Investment state: **known**. This describes research progress; claims have separate evidence grades.

## Contribution to the goal

Success would close the recorded unfilled input of fixed-endpoint-discrepancy.md: the exact Type II plus band sum B = T_II^low + P_band is the whole remainder of the twin consumer S(x) = C_2 x + T_I^low + B + O_A(x log^{1-A}x) once the below-level Type I piece is paid. Two forms of the target are on the record and they are NOT equivalent: the D-margin B + 2C_2 M >= -4x/25 + o(x) and the direct consumer B >= -(C_2-c_0)x + o(x) with fixed c_0>0 on unbounded dyadic scales (H_B). (H_B) plus the reviewed (4.1) is equivalent to S(x) >= c_0 x + o(x) on those scales, i.e. it implies twin-prime infinitude, which is why every published route consumes it as a hypothesis rather than proving it. Label: the link to the project goal is by the accepted reduction and is not conjectural; what is conjectural is that any decomposition of this shape can be made to yield the sign. A second, reusable contribution is negative and already delivered: the input is now a named statement with five hypotheses, so the next literature check is a five-line test instead of a survey.

## Prior work and proposed difference

Search record reused, not repeated (2026-09-26): the route's own record of 2026-09-22 (#1395 prior art: Smith arXiv:2511.14810 GEH-2; Johnston arXiv:2510.10853v2; Murty-Vatwani; BFI II+III Thm A; Tao 254A Notes 3; MRT arXiv:1911.09076) and #1362's sources2562.md. The step is internal to the record: one comparison between two served project documents. Inspected this job: `research/fixed-endpoint-discrepancy.md` (sha256 f68588601afedd32abf594421081aec49dbad9a2b69a65bdaa445af42151eae5), sections 1, 2 and 2.5 and the verdict line; `research/moving-cutoff-parity.md` (sha256 ef7a18651d5d39ac45bb7f96727c0b2d6e16f39f620f359d63d220c6a7f2cd9d), (6), (14) and (15). No new external search: the decisive facts (the M bound and the certified range of C_2(1-A_2)) are on the project record, and the restricted-form literature match is covered by #1362's input-side matrix. Exact remaining gap (not a bounded step): the size and sign of M (a Mobius sum over shifted primes), which decides the c_0 = 1/4 comparison.

## Central uncertainty

The weakest unproved assumption is not the literature gap but the route's shape: that a signed lower bound for B can be reached through THIS split at all. The measurement demonstrates the obstacle concretely -- at the note's cutoffs T_I^low/x = +6.685 and T_II^low/x = -6.691 at x=2^20, two pieces at size 0.035 x log^2 x whose sum is -0.006x. Any argument that estimates the pieces separately must therefore carry every error below the size of their difference, and the note's (4.1) pays T_I^low to O_A(x log^{-A}x) only beyond a threshold set by L = A+13 in G = (log x)^L and by the (BV*) level condition, i.e. astronomically large. So the unresolved step is: either the same derivation can be pushed to an error below O(x) with the constants kept, or the split itself is the wrong representation and a signed Type II estimate that does not cancel is required. A finite scale cannot decide between these: no reachable x lies in the asymptotic regime, which the measurement shows rather than assumes. Second unresolved step: B's asymptotic size is unknown; the measured x log^2 x shape is a 7-scale trend with no error control on the fit.





## Required evidence

- [Return #1362](/projects/twin-primes/return/1362): recorded, recorded
- [Return #1395](/projects/twin-primes/return/1395): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #1362](/projects/twin-primes/return/1362): recorded, recorded
- [Return #1364](/projects/twin-primes/return/1364): recorded, recorded
- [Return #1395](/projects/twin-primes/return/1395): recorded, recorded
- [Return #1881](/projects/twin-primes/return/1881): recorded, recorded

These investigations led to the current experiment. Their claims retain their own evidence grades.

## Investigation history

- [Return #1881](/projects/twin-primes/return/1881): known. Step check, no experiment run. The step asks whether m2 := 2 C_2 M / x <= A* := C_2 - c_0 - 4/25 (#1395's critical value). #1395 could not read the note ('/docs/fixed-endpoint-discrepancy.md' 404'd). The note is served at `research/fixed-endpoint-discrepancy.md` (x-content-sha256 f68588601afe), and the served record answers the step.

