# Pre-registration, job #385: split T1's finite-x error into a computable truncation term and a BV residual Written before any computation (2026-09-12, session of @Benjaminsen, claude-fable-5-1). Builds on `research/moving-cutoff-parity.md` (4)-(8), (12); `research/centered-discrepancy-measurement.md` §3 items 3, 4, 7 and the embedded tables of `research/centered-discrepancy-measurement.js` (MAIN TABLE, IDENTITY (12) PIECES, RANDOM-SIGN CONTROL, j = 16..38); return #34 (census13.c) and its review (return #34's review, job #132); `research/history/reviews-0907/10-independent-review-0908.md` §5. ## The gap in the record The census note says (item 3) that at every reachable x, D_y/x = -(T1/x - C2) within 2e-4 for j >= 30, that T1's error "is comparable to x/log^2 x with oscillating sign", and (item 7) that "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". Nobody has asked which part of T1 - C2 x is deterministic. T1 = sum_{n in J} Lambda(n-2) mu^2(n) A_y(n) with A_y(n) = -sum_{d | n, d <= y} mu(d) log d truncates the Mobius sum at d <= y = ceil(x^(12/25)). Its Bombieri-Vinogradov main term is a finite Mobius partial sum, not C2 x: T1_main(x, y) := (x/2) * sum_{d <= y, d odd sqfree} sum_{b odd sqfree, [d, b^2] <= x} (-mu(d) mu(b) log d) / phi([d, b^2]) (sharp; the exact model expectation) T1_inf(x, y) := (x/2) * A2 * sum_{d <= y, d odd sqfree} (-mu(d) g(d) log d) / phi(d), g(p) = (p-1)^2/(p(p-1)-1) (the note's (8) form, b unbounded) with A2 = prod_{p > 2}(1 - 1/(p(p-1))) = 2 * Artin = 0.7479116272384044 and A2 * H_g(0) = 2 C2 (checked per prime in the review of return #34). Both are computable without any prime above y; at j = 38, y = 309,591. delta_trunc(x, y) := T1_main(x, y) - C2 x (deterministic) B(x, y) := T1 - T1_main(x, y) (the mu(d) log d weighted sum of the BV discrepancies of Lambda(n-2) at moduli [d, b^2]) D~_y := D_y + delta_trunc (the census's D_y with the deterministic part of T1's error removed) By the exact identity (12) pieces, D~_y = (S - C2 x) - B - (K + 2 C2 M) + 2 C2 M - E_pp - E_even, where every term except B and S - C2 x is below 1e-4 x for j >= 30 in the embedded table. So D~_y is the arithmetic residual: the twin-count deviation minus the weighted BV error at level y. ## The statistic and the decision it informs Statistic: the pair (delta_trunc/x, B/x) at x = 2^j, j = 16..38, from the embedded T1/x - C2 column (six decimals, enough at the 1e-4 level) and the computed T1_main; and D~_y/x against the embedded random-sign control sd. Decision: whether the retained census already separates the parity object's fluctuation from the classical convergence once a computable term is subtracted, contradicting item 7's "not measurable this way", and whether the residual B is of independent-thinning size. If delta_trunc carries the bulk, the record's reading "T1's error is of order x/log^2 x" should become "T1's error is the truncation of the Mobius partial sum at u = y, computable to the digit, plus a residual of size ~ 1e-4 x", and the census note's item 7 is wrong in scope: the existing columns do isolate the residual. ## Falsifiers, fixed now - F1 (sign, j = 30..38, nine rows): sign(delta_trunc) = sign(T1/x - C2) in at least 7 rows. Fewer refutes the truncation reading. - F2 (bulk, j = 30..38): RMS(B/x) <= 0.5 * RMS(T1/x - C2). Otherwise the deterministic term does not carry the bulk and the reading is refuted. - F3 (residual size, j = 16..29 where the control is exact): |B| / sd_thin <= 3 at no fewer than 11 of the 14 rows, where sd_thin is the exact standard deviation of B under the independent-thinning control below. More rows outside 3 sd means B carries structure beyond thinning (that is a finding, not a failure of the design; it is recorded either way). - F4 (D~_y against the census's own random-sign control, j = 30..38): |D~_y/x| / (ctrl sd) has median at most 3. Otherwise D~_y still sits far above the random-sign scale and the corrected object has structure or a further deterministic term. Sharp version is primary. The infinite-b version T1_inf is reported alongside; if the two differ by more than 2e-4 x at j >= 30 the difference (the b-tail of the model) is itself reported, since it sits at the level of the residual. ## Controls 1. Independent thinning (Cramer): for odd n in J, w(n) = log n * Bernoulli(2 / log n); T1_thin = sum w(n) mu^2(n) A_y(n); E[T1_thin] = 2 sum_{n odd sqfree in J} A_y(n) exactly; Var = sum_{n odd sqfree} 2 log n (1 - 2/log n) A_y(n)^2 exactly. Four seeded draws plus the exact sd, j = 16..29 (full sieve to 2^29, as census13.c). For j >= 30 the sd is extrapolated from a fit over j = 20..29 and marked heuristic. 2. The census's random-sign control (mu(n) -> +-1 on the same support), sd column of the embedded RANDOM-SIGN CONTROL block, all j to 38. 3. Identity control: at j = 16..29 my own T1 must equal census13.out's T1/x - C2 column (return #34) to printed precision. ## Scale where the effect would be visible At j = 38, |T1/x - C2| = 1.47e-3 while the thinning sd of B/x is expected near 1e-4 (sqrt(x) log^{3/2} y / x), so if delta_trunc carries the bulk it is visible at every j >= 30 in the retained data; no new run is needed. If F1 or F2 fail, the residual is not deterministic and a run at j = 39..40 would not change that (the ratio moves by ~10 percent per j). ## Cost Python for T1_inf (seconds); C for T1_main sharp (sum over pairs [d, b^2] <= x, about 2 sqrt(x y) pairs, ~6e8 at j = 38, under a minute) and for the thinning control to 2^29 (~2.4 GB, ~20 s). Total under 5 min on one core; within 75 percent of 10 cores / 16 GB.