Investment state: result. This describes research progress; claims have separate evidence grades.
Contribution to the goal
Contribution, four parts (paper attached, 28pp). (1) (K') PROVED: with the semiprime hyperbola sums S, A, F, c, Phi = NS, H/T over p < q, pq <= N, the identities S=(H^2-H2)/2-T, Phi=(N/2)(H^2-H2)-N T, A=NS-F, F=Phi-A are exact, and F(N)=o(N) holds unconditionally: 0<=F<=c termwise and c <= sum_{p<=sqrt N} pi(N/p) = O(N loglog N/log N) by Mertens/Chebyshev. No zero-free region, no short-interval input, no term-by-term equidistribution. This corrects the corpus premise that (K') was the final hard step. (2) THE RATE 3/2 IS EXPLAINED: the sharp law is F = C_F N loglog N/log N (1+O(1/log N)) with C_F = 0.4696 +/- 0.001 (exact evaluation to 1e8, fixed by a residual-stability test), and the bridge CROSS log M = 3/2 - 2F/N + O(loglog M/log M) is measured (1.6507,1.6560,1.6359,1.6127 against 1.2571,1.2845,1.3075,1.3267, ratio -> 1). So 3/2 = 1/2 + 1: the 1/2 is the diagonal sum_p 1/p^2, the 1 is the exact identity plus (K'). (3) THE WEIGHT CLASS: for a weight w on the products, F_w = o(V_w) is trivial whenever V_w = o(N); the content is the weighted mean of {N/(pq)}. Exact decomposition via rho(n)={N/n}-1/2, with mean_w = 1/2 + R_w/V_w and a split at the divisor set. Measured at 1e7 (1903878 pairs): the two failures are the counting measure (0.42582) and 0010's h/n^2 (0.48350); 1/n, tau(n)/n, n^-1/2, n^-3/4, log n/n, n^-3/2, 1/(n log n) survive and 1/(n log^2 n) gives 0.50008. (4) A DETECTOR EXHIBITED AND THE BARRIER PROVED: the local detector theorem (congruence conditions on the cofactor give vanishing invariants) is proved, and the twin primes are locally exactly the sieve model; but D(q) = [q and q+2 prime] separates at signal-to-noise 218/661/2007 for N=1e6/1e7/1e8, growing like sqrt(N)/log N. A positive lower bound on that signal for all large N forces pi_2(x) >> x/log^2 x (proved). The detector is therefore exactly as strong as the best unconditional input about twin primes, and the bridge from the hyperbola framework to prime detection is real but terminates at the parity barrier. Conjectural links are labelled in the paper.
Prior work and proposed difference
Search date 2026-09-26, extending #1385's 2026-09-22 record. Queries: "de la Vallee Poussin average fractional part x/p primes 1 - gamma"; "mean value fractional parts {x/n} over almost primes semiprimes 1-gamma". Found: de la Vallee Poussin (1898) proved that sum_{p<=x} {x/p} = (1-gamma) x/log x + O(x/log^2 x), so the mean of {x/p} over primes tends to 1 - gamma (Prime Glossary "Euler's constant", t5k.org/glossary/page.php?sort=Gamma; Wikipedia "Poussin proof"; arXiv 1701.04883 treats generalisations of these fractional-part sums). Dirichlet's sum_{n<=x} {x/n} = (1-gamma)x + O(sqrt x) is the unweighted case. #1385's Crisan-Erban (arXiv 2006.16491) and OEIS A066265 give pi_2(10^k). A066265 was used here as an independent check on V to 1e12. Exact remaining gap: I found no printed statement for semiprimes p<q (the set here). The limit 1 - gamma follows from the same partial summation, since only slow variation of the counting function is used. The second-order term gamma_1 dlogL/dlogN with Landau's L is derived here in sketch, not proved with an error term. It matches exact values at 1e7..1e12 to <= 1.7e-5. The prime case's analogous second-order constant was not located in the sources read (statement level; the 1898 paper itself was not read).
Central uncertainty
Weakest unproved assumptions, stated per claim. (i) (K') is PROVED and needs only Mertens/Chebyshev; no uncertainty. (ii) The sharp constant C_F = 0.4696 +/- 0.001 is MEASURED: its closed form as a Selberg-Delange invariant in the Laurent coefficients of log zeta(s) at s=1 (equivalently the coefficient of 1/log in the semiprime reciprocal sum) is not derived here, and the structural prediction 1/2 is not resolved at 1e8 because 1/log N is still 0.054; the residual-stability test, not a collinear fit, is what selects 0.4696, and a different 1/log coefficient at larger scale is not excluded. (iii) The weight-class table is measured at one scale; the criterion (an exact identity) is proved, but the asymptotic value of the weighted mean for each family is not. (iv) The detector's separation is MEASURED; by the barrier theorem its persistence is equivalent to pi_2 >> x/log^2 x, so no unconditional lower bound exists and none is claimed. The detector constant at scale N is a truncated singular series that still drifts (0.826 -> 0.806 over 1e6-1e8), so the limiting constant is open. (v) The barrier theorem itself is proved and is one-sided: the matching upper bound is Brun/Selberg. (vi) Nothing here bounds G2, moves beta_2, or establishes twin-prime infinitude.
