# A finite statistic for the signed singleton remainder, with a pre-registered falsifier Job #1393 (explore, discovery, lane formalize), attempt `807baf450b9ff6195cfef926378aaa4d`. Rung per section is stated in place. Producer of the admissibility figures: `new-statistic-degeneracy.js` (output reproduced in §5 and in `degeneracy.json`). ## 0. What is already on the record Cited by file, because the design is a response to what these three records could not decide. - `singleton-fiber-audit.md` (Q-singleton-fiber-audit). Establishes, from named classical inputs (prime BV after Tao Notes 3 Thm 17, PNT, Σμ = o(t)), that each sign's **ungrouped** singleton mass satisfies P₁(x) ≥ c·x·log x and N₁(x) ≥ c·x·log x eventually, with c = 1/40960 a non-optimised choice. Its own verdict: "The signed singleton and longer-fiber contributions remain OPEN". The result "refutes O(x) control of the ungrouped negative singleton mass, not cancellation of the signed remainder". - `corner-measurement.md` (Q-corner-measurement). MEASURED at x ≤ 2³⁶. Its **leading** result is disconfirming for the measurement, not the corner: at every reachable x the actual right band (⌊x^(1/20)⌋, ⌊(x−2)/(E₁+1)⌋] contains at most one prime and is empty at several dyadic blocks, so the actual-parameter rows are "a one-prime object rather than the asymptotic corner". Second reading: |K|/mass falls no faster than the matched random-sign control. - `kernel-sign-control.md`. MEASURED at x ≤ 2³⁰ on the (8/25, 9/20) box: no advantage over random signs in any of four families, in a regime where the R = 0 class carries about all of the moment and the kernel is invisible — "a failure to detect, not a refutation". ## 1. The gap, and the decision the statistic informs The audit bounds P₁ and N₁ *separately, from below*. Two quantities that each exceed c·x·log x admit **any** signed difference in [−(P₁+N₁), P₁+N₁]; those inequalities therefore decide nothing about R₁ = P₁ − N₁, its size, or its sign. That is the uncovered question, and it is stated OPEN by the audit itself. **Decision informed.** Whether the signed singleton remainder carries arithmetic structure — cancellation below the random-sign null — at a scale a run can actually compute. This is the only kind of finite evidence that can support or undercut the heuristic that E_⧫ is payable, and it is the input the audit leaves OPEN. A reading of "no structure" is a real decision: it removes an anticipated piece of support and tells the programme not to price the signed remainder by assumption. Why not reuse the corner statistic for it: §5 measures that the corner statistic is *inadmissible* at every reachable x, while this one is admissible at every reachable x. That is the concrete difference between the two proposals. ## 2. The statistic Cutoffs and partition read from the audit §1, unchanged: U = V = ⌊x^(6/25)⌋, Y = Z = ⌊x^(1/20)⌋, J_x = (x/2, x] ∩ ℤ, β_W(m) = Σ_{r|m, r>W} Λ(r) (r prime), β_V(k)β_Z(v) the weight of a fiber, and P₁ = Σ_{|I_{k,v}|=1} 1[μ(d)μ(e)=+1] β_V(k)β_Z(v), N₁ = Σ_{|I_{k,v}|=1} 1[μ(d)μ(e)=−1] β_V(k)β_Z(v), R₁ = P₁ − N₁, over the fibers I_{k,v} = {(d,e) ∈ ℕ² : d>U, e>Y, dk−ev=2, dk ∈ J_x}. Define, per dyadic block, S(x) = R₁(x) / sqrt( Σ_{|I_{k,v}|=1} (β_V(k)β_Z(v))² ), ρ(x) = |R₁(x)| / (P₁(x) + N₁(x)) (signed / unsigned), M(x) = P₁(x) + N₁(x), Σw²(x) = Σ (β_V β_Z)². **S is the signed sum normalised by its exact random-sign null standard deviation** — the same object `corner-measurement.md` calls "the exact second-moment standard deviation sqrt(sum w²)". Under the random-sign null S has mean 0 and variance 1 by construction, so |S| ≤ 2 is the pre-declared null band and no simulation is needed to interpret it. The 16 seeded draws exist to check that normalisation, not to replace it. Rung: **conjectured / design.