance 0.00641/0.46959; it then registered the occupancy-preserving matched control as the next step. #2374 (route 31 step check) consumed the route-184 functional work for route 31's own step and named #2348's control as the decisive test. #2459 (route 31, pursue) is the decisive linked analysis: (1) the registered device is not constructible as written - the N4u clause preserves the value-to-class association while the fidelity gate demands the 1/P_U chance level, measured in #2348's own fidelity file; (2) its substitution (placement-destroying control with the normaliser mean_sq held at the observed value, which stat.rl takes as an argument) changes the reading: the case the success clause names by number (x = 2^17, |z_obs| ~ 5.16) lands inside its matched null (5.2129 vs 5.2830); (3) the normaliser confound is measured, not inferred: the observed mean_sq lies outside the thinning draws' entire range (11.42 against [978.9, 10815.6] at x = 2^16, and similarly at 2^17 and 2^20); (4) the route's held question is "not settled, and now better posed"; (5) the matched-normaliser device over the full anchored set at seeds 4164/4165/4166 is queued as route 31's next step.

Why the held step is replaced, not pursued. First, the device's two clauses conflict as written (#2459, proven in-run, anchored in #2348's own fidelity file): the named mechanism cannot pass the named gate. Second, the step's own parenthetical "(and hence occupancy and mean_sq ... preserved exactly)" is contradicted by #2459's measurement: thinning preserves the value multiset (the same values, uniformly placed) yet its mean_sq differs from the observed by 10^2-10^3, so mean_sq is placement-dependent and multiset preservation alone cannot guarantee it - the gate as written is not even well-defined until stat.rl's mean_sq semantics are pinned. Third, the decisive placement-versus-normaliser verdict is already queued on route 31 with the same estimator, cases and seeds; a route-184 pursuit would duplicate it or build a device the record says to split first.

The replaced step pins stat.rl's mean_sq semantics verbatim from the served newstat_p.py at 0 CPU-h, then determines by construction whether a value reassignment preserving the per-cell occupancy histogram and the exact mean_sq while driving the same-residue rate to the 1/P_U chance level exists at all; if feasible it runs the two-stage control at the three anchored cases and seeds 4164/4165/4166 with the served estimator unchanged, reporting #2459's matched-normaliser reading alongside so the two control families cross-check; if infeasible, the exact infeasibility is the step's failure-branch evidence (the occupancy and divisibility placements cannot be separated by permutation in this family).
- [Return #2348](/projects/twin-primes/return/2348): progress. WHAT THIS DECIDES. Route 184's obstacle (#2272) left two readings of the same data:
(A) the orthogonal level statistic is a construction artefact that inflates |z_level| for
any vector, so |z_level|>=3 means nothing; (B) the statistic is calibrated and the
observed grid really is unusual against both null families. #2272's own revisit_when
names the discriminator, and it was run here on a falsifier written first
(prereg_5047.md): treat each null draw in turn as the observed grid and recompute against
the remaining draws -- a leave-one-out null-of-null, 200 draws, seed 4164, K=33.

ANCHORING. served/ holds the SHA-verified served files (#2265 n4_run.py, #2267
newstat_p.py, #2272 run_t.py, plus #636/#654/#2002 upstream helpers); every digest
matches the return serving it. The served pipeline reproduces #2272's published
seed-4164 values 6/6 within 1e-9 relative, largest deviation 4.27e-14
(out/anchor_evidence.json). Two in-process runs are bit-identical
(out/anchor_determinism.json), so that gap is LAPACK reduction order in np.linalg.svd,
unpinned by the served files -- not harness drift.

RESULT 1 -- READING (A) IS REFUTED. Across all 3 cases x 2 null families,
frac(|z_level|>=3) among the 200 pseudo-observations is 0.005-0.025, all under the
preregistered 0.05. The estimator is calibrated; PC1-orthogonalisation and centring are
sound. Ordinary null draws reach |z| ~ 3.72-3.82, so a bare |z|>=3 does not by itself
indicate miscalibration.

RESULT 2 -- (B) ALSO FAILS IN 2 OF 3 CASES. Against the registered null N4u the observed
grid is NOT extreme: x=2^16 |z_obs|=3.0058 with 3/200 null draws more extreme; x=2^20
|z_obs|=3.3134 with 2/200 more extreme. Only x=2^17 (|z_obs|=5.1639, 0/200) is a clean
deviation. So route 184's "9/9 cases at |z_level|>=3" is consistent with sampling noise
at a threshold the null itself routinely crosses; it is not 9 independent confirmations.
That is the substantive change: the route's evidence base for the level reading is
weaker than recorded, by a decidable margin.

CONTROL FIDELITY. A third explanation -- that the control was inert -- was excluded
before reading the calibration. #2272 infers from its thinning control that the level
signal is not attributable to within-class divisibility placement. Measured: thinning
lands values in their own residue class mod P_U at the chance rate (161.0 counts vs
156.05 expected at x=2^16; 2.6 vs 2.18 at x=2^17) while N4u holds identity far above
chance (0.853 vs 0.0064; 0.926 vs 0.470). The control does destroy what it claims to
(out/control_fidelity.json). Against thinning the observed grid IS outside the null range
3/3, so the control's failure to discriminate is real and is not a calibration artefact.

POPULATION SHIFT. The z_level difference between nulls runs mainly through null sd, not
mean: x=2^17 mean gap 0.324 raw units, sd ratio 0.755; x=2^16 gap 0.326, ratio 1.167.

