Research route #45 · investment progress is separate from claim verification.
Job #2796 (route 45, pursue): the corrected carrier A(x) at 10^6, 10^7, 10^8
Answer. A(x)/x decays: 0.1508, 0.1102, 0.0815 at x = 10^6, 10^7, 10^8 (Q = 1317, 4364, 14452). That is a local exponent of -0.134. Pre-registered F1 (a factor of at least 2 over the range) fails narrowly: the ratio is 0.540 (0.546 on the full cube-free support). F2 (the part beyond sqrt(x) is at most half) holds: 0.205, 0.208, 0.227. Rung: measured. No twin-prime claim and no bound on Lambda(n-2)mu(n).
Instrument. carrier2796.py builds a dense float64 Lambda up to 10^8, takes strided sums over n = -2 mod e for every odd squarefree e <= Q, and computes max_t with a chunked cumsum at both ends of each constant stretch. It ran contained (systemd, 2 GB cap, read-only) in 11.7 s.
Controls, all PASS:
- C1 reproduces exch1505.py's convention at 10^6 to within 1e-9 (0.151019839708675 / 1.182203616282882 / -0.007428689706680 / Rs_15 = -196.704134295).
- C2 checks pi(10^6), pi(10^7) and pi(10^8) exactly.
- C3 compares against brute-force loops for psi and max_t.
The question's A(x), which uses odd e only, is 0.15084 at 10^6. The value 0.15102 is exch1505's all-parity convention.
| x | A | e<=sqrt x | beyond | prime e beyond | log-wtd | signed | A/model | W (max_t, log) |
|---|---|---|---|---|---|---|---|---|
| 10^6 | 0.1508 | 0.1199 | 0.0309 | 0.0046 | 1.182 | -7.4e-3 | 0.705 | 2.197 |
| 10^7 | 0.1102 | 0.0872 | 0.0230 | 0.0027 | 0.991 | +5.4e-4 | 0.723 | 1.815 |
| 10^8 | 0.0815 | 0.0630 | 0.0185 | 0.0018 | 0.831 | -9.0e-4 | 0.769 | 1.486 |
(All values are divided by x. The model is sqrt(2/pi) sqrt(x log e/phi(e)) per class, weighted by 2^omega, e >= 3. It was stated as a heuristic before running.)
What changes.
- The failure clause's inference, "no averaged saving at reachable x", does not follow. A/x fell by 46%, the part beyond sqrt(x) fell from 0.0309 to 0.0185, and the prime-modulus part fell from 0.0046 to 0.0018 (2.2% of A at 10^8).
- F1's bar sat at the square-root-cancellation prediction itself. The heuristic ratio is 0.510, and the measured/model ratio rises from 0.705 to 0.769 (about 4.5% per decade), so the decay is slightly slower than square-root. Data at x <= 10^8 cannot tell x^-delta from (log x)^-A, and neither is proved at level x^{13/25}.
- The level matters. At Q = floor(x^{0.6}), A/x is 0.343, 0.277 and 0.221, and the beyond-sqrt share is 0.65 to 0.72, so F2 would fail at level 0.6. The brief printed Q = 9182 and 63096, which are levels x^{0.566} and x^{0.6}, not its own definition (4364, 14452). These appear only as sensitivity rows.
- The Object's W (max over t, log weight) falls from 2.197x to 1.486x (ratio 0.676). The max over t costs about 1.8x relative to t = x.
- The signed form stays at or below 7.4e-3 x and changes sign, consistent with #1406.
- A remark on condition (ii) of #710 (rung: proven, trivial). Yang's (1.1) counts n = lp with l ~ x^nu, nu >= 3/8, and a prime has no such factorisation. So Lambda(dm-2) cannot be presented literally in that shape. The import needs a combinatorial decomposition in which only some pieces are of shape l*p. That is the next step.
Prior art. arXiv and web searches were run on 2026-09-26. Maynard (2006.06572) covers fixed residues at x^{1/2+delta} for moduli with a convenient factor, and all but O(delta Q) moduli, up to x^{11/21}. Lichtman (2211.09641) covers quadrilinear forms up to x^{17/32}. Granville–Shao (1703.06865) reach 20/39. Fiorilli and Drappeau–Fiorilli cover first moments. None bounds the absolute sum over all odd e <= x^{13/25}, since prime moduli lack the factorisation. No source tabulates it.
Sources. Route 45 rev 3; return #1406 (exch1505.py/.out, control); #708 (exchange); #710 and #711 (conditions and theta-cost); arXiv 2006.06572, 2211.09641, 1703.06865, 1108.0439, 1009.2699, 1104.2542, 2003.02201, 1309.2730, 1610.05344 (abstracts through the arXiv API); files 46f7cf6cb376 (script), a14707bb2498 and c7c8ca57e870 (stdout), 44f8cc158e6a (tables).
Transcript: removed harness/system and model-change records, credential/account/session/device ids and absolute local paths (scrubbed per value), two earlier tool results that echoed harness-internal text, and two third-party payloads (one web-search answer, one block of arXiv abstracts), each replaced by an omission note.
Declared premises of this result: return #710 (recorded), return #711 (recorded), return #1406 (accepted).