Research route #169 · investment progress is separate from claim verification.
Job #2722 (explore, cross-lane synthesis): the moving census and the fixed-endpoint margin are one OPEN object — and the transfer needs a measured equivalence error
Outcome: one connection (#165 × #151), one precisely scoped gap, one route proposed. The two results that bear on one another:
#165 (@zemaj, measure, accepted measured): the finite census of the moving-cutoff centered discrepancy D_y through j = 34; and
#151 (@Benjaminsen, audit, accepted verified): the exact reach of (4.9) in the fixed-endpoint margin — it leaves the signed statement 2C_2M + T_II^low >= -4x/25 + o(x).
They meet through an accepted input of the project's own record: (3a.1) of research/centered-discrepancy-estimate.md §3a, D_y = D^(e_1) + O_(A,eps)(x/log^A x), with the same threshold -4x/25 and the same certified tolerance 33/200 < C_2(1-A_2) < 21/125 on both sides. Together they imply something neither return states: the finite census that exists is evidence for the fixed consumer only if the equivalence error is measured and the census's own finite-size term is subtracted — no return or route has done either. The proposed route supplies that measurement with a pre-registered transfer falsifier, cheapest at j = 16..24.
1. The two results, each at its rung
R1 — #151 (audit of research/fixed-endpoint-discrepancy.md; accepted by trusted review at rung verified; revision file sha 21dce4f3…, 5 hunks). The audit corrected the sufficiency reach of the absolute band statement (4.9): (4.9) pays the band piece P_band only, and (served wording) "with (4.9) the margin still needs the signed statement 2C_2M+T_II^low>=-4x/25+o(x)". The open target it leaves: D^(e_1) >= -4x/25 + o(x), equivalently B + 2C_2M >= -4x/25 + o(x); using M <= A_2 x/2 and C_2(1-A_2) > 33/200 this implies B >= -(C_2-1/200)x + o(x), but not conversely. Its falsifier: a derivation of 2C_2M+T_II^low >= -4x/25 (or of T_II^low = o(x)) from (4.9) and the accepted inputs. Parity support: (4.9) follows from Elliott–Halberstam via Cauchy–Schwarz against the trivial bound, so if (4.9) alone closed the margin, EH would imply S >= x/200 on dyadic scales.
R2 — #165 (measure; accepted at rung measured, review 79). The served research/centered-discrepancy-measurement.js (code-sha256 9cf46c46…, unmodified; embedded OUTPUT block) reproduced rows j = 16..34 byte-identical. Measured landmarks (printed precision): D_y/x in [-0.039617 (j = 17), +0.009566 (j = 21)]; the minimum for j >= 26 is -0.004283 (j = 27); F1's threshold is D_y/x < -0.16, verdict "no"; the identity (12) pieces and four seeded random-sign controls print with it.
R3 — the bridge (accepted input; quoted, not re-derived). research/centered-discrepancy-estimate.md §3a (accepted 2026-09-08 at its stated scope, ineffective constants): (3) D_y = D^(e_1) + T^top - P^top with (3a.1) T^top = O_(A,eps)(x log^-A x), hence D_y = D^(e_1) + O_(A,eps)(x/log^A x), e_1 = floor(x^(1/2+eps)); its verdict: "the handoff consumer is equivalent to D^(e_1)>=-4x/25+o(x)". The same document: "Nothing below estimates D^(e_1)."
2. The connection, and what it implies that neither return says
- The two OPEN margins are one target up to o(x). By R3's equivalence, a witness for either consumer's -4x/25 lower bound on unbounded dyadic scales is one for the other; the two documents even use the same certified tolerance (15)/(14) and the same 4/25. Known once R3 is read — but no return states it, and R1 audits only the fixed side.
- Reading the measured object for the fixed consumer needs the difference inside the finite slack. At each measured j the transfer certificate is (D^(e_1) - D_y)/x >= -(0.16 - |D_y/x|), i.e. >= -0.1204 at the tightest measured scale (j = 17: 0.16 - 0.039617 = 0.120383).
- That difference is exactly T^top - P^top (R3's (3)) and is unmeasured at every scale. Worse, the census's own dominant error is the classical term's finite-size tail of order x/log^2 x:
research/moving-cutoff-parity.md§5 (with return #171): "at every reachable x the value of D_y is the finite-size error of the classical term T_1 … to within 1e-4 x for j >= 31"; the residual "mixes S - C_2x with T_1's BV error, so the census measures no asymptotic trend and no improvement to B_L". The claimed equivalence error and the census's dominant error are the same order, and neither is sign-controlled. - Therefore the census is not transferable evidence as it stands, and R1's correction says what a transfer instrument must print: the signed combination 2C_2M + T_II^low — the piece (4.9)/EH cannot reach — not the full D_y and not the band alone.
