# Job #1629 — PRE-REGISTRATION (written BEFORE any computation on the retained census) - attempt `87df287927e37cbff72ae5bb5ce0a0ad`, session `35883a232f8b9bc51ea0fbca`, department `dept_c326cb5ae203e5d0d94f8db1`, general mode, job **#1629**, explore/discovery, lane `formalize`, NO ROUTE. Written 2026-09-17 ~09:17Z, before fetching or reading `research/centered-discrepancy-measurement.json`. - Object of study: return #165 / job #? census note `research/centered-discrepancy-measurement.md` (verdict `MEASURED to x=2^38`) and the OPEN sufficient consumer **(16)** `D_y(x) >= -(4/25) x + o(x)` of `research/moving-cutoff-parity.md`. ## 1. The gap being addressed (quoted, not paraphrased) Census note §3 item 3: "The real D_y's excess over the control is therefore the slow convergence of a proved classical asymptotic, and **the parity object's own fluctuation, which is what a census was meant to look at, is not isolated by this raw comparison**." Census note §3 item 7: "**no census of D_y at reachable x can separate the parity object from the classical convergence**, so a larger run has no decision attached and should not be made." Census note §3 item 3 also records, from the *same retained columns*, the exact identity D_y/x = (S/x - C2) - (T1/x - C2) - P/x - E_pp/x - E_even/x, in which every term but `-(T1/x - C2)` is below 2e-4 in magnitude for j >= 30. ## 2. The new statistic (definitions fixed now) Using only columns the census already retains (`D_y`, `T1`, `C2`, `S`, `P`, `E_pp`, `E_even`, and the matched random-sign control's `D_y`/rms per j), define at every retained dyadic `x = 2^j`: c_D(j) := D_y(2^j) / 2^j c_T1(j) := T1(2^j) / 2^j - C2 (the classical companion's finite-size error) R(j) := c_D(j) + c_T1(j) (companion-removed, "detrended" parity residual) rho(j) := |R(j)| / ctrl_rms(j) (R in units of the census's own matched control) lam := -slope of log|R(j)| against log(2^j), fitted by least squares over j >= 30 lam_ctrl := the same slope for the control's own D_y (census reports 0.544 over j >= 26) lam_raw := 0.863 over j >= 26 (census's reported slope of log|D_y|, which is what the raw statistic sees) ## 3. Decision this statistic informs Whether the census's own raw "excess" reading (|D_y|/ctrl_rms = 6.014, 12.849, 8.252, 7.703, 7.374, 8.580, 21.816 at j = 32..38, which "read alone would look like arithmetic structure") is **entirely** the classical companion's convergence, or whether a signed parity-object fluctuation survives removal of the companion. In the second case the census's §3 item 7 / §4 "bounded negative about the method of measurement" is refuted at the retained scales, (16) becomes numerically addressable there, and the retained census already contains a usable handle. ## 4. Hypothesis under test (taken from the census, not from me) H0: the companion-removed residual is noise-sized, i.e. `rho(j) = O(1)` for every retained j >= 30, and `lam` is within 0.15 of `lam_ctrl = 0.544`. Under H0 the census's diagnosis is confirmed and the raw |D_y|/ctrl_rms ratio is retired as a measure of arithmetic structure. ## 5. PRE-REGISTERED FALSIFIER (F*) F*: if for some retained j >= 30, `rho(j) > 3`, **and** the fitted `lam > 0.544 + 0.15`, then H0 is **falsified**: the companion removal is insufficient, the retained excess is not (only) classical, and this job must return an **audit** of `centered-discrepancy-measurement.md` §3 item 3/7 rather than a design. Any other outcome (including `rho > 3` at an isolated j with `lam` inside the band, or the residual falling below the control) **confirms H0** and the return is the design below. Secondary, algebraic, pre-registered: F**: `|R(j) - (c_D(j) + c_T1(j))| < 1e-12` for every row, and the recomputed identity residual `|c_D - ((S-C2)/x - c_T1 - P/x - E_pp/x - E_even/x)| < 1e-9` — this is a transcription check on the retained JSON, not a new claim. ## 6. Matched control The census's own **four seeded random-sign draws** (`mu(n) -> +-1` on the same support, `mu(e)` kept) supply the matched null: the same functional, the same support, the same sieve. The design below adds an **independent-thinning** control (keep each slot independently with the density the census's own count implies) as the confirmatory run's null, and a **shift-4** negative control (already in the census) to show the statistic is not shift-blind. ## 7. Scale at which the effect would be visible Companion removal leaves a residual of size `ctrl_rms` under H0. At j = 38 the census's control rms is about `1e-4 x`, and the largest competing term of the identity is `1.84e-4 x` at j = 30. So a non-classical signed fluctuation is visible from **j >= 30** onwards at any size above `3 * ctrl_rms / x ~ 3e-4`, i.e. about 1/500 of the `4/25 = 0.16` threshold of (16). Below j = 30 the identity's neglected terms are not small enough for the residual to be interpretable. ## 8. Cost if the retained columns suffice; cost if they do not Retained-column evaluation: **0 CPU-hours**, seconds, stdlib Python, no new arithmetic. If the retained JSON lacks the `T1` or control columns, the missing part is exactly one column re-derivation on the census's own fixture (j <= 38), which the census priced at 8223.5 s wall on 8 workers (~18 CPU-h) and which **exceeds** this assignment's 4 CPU-hour budget — in that case this return is the design plus the explicit cost, and no run is made.