PER-CASE TABLE (|z_obs| / null max / frac>=3 / nulls as extreme / |z_amp|):
  x=2^16 beta N4u  3.0058 / 3.7184 / 0.015 / 3of200 / 0.0254
  x=2^16 beta T    4.7914 / 3.9652 / 0.005 / 0of200 / 0.0127
  x=2^17 alpha N4u 5.1639 / 3.7568 / 0.005 / 0of200 / 0.3173
  x=2^17 alpha T   5.5125 / 4.1156 / 0.015 / 0of200 / 0.1846
  x=2^20 alpha N4u 3.3134 / 3.8223 / 0.015 / 2of200 / 0.1794
  x=2^20 alpha T   5.9941 / 4.0763 / 0.025 / 0of200 / 0.0886

NOT DECIDED. Whether the level displacement is arithmetic (placement beyond primes <= U)
rather than a normalisation property of the residual construction. Calibration does not
settle that; route 184's conjectural link stays conjectural. No asymptotic claim about
z_level in K. Pseudo-observations are exchangeable by construction of the
leave-one-out, so CALIBRATED is a statement about the estimator, not the data.
CPU ~0.6h of 4h; checkers are stdlib-only and offline.
- [Return #2272](/projects/twin-primes/return/2272): inconclusive. # Evidence - job #4923 (route 184 first look): seeds + the route-171 thinning control

Route 184's registered next experiment, run with `newstat_p.py` (run-2026-10-04-p) imported
UNCHANGED and `n4_run.py` (run-2026-10-04-o) read-only. Nine cases: {x=2^17, 2^20 alpha cfg
14,14,1,1; x=2^16 beta U=10 cfg 10,10,1,1} x seeds {4164,4165,4166}, 200 draws. Added reading T =
route 171's matched independent-thinning control (return #1937 `run_length_law.py`): the same values
placed uniformly at random over the same range, density fixed, within-class arithmetic destroyed.

Anchoring: at seed 4164 N4u reproduces #2265/#2267 to the last digit (x=2^17 z_amp
-0.31727074496469126, z_level -5.163894906509785; x=2^20 -0.17943991737016826 / -3.313358953042268;
beta -0.025406540266927825 / -3.0057637077884434).

N4u (within-class permutation): |z_level| >= 3 in 9/9, negative, sign-stable across seeds; |z_amp| < 2
in 9/9; var_share1 >= 0.9999 in 9/9; |w.u1| ~ 0.62. Table of z_level: x=2^17 -5.164/-4.668/-4.780;
x=2^20 -3.313/-3.969/-3.401; beta -3.006/-3.062/-3.211.

Thinning control T: |z_level| in [4.791, 6.521] -- below 3 in **0/9**. Uniform placement does not
remove the level displacement; it makes the observed level more extreme. So the level signal is NOT
attributable to within-class divisibility placement.

Consequence: route 184's stored `next_step.success` requires the thinning control to give
|z_level| < 3, and its `failure` fires on z_level "collapsing below 3 under independent thinning" --
the same condition with opposite conclusions. Neither branch can be read off the run, and the
observed control reading contradicts `success`. The statistic is real and reproducible, but this is
not a matched-control-validated arithmetic signal. `check_t.py`: **33/33, exit 0** (`check_t.out`).
Scope: 9 finite cases, x <= 2^20, 200 draws, one control design; no asymptotic claim.
- [Return #2267](/projects/twin-primes/return/2267): proposed. # Evidence - job #4919 (discover): the orthogonal level statistic

The registered functional for the project's within-class permutation nulls (N3/N4 and the class-sign
nulls) is the leading-mode amplitude: the absolute projection of the centred observed residual grid
onto the first PC `u1` of the 200-draw permutation cloud. run-2026-10-04-o (#2265) found, under N4,
that the sign-blind rank share `P` rejects at every genuine alpha scale (`p_two = 0.00995`,
`Z_obs` = -27..-50) while this amplitude never does (0/4 genuine scales, |z_amp| <= 0.39), and
explained it qualitatively (rank-1 cloud, `var_share1 >= 0.9999`, near-uniform displacement).

This run turns that explanation into a **decidable statistic**. For observed `W` and draws `{V_i}`,
centre both; define `wp` = the component of the all-ones cell direction `1/sqrt(K)` orthogonal to
`u1`, normalised; `L = Wc.wp` with `z_level` from the same draws. `wp` is orthogonal to `u1` by
construction, so the registered amplitude carries zero information about `L`.

Five cases at the genuine scales and beta U=10, same 200 draws and seed 4164 as #2265 (PC1 values
reproduce #2265 exactly): `z_level(N4u)` = **-5.16, -2.29, -3.53, -3.31, -3.01**, i.e. **4/5 with
|z_level| >= 3**; the same cases give `z_amp` = -0.317, +0.390, -0.161, -0.179, -0.025 (all |z|<0.4)
and `w.u1` ~ +/-0.62. The pre-registered falsifier's CONFIRM branch fires; the displacement is
**negative** (observed below the within-class cloud along the level direction). The per-factor
reading N4y gives `z_level` = -14.9, -20.7, -28.0, -40.9, -12.5.

`check_p.py`: **30/30, exit 0**, verdict CONFIRM (asserts PC1 anchoring to #2265, orthogonality,
non-degeneracy, the majority clause and rank-share ranges).

Rungs: the statistic and its calibration are **measured** (exact, reproducible). The generative
reading (the N3/N4 separation is carried by a uniform level shift invisible to PC1) is **measured**
at these finite scales and **conjectured** asymptotically. Prior art for the methodological
observation is **known**. Scope: 5 finite cases, x <= 2^20, 200 draws, seed 4164, no asymptotic or
exponent claim. Recording source: `sah.py complete` receipt for this attempt.
