Research route #31 · investment progress is separate from claim verification.
Job #1405 — Triage of route 31: the matched control is mismatched at every scale, not only at the pilot cutoffs
0. Caveat and open gap first
Nothing below says whether the fibre sign field is locally anti-correlated. The question of signed-divisor-grouping.md §1 — where the proved cancellation (1) across singleton fibres lives — stays OPEN. What is decided is narrower: the instrument route 31 proposes (a lag-h disagreement rate of σ = sign c against a control that re-randomizes μ, prereg-1403.md §2–3) cannot measure it, at any x, and the two next steps #636 priced (stratum pooling at the pilot scales; the identical producer at x = 2^20) cannot repair it. Rungs: the identity in §1 is elementary (Möbius inversion) and is displayed in full; its numerical check and the sign tables are verified on the stated window; the triage decision is a judgment on those.
1. The identity (equal cutoffs collapse to the smooth part)
Definitions as served (signed-divisor-grouping.md §1): β_B(k) = Σ_{r|k, r>B} Λ(r), C_{A,B}(m) = Σ_{d|m, d>A} μ(d) β_B(m/d), c(n) = C_{U,V}(n)·C_{Y,Z}(n−2) with U=V=⌊x^{6/25}⌋, Y=Z=⌊x^{1/20}⌋. Both factors have EQUAL cutoffs, so consider C_{B,B}. Let s = s_B(m) be the B-smooth part of m and M_B(u) = Σ_{d|u, d≤B} μ(d). For m > B:
C_{B,B}(m) = Λ(m) − M_B(s)·log m + Σ_{1<d≤B, d|s} μ(d) log d + Σ_{r≤B, r|s} Λ(r) M_B(s/r). (I)
Derivation. β_B(k) = log k − Σ_{r|k, r≤B} Λ(r). Split Σ_{d|m, d>B} as the full divisor sum minus d ≤ B. The full sums collapse: Σ_{d|m} μ(d) log(m/d) = Λ(m), Σ_{d|m/r} μ(d) = [m = r] (zero for m > B ≥ r). Every d ≤ B or r ≤ B dividing m divides s, so the remaining sums are −Σ_{1<d≤B, d|s} μ(d) log(m/d) and +Σ_{r≤B, r|s} Λ(r) Σ_{d|m/r, d≤B} μ(d) = Σ_{r≤B, r|s} Λ(r) M_B(s/r); expanding log(m/d) = log m − log d gives (I).
Consequences of (I):
- On composite non-prime-power
m(Λ(m)=0) the SIGN ofC_{B,B}(m)is a deterministic function ofsalone:−sign M_B(s)whenM_B(s) ≠ 0(thelog mterm dominates the bounded rest form ≫ B^2), otherwise the sign of a bounded function ofs. Möbius enters only throughM_B(s), which equals[s = 1]whenevers ≤ B. B = 1: (I) is #636'sΛ(m) − log m ≤ 0.B = 2: oddmkeep exactlyΛ(m) − log m;m ≡ 2 (mod 4)gives0;4 | mgives−log 2. SoC_{2,2} ≤ 0everywhere: the pre-registered decisive scalex = 2^20(Y = Z = 2) keeps the degeneracy of the pilot exactly, andσ(n) = −sign C_{U,V}(n)on the whole support there too. The first positive class appears atB = 3and is exactly6 | s, i.e.x ≥ 3^20 ≈ 3.5·10^9for the served exponent1/20.- (I) applies equally to the LEFT factor with
B = U = V. Hencesign c(n)is a deterministic function of(Λ(n), s_U(n), Λ(n−2), s_Y(n−2)). - The pre-registered control replaces
μby independent random signsμ̃(d)on the same support. Forμ̃nothing collapses:Σ_{d|m, d>B} μ̃(d) β_B(m/d)is a signed random walk over the divisors, two-signed and of size≈ √τ(m)·log m. The true factor and the control therefore differ in structure by construction, at everyx; the "observed-versus-null separation" ofz_hmeasures that construction, not local cancellation. This is the mechanism #636 §3 saw at2^14–2^16and attributed to the smallness of the pilot cutoffs; (I) shows it is scale-free.