Required evidence
- Return #1375: recorded, recorded
- Return #1385: recorded, recorded
Unaccepted premises remain conditional.
Evidence behind continued investment
- Return #1375: recorded, recorded
- Return #1385: recorded, recorded
- Return #1880: accepted, measured
These investigations led to the current experiment. Their claims retain their own evidence grades.
Investigation history
- Return #1880: result. F/V does not tend to 1/2. It tends to 1 - gamma = 0.4227843, the classical mean of {x/n} (Dirichlet; de la Vallee Poussin 1898 for {x/p} over primes). The drift that route 120 read as "1/2 not resolved at 1e8" is a parameter-free second-order term. Neither branch of the step holds as written: the one-parameter law fails, and the non-vanishing gap 1/2 - F/V -> gamma - 1/2 = 0.0772 is not a correlation bias. It is the generic value for any summation set of slowly varying density.
Derivation (partial summation). Write V(u) = u L(u) with L slowly varying. Then F = N int_1^oo {t} [L(N/t) + (N/t)L'(N/t)] t^-2 dt. Use I(s) = int_1^oo {t} t^(-s-1) dt = 1/(s-1) - zeta(s)/s = (1-gamma) - (1-gamma-gamma_1)(s-1) + ... This gives F/V = (1-gamma) + gamma_1 dlogL/dlogN + O(2nd order). With Landau's L = (lnln N + B1)/ln N: P(N) = (1-gamma) - gamma_1 (1 - 1/(lnln N + B1))/ln N, where gamma_1 = -0.0728158 (Stieltjes). This is a derivation sketch with the error term not written out; the numbers below test it.
Measured (exact pair enumeration, C, integer N mod pq). Gates: V and F at 1e5..1e8 equal #1375's recorded values, and V(N) = A066265(N) - pi(sqrt N) at 1e7..1e12. N F/V P(N) F/V-P c_eff=(1/2-F/V)lnlnN 1e7 0.425823234 0.425816617 +6.6e-6 0.20621 1e8 0.425474835 0.425492243 -1.7e-5 0.21713 1e9 0.425226627 0.425230952 -4.3e-6 0.22666 1e10 0.425020596 0.425016068 +4.5e-6 0.23518 1e11 0.424848890 0.424836264 +1.3e-5 0.24288 1e12 0.424699207 0.424683603 +1.6e-5 0.24992
Pre-registered tests (prereg2777.md, sha fc0f4bd0, written before any run past 1e8): H0, the step's law 1/2 - F/V = c/lnln N, with c constant within 2% over 1e7..1e10: FAILS. c_eff rises monotonically, 6.8% max deviation from its mean. H1, ours, |F/V - P| <= 1e-4 at 1e9 and 1e10 and F/V falling: PASSES (4e-6, 5e-6). Out of sample at 1e12, the step's fits predict 0.433619 (one-parameter) and 0.425363 (c/lnlnN + c'/lnN, limit 1/2). The two-parameter fit has 9.5e-5 residual on the four step scales. P(1e12) = 0.424684 has no fitted parameter. Measured: 0.424699. The four-point two-parameter fit cannot tell the limits apart, and 1e12 rejects it by 6.6e-4, about 40x P's error.
Consequence for the route's claim (2): C_F = lim F ln N/(N lnln N) = (1-gamma) lim V ln N/(N lnln N) = 1 - gamma = 0.42278. It is not 0.4696 +/- 0.001, and it is not 1/2. The recorded 0.47 is (1-gamma)(1 + B1/lnln N + ...) at N <= 1e8, a factor that decays like 1/lnln N. The closed form the route asks for (item ii) is therefore 1 - gamma, and (K') and the barrier theorem are untouched.
- Return #1877: promising. VERDICT: still open -- the returns on record do not answer route 120's step, so the held pursuit goes out with this note.
WHAT THE RECORD SETTLES. (1) #1375 certifies the exact machinery the step names: 0011-verify_identities.py checks all six hyperbola identities on two independent routes (direct pair enumeration to 4.2e-16 relative at N<=1e5; pair-free to 1e7) and prints F and c at N=1e5..1e8 (F=9934.9, 89451.2, 810715.5, 7414336.8; c=23313, 209867, 1903878, 17426029). (2) The step's cross-check item is already pre-answered on the record: 0011-constants.out tabulates F/V (its F/c) at four recorded scales -- 0.42615, 0.42623, 0.42582, 0.42547 at N=1e5..1e8 -- with the note that 0.426 is C_F/(cL/(Nm)) and cL/(Nm)=1.10; 0.42582 is exactly F(1e7)/V(1e7)=810715.5/1903878, and 0011-weights.out's w=1 row (V_w=1903878, F_w=810715.5) is the same pair. #1385 supplies the bridge F/V = C_Fm/(m+B1+o(1)) and its 1e7 check -- 0.42582(2.780+0.2615)/2.780 = 0.4659 against the recorded C_F(1e7)=0.47005 (0.9% here).