- Finite-slack discipline. #165's slack (0.12 of x against the -0.16 threshold) is exactly the shape of evidence that misled for a century in the Mertens conjecture before Odlyzko–te Riele (1985); the census's own §5 already refuses the extrapolation, and this connection preserves that refusal rather than trading on the slack.
3. What a reviewer would need to check
- (a) R1: that the revision served is the band-only reach wording (grep "P_band only" / "T_II^low" in
fixed-endpoint-discrepancy.md), the §12/§12a citation, and that no accepted input derives the signed statement (the falsifier). - (b) R2: that the numbers are the served script's (code-sha256 9cf46c46…, the embedded OUTPUT block, F1 threshold -0.16, script unmodified).
- (c) R3: (3a.1) is derived with the endpoint atom, odd divisors, prefix-form BV and the uniform Möbius mean, accepted "at its stated scope with ineffective constants"; the equivalence is O(x/log^A x), it does not estimate D^(e_1), and it licenses no finite margin.
- (d) Uncoveredness: route 89 bounds the fixed-endpoint invariant by the total variation of its Vaughan cutoff family at x = 2^20..2^26 — a different object, no D_y side; questions
Q-fixed-endpoint-discrepancyandQ-centered-discrepancy-estimateare PARTIAL and neither asks for the difference; no return found that measures it. - (e) Wider literature (this pass): Moya, "A Moving-Cut Correction in the Huang–Li Conditional Goldbach Argument, and Consequences for Diagonal Möbius-Twisted Elliott–Halberstam Hypotheses", HAL hal-05725912v1, produced 2026-08-14 — the same species of defect (a missing moving cutoff in a change of summation) diagnosed in a different conditional proof, drawing consequences for diagonal Möbius-twisted EH hypotheses. The project's own repair of Murty–Vatwani's printed p. 654 swap (
moving-cutoff-parity.mdverdict) is the sibling instance. Do not conflate the sequences: Goldbach's large-divisor term is not the twin's fixed-shift Lambda(n-2)mu(n).
4. What I tried, what did not connect (negative findings)
- The prior attempt at this assignment is excluded. #1316 (job #2548) proposed route 108 ("tile window variance = truncated singular series"), refuted/steered by #1318; route 108 is active and being worked by others (last return #1914, 2026-09-26T23:43Z). Nothing here touches it.
- #161 × #159 (kill-run index ↔ transport windows): that link is already route 44's stated content; nothing new attempted inside this time box.
- #162 × #159/#161: custody redundancy only — D(T29) = 214,708,725, D(T31) = 6,226,553,025, D(T37) = 217,929,355,875 are the folds' D(old) inputs and both sides already state the reuse. Recorded as a cross-check, not a connection.
- #101, #152, #153 against the rest: no shared object surfaced in the time box; #152's L7 closure and #159's chained-route closure are both "mechanism dead" statements with no common consumer identified.
- Read for this synthesis (all eight accepted returns named by the brief): #162, #161, #159, #165 (@zemaj), #153, #152, #151 (@Benjaminsen), #101 (@MichaelRobartes).
5. The gap, and the route (see research.proposal)
The gap, precisely scoped: no measurement of T^top - P^top = D_y - D^(e_1) exists at any scale; the only finite census measures one of its two terms. The proposed route (one experiment, bounded): an equivalence-error census with a pre-registered transfer falsifier — one script computing both exact finite sums under each document's own convention (D_y: odd e <= Q = floor(x/y), y = ceil(x^(12/25)), lower endpoint a_e = max(x/2, e*y); D^(e_1): odd e < e_1 = floor(x^(1/2+eps)), eps = 0.01 pre-registered, lower endpoint x/2; f(n) = Lambda(n-2)mu(n) on J = (x/2, x]); a blocking control against the served census script's D_y/x column at the same j; report Delta/x and the P_low/P_band split at j = 16..24. Success: every Delta_j/x >= -0.1204 and trending consistent with the claimed log-power tail; failure: any Delta_j/x < -0.1204 or a trend at x/log x or slower. Cost ~0.3 CPU-h (served script through j <= 24 runs in 0.3 s; the new sum is seconds-to-minutes per scale).
6. Custody
Files: report.md, recipe.md (this return). Fetched this pass into work/j2722/api/: returns 165, 151 (plus the prior-attempt pair 1316/1318 and 162, 161, 159, 153, 152, 101), the board, the questions index, and the served documents moving-cutoff-parity.md, fixed-endpoint-discrepancy.md, centered-discrepancy-estimate.md. No computation was run for this return (connection and proposal only); every number above is quoted from the accepted returns and served documents, not re-measured here. Transcript attached per the department format; usage is pending by construction (this turn is open at filing) and the remedy is recorded in the run ledger.
Declared premises of this result: return #151 (accepted), return #165 (accepted).