** No value of S has been computed here. ## 3. Pre-registered falsifier — fixed before any run Written now, before any block is computed, and not to be modified after: - **F1 (primary).** If |S(x)| ≤ 2 for every admissible dyadic block x ≤ X_max, the statistic is consistent with the random-sign null at every scale reached. The report then says "no detectable signed structure" **first**, claims no heuristic support for cancellation in the signed remainder, and does not report a rate. - **F2.** Fit s = d log(|R₁|/M) / d log x by least squares over the admissible blocks. If s ≥ 0 within the spread of the random-sign control, the run gives no heuristic support for cancellation and says so first. If s < 0, the reading is "consistent with" at most, and s is compared with the control's own decay: a random-sign sum of N terms decays like N^(−1/2) relative to its mass, so **a decay no faster than the control's is not evidence of arithmetic structure**. (F2 is `corner-measurement.md` §2's rule reused verbatim, because it is the correct rule and re-deriving a weaker one would be a defect.) - **F3 (sign).** sign(R₁(x)) is pre-registered per block. If the block signs are not balanced beyond the control's spread, that is anti-cancellation evidence and must be reported as such, not summarised away. - **F4 (admissibility).** A block is admissible only if its right weight is non-zero on more than one n in the block. Inadmissible blocks are reported as inadmissible and excluded from F1–F3, never silently averaged in — which is the failure mode §5 shows the corner statistic cannot avoid. ## 4. Matched control Three controls, in the repo's own style, all pre-registered: 1. **Random-sign.** 16 seeded draws replacing μ(d)μ(e) by ±1 with equal probability while preserving every magnitude, so the mass is unchanged and only the signs move. Report the mean and sample sd of S across draws; the mean must lie within a few standard errors of zero and the sd within a few percent of 1, or the normalisation in §2 is wrong and the run stops rather than reading F1. 2. **Independent thinning.** Keep each term independently with probability 1/2. The null normalisation predicts E[S] = 0 and sd(S) = 1 at every thinning level; a departure is a defect in the statistic, not a finding about Möbius. 3. **Negative control.** Empty the e > Y band. Then no fiber is a singleton, R₁ = P₁ = N₁ = 0 and S = 0 with zero contributing terms. (Reused from `corner-measurement.md`'s pre-registered control (iv).) ## 5. Admissibility, measured here — and why this statistic does not inherit the corner's degeneracy The two admissibility criteria differ in kind, and that is the design's whole claim: - (A) corner statistic: needs a **prime in the narrow interval** (⌊x^(1/20)⌋, ⌊(x−2)/(E₁+1)⌋] with E₁ = ⌊x^(93/100)⌋ — a *band* condition. - (B) this statistic: needs a **prime factor above ⌊x^(1/20)⌋ in the cofactor v** — a *support* condition. `new-statistic-degeneracy.js` computes both exactly for dyadic x. Rung: **measured** (exact integer sieving, no floating-point decision on an integer condition; 1.83 s wall, peak 111 MB, run under the Windows job object at --timeout 300 --mem-mb 2048 --cpu-s 240, limit states `enforced`). | x | (A) primes in band | (A) admissible | (B) v in (x/2,x] with a prime factor > ⌊x^(1/20)⌋ | (B) admissible | |---|---|---|---|---| | 2²⁰ | 0 | **no** (band empty) | 524287 / 524288 = 0.999998093 | yes | | 2²² | 0 | **no** (band empty) | 2097151 / 2097152 = 0.999999523 | yes | | 2²⁴ | 1 (the prime 3) | **no** (one-prime object) | 8388607 / 8388608 = 0.999999881 | yes | | 2³⁰ | 1 (the prime 3) | **no** | not computed (above the exact-sieve range) | — | | 2³⁶ | 1 (the prime 5) | **no** | not computed | — | The single inadmissible cofactor in each block is the top power of two (v = x, largest prime factor 2), which is exactly the one v with no prime factor above ⌊x^(1/20)⌋ = 2. So the singleton criterion fails on 1 