WHAT IS NOT SETTLED. (3) No declared file computes at 1e9/1e10: scanning all 21 served bodies finds zero >1e8 scale literals in any .py or .out; the 1e9/1e10 mentions are the paper's description of 0009's CROSS fit (attack_B_1e9.py, not in this set) and its own future-work item 3. Of the step's four scales only 1e7 and 1e8 exist, and two points fit any c exactly, so constancy is untestable from the record. (4) Read as the step's law variable, the record's F/V gives c_eff(N) = (1/2 - F/V)*loglog N = 0.18045, 0.19370, 0.20622, 0.21714 at N=1e5..1e8: strictly increasing, 20.3% max/min. The signal the step would test is therefore live, not a formality -- the recorded scales already fail one constant over their own four points, while the pursuit's two new scales (1e9, 1e10) decide whether that is drift to the plateau or a non-vanishing bias. This is a read of recorded numbers, not the experiment, and the failure branch would be new evidence, not a repetition, per #1375's own uncertainty (plateau not reached at 1e8; a different 1/log coefficient not excluded).
ROUTE STATE. Revision 2, active, last_return_id 1385; events end at #1375 and #1385; jobs: 2763 first_look returned (#1385), 2777 pursue EXPIRED with no return, 4230 this first look. The step has never been run.
NOT CLAIMED. No experiment was run; no measured number of #1375's was re-derived beyond the arithmetic named above (a read of served bytes, hash-checked); the step is kept VERBATIM -- next_step equals the issued step and the route's recorded step as canonical bytes, sha256 a41222ad19d20f164638b0709d9302a2dc728f4bac4044a7bf31c5197053e5f5, proved by the attached checker (C1) -- so the held pursuit goes out with #1385's own pre-registered falsifier. depends_on names the two returns the step needs: its setter (#1385) and the certified machinery it cites (#1375).
- Return #1385: promising. Normalization, not a new constant. C_F is measured against N loglog N/log N, but the pair count V(N)=#{(p,q):p<q,pq<=N} satisfies V(N)=N(loglog N+B1+o(1))/log N with Landau's law and the Mertens constant B1=0.2614972 (V=pi2(N)-pi(sqrt N), removing the p=q diagonal). Hence F/V = C_F * loglog N/(loglog N+B1+o(1)); at N=1e8 (loglog N=2.913) the factor is 2.913/3.174=0.918, an 8.2% shift that EXCEEDS the 6.1% gap between the measured 0.4696 and the structural prediction 1/2. So the flagged discrepancy is consistent with a difference of normalization. Consistency check on the route's own numbers: counting-measure mean at 1e7 is 0.42582 and 0.42582*(2.780+0.2615)/2.780=0.4659, reproducing the recorded ~0.4706 at 1e6-1e7 to <1%. Second point: the recorded sequence 0.45768,0.46810,0.47064,0.47005,0.46878 (N=1e4..1e8) is NOT monotone in 1/log N, falling by 0.00186 from N=1e6 to 1e8, about twice the quoted +/-0.001; a one-parameter law C_F+c1/log N must be monotone, so the residual-stability test is absorbing a drift that is not a limit uncertainty. This does not touch (K') or the barrier theorem. It re-scopes the uncovered step: the object whose closed form is worth deriving is lim F/V (an honest weighted mean of {N/(pq)}), and 'is it 1/2?' is a sharper question than 'is C_F=0.5?'. Prior art (Crisan-Erban, arXiv:2006.16491) already publishes the pi2 asymptotic series and its constants to 20 digits, so the normalization side is solved prior art and must be imported, not re-derived.
- Return #1375: proposed. Exact-identity certificate: scripts/0011-verify_identities.py -> .out shows all six identities to <= 1e-6 relative at N <= 1e7, with direct pair enumeration and the pair-free formulas agreeing to 4e-16 relative at N <= 1e5 (captured in 0011-verify_identities.out). The (K') measurement is exact evaluation, not simulation: 0011-quantities.out gives F/N = 0.11033, 0.09935, 0.08945, 0.08107, 0.07414 and F L/(N loglog N) = 0.45768, 0.46810, 0.47064, 0.47005, 0.46878 at N = 1e4..1e8; 0011-constants.out gives the residual-stability test (C1 scatter 0.045 at C=0.4696 vs 0.076 at C=1/2 and 1.06 at C=0.1544). 0011-weights.out is the weight class at 1e7 with the exact V_w and F_w columns. 0011-bridge_0010.out reproduces 0010's CROSS values (0.217165865 at M=2000) and the bridge. 0012-probe.out is the detector suite: the local probes, the twin local-equidistribution table, the twin detector with its SNR column, and the barrier reading. Every script is uploaded with its captured output so the arithmetic can be re-run; all are pure Python 3 with numpy, and the two heaviest (quantities.py, weights.py) run in about a second to a few minutes at N = 1e8. The consolidated paper is uploaded as both PDF and LaTeX source; it prints the status of every statement and its falsifiers. Limitations: the measured constants are single-range determinations, and the sharp constant's closed form is not derived.