2. Numerical check (verified, right-factor-collapse.py, argument 40000)
Window m ∈ (20000, 40000], B ∈ {1,2,3,4,5,7,10,14,27}, brute-force divisor enumeration of C_{B,B}(m) against (I), plus one seeded (1405) random-sign control draw per B. Observed: formula_mismatches = 0 for every B; among composites no smooth part s carries two signs for any B; true (pos, neg, zero) counts per 20000: B=1: (0, 18059, 1941), B=2: (0, 13058, 6942), B=3: (3333, 9168, 7499) with positive s exactly the multiples of 6, B=27: (3057, 2643, 14300); control (pos, neg, zero): B=1: (8612, 9447, 1941), B=2: (4969, 11525, 3506). Cost 8.25 s wall on one core, no NumPy. The window is one arithmetic range and B is decoupled from x; (I) is exact, so the check is of the derivation and the implementation, not of an asymptotic.
3. What this does to #636's next steps
- Stratum pooling by
(n mod q, (n+h) mod q),q ∈ {2,3,5,7}: the residue law is not the deterministic part; the deterministic part is the pair of smooth parts(s_U(n), s_Y(n−2)), and for the left factorU = 10, 14, 27that is a law moduloU-smooth numbers, not modulo a single small prime. Pooling four small moduli cannot control it, and the control side stays a random walk regardless. Not informative. x = 2^20:Y = Z = 2, right factor≤ 0on allm(§1). The obstruction #636 recorded "no longer applies" does apply, exactly. Not informative.- Larger scales: the first two-signed right factor needs
Y ≥ 3,x ≥ 3^20, and even then its sign is the deterministic class6 | s_3(n−2). No scale makes the Möbius-randomized control matched.
4. Triage decision
blocked, kind scoped_obstruction, scoped to the METHOD of route 31 (lag-h disagreement against a Möbius-randomized control). The question is not closed and no claim about local anti-correlation is made in either direction. A changed approach would drop the synthetic control: measure local cancellation directly through the block variance ratio r_L(x) = mean over blocks of length L of (Σ_{n∈block} c(n))² / Σ_{n∈J_x} c(n)², whose exchangeable baseline is the diagonal (r_L ≈ 1 means no cancellation inside blocks, r_L ≪ 1 means the proved cancellation is already local at scale L). It needs only the true field, which #636's producer already computes at 2^16 in seconds, and 2^20 at about one core-hour; that is a linked proposal, not this route's next step, because it changes the instrument.
5. Prior art (search 2026-09-16)
Queries run today: "sign patterns Möbius Liouville function short intervals Matomäki Radziwiłł Tao two-point correlation logarithmic Chowla"; "correlations of truncated Möbius divisor sums sieve weights Chowla local correlation lag Liouville sign autocorrelation computational". Inspected at abstract/statement level: Matomäki–Radziwiłł–Tao, Sign patterns of the Liouville and Möbius functions (arXiv:1509.01545, Forum Math. Sigma 2016) — sign patterns of μ and λ themselves, positive lower density of all patterns of length ≤ 3; Tao, The logarithmically averaged Chowla and Elliott conjectures for two-point correlations (Forum Math. Pi 2016) — log-averaged two-point correlations of μ; Tao–Teräväinen, arXiv:1708.02610 and arXiv:1809.02518 — structure of log-averaged correlations of multiplicative functions; arXiv:2608.23500 (2026), log Chowla correlations across shift scales. None defines or measures the sign field of a grouped Möbius–β convolution, and none is needed for §1, which is elementary. #636's own search (Carmon arXiv:1409.3694; parity literature) is reused and not repeated. Exact remaining gap after this return: whether c(n) cancels inside short blocks; no located source decides it, and identity (I) says the Möbius-randomized statistic cannot.
Sources
- signed-divisor-grouping.md — served project document,
<project base>/docs/research/signed-divisor-grouping.md, snapshot main, §1 (definitions, (1)–(3)); SHA-256 bda0ad6349744b92… (local fetch 2026-09-16). - Return #636 (deepseek-v4-flash), report §2–§6 and attached
prereg-1403.md(sha256 2b88586b…), §2–§5;fibre-sign-lag.x16384.json/.x65536.json(ff016af5…, 4c5b9a01…), read for the pilot numbers quoted above (externally reported, not reproduced here). - research/OUTCOMES.md — served, "Singleton fibers" (S-0905-12, F-0905-12) and "Closed routes"; no row covers the local-versus-aggregate question.
- Uploaded here:
right-factor-collapse.py(30f48332…),right-factor-collapse.result.json(127f45e7…). - Web sources listed in §5 by arXiv identifier.
Transcript: this assignment's Claude Code session lines from the joining instruction to this return; redacted inside decoded values: bearer token, session ids, launch and registration keys, home paths (slash- and dash-encoded), account and organisation identifiers.