integer per block where the corner criterion fails on all of them. This reproduces `corner-measurement.md` §3 independently — the band holds ≤1 prime and is empty at 2²⁰ and 2²² — and shows the degeneracy is a property of the *band* condition, not of the scale. Beyond the exact rows the claim is qualitative and stated as such: for Z = x^(1/20), the integers v ≤ x with no prime factor above Z are the Z-smooth ones, counted by Ψ(x, x^(1/20)) = x·ρ(20) in Dickman's convention, which is astronomically small — so (B) is admissible for all but a vanishing proportion of v, at every scale. **Not claimed:** that ζ-Dickman applies at every scale here; the exact rows above are what is measured, and the extension is the standard heuristic marked as such. ## 6. The scale at which the effect would be visible Sensitivity is a property of the statistic and is decided before any data: with N_eff(x) = (Σw)² / Σw² the effective term count, a relative signed bias ρ is detectable exactly when ρ · sqrt(N_eff(x)) ≳ 2, i.e. ρ_min(x) = 2 / sqrt(N_eff(x)). So each block reports its own floor ρ_min(x), and the target bias must be compared with it. For illustration only, taking the reference case N_eff ≈ x/10 gives ρ_min ≈ 6.3/√x, which is below 1/log(x)^K for K ≤ 4 already at x ≈ 10⁸: | target bias | smallest x with ρ_min(x) ≤ 1/log^K x (reference N_eff = x/10) | |---|---| | 1/log x | ≈ 4·10³ | | 1/log² x | ≈ 10⁵ | | 1/log³ x | ≈ 3·10⁶ | | 1/log⁴ x | ≈ 10⁸ | **Rung: conjectured.** N_eff is *not* measured here, and the table's dependence on N_eff ≈ x/10 is an assumption; the run must measure N_eff per block and report the floor it actually got. The consequence is honest in either direction: at reachable x the statistic's sensitivity is not the binding constraint, so a null reading from F1 is informative rather than a statement about insufficient compute. That is precisely what `kernel-sign-control.md` could not say, having found its regime dominated by a class where the kernel is invisible. ## 7. Cost The pipeline is `C_{U,V}(n)·C_{Y,Z}(n−2)` over n ≤ X — the correlation form already validated in the audit's companion (`singleton-fiber-validation.js`, per `data-reuse-audit.md`) — extended by the R₁ / R_{≥2} partition and the singleton enumeration of I_{k,v} per fiber. Cost is a segmented sieve of μ and the β weights, O(X log X) elementary operations with O(block) memory; there is no CRT or DFS component. | X | single-core estimate | memory | |---|---|---| | 2³² | ≲ 1 CPU-h | bounded (segmented) | | 2³⁴ | ≈ 3 CPU-h | bounded (segmented) | **Does not fit this assignment.** Two hours of wall time with 4 CPU-h of *grant* remains, but the partition and the admissibility gate must be built and validated against the audit's own finite checks before any S(x) is read, and a design that is wrong in the partition would read a false null. This return is therefore the design with its cost, as the brief permits, plus the one decisive measurement (§5) that the design depends on. Rung: **design + measured admissibility; no S(x) computed.** ## 8. What remains open after this - The actual values of S(x), ρ(x) and the F2 slope — the run itself. - Whether the singleton partition R₁ / R_{≥2} at these cutoffs is populated at reachable x (the audit validates the partition on finite intervals but does not report the singleton share at the cutoffs used here). - Nothing in this note changes the OPEN status of the signed singleton estimate, of the sufficient twin margin, or of twin-prime infinitude. No novelty is claimed for the statistic: the random-sign normalisation and the |K|/mass-vs-control rule are the repo's own, and the review convention is Chowla/Elliott-style autocorrelation, e.g. arXiv:2206.12956 (Carella, 2026) for the shift-a autocorrelation of μ and λ — prior art for the *shape* of the question, not for this weighted determinant-2 restriction.