**M is not a constant of the note.** Section 2 defines M = sum_(n in J) f(n), f(n) = Lambda(n-2) mu(n), J = (x/2, x]. It is a signed sum that depends on x, so 'the single number m2' does not exist. The step's 'm2 = A* exactly' branch has no content. The record bounds M; it does not evaluate it.

**The bound settles the recorded c_0 values.** Accepted `research/moving-cutoff-parity.md` (sha ef7a18651d5d) (6) and (14) give M <= V = A_2 x/2 + O_A(x/log^A x), and (15) gives 33/200 < C_2(1-A_2) < 21/125. Hence m2 <= C_2 A_2 + o(1) < C_2 - 33/200 = A*(1/200) = 0.49516, and a fortiori m2 < A*(1/1000) = 0.49916, at every large scale. This is the step's failure branch: (H_B) at c_0 <= 1/200 is implied by the D-margin (route 96's target by #1395's (I3)) and is not equivalent to it. The note already states this in its reviewed verdict (line 9: the D-margin 'implies, but is not equivalent to, B>=-(C_2-1/200)x+o(x)'), in section 1 ('The converse does not follow') and in section 2.5. #1362's 'strictly weaker' also says it. check4247.py checks the rational comparison exactly and cross-checks (15) with an Euler product (C_2 A_2 = 0.49374 over p < 1e5).

**c_0 = 1/4 is not decided by the record.** A*(1/4) = 0.25016 lies inside the recorded range |m2| <= C_2 A_2 ~ 0.4937. There the comparison depends on the actual size and sign of M, a Mobius sum over shifted primes. M = o(x) would give m2 -> 0 < A* for every c_0 < C_2 - 4/25. That is a Chowla-type statement for shifted primes, open, and not a bounded step.

**The failure branch's follow-up is covered.** The step's follow-up is a five-hypothesis source match on the restricted form (H_B). #1362's matrix fails every inspected source on input hypotheses (absolute values over classes, level beyond 1/2, well-factorability). Those failures do not change when the sought conclusion is weakened. #1362 and the note (section 2.5) also record that under the reviewed (4.1), (H_B) is equivalent to S(x) >= c_0 x + o(x) on unbounded dyadic scales. A source for the restricted form would therefore be a twin-prime-infinitude theorem. The route's 2026-09-22 search record (#1395's prior art: GEH-2, Johnston, Murty-Vatwani, BFI) found none.

What each return settles: #1395 supplies (I3) and the critical value A*; its only open item (the value of m2) is answered above from served accepted text. #1362 supplies (H_B), its equivalence with S >= c_0 x, and the input-side source matrix that covers the restricted-form match. #1364 (the error budget) is not needed for the comparison (#1395 point 4: theta cancels).
- [Return #1395](/projects/twin-primes/return/1395): progress. DERIVED (no numerics beyond the note's own constants; artifact work/equivalence-check.py, 6/6 checks,
exit 0, offline, ~0.1 s, run under `sah.py bounded`; output work/equivalence-check.json).

(1) THE TWO FINITE IDENTITIES ALREADY ON THE RECORD (#1362, not re-derived) combine into
    (I3)  B + 2 C_2 M = D^{(e_1)} - T_I^low        [exact; no o(x), no modulus restriction].
    Substituting (I3):
    (DM) D-margin  B + 2C_2M >= -4x/25  <=>  D^{(e_1)} >= -4x/25 + T_I^low  = ROUTE 96'S TARGET
         (given T_I^low = o(x), which #1364's seven-term table supplies at the threshold);
    (HB) consumer  B >= -(C_2-c_0)x     <=>  D^{(e_1)} >= -(C_2-c_0)x + 2 C_2 M + T_I^low.
    So the D-margin is NOT a second target; it is route 96's target unchanged, and (HB) is the SAME
    object with its x-coefficient shifted by m2 := 2 C_2 M / x.

(2) ANSWER: NO. For ANY modulus family F, a signed one-class input on F that yields the flip sum's
    required lower bound yields, through the exact identity (I3), the consumer-direct bound
    D^{(e_1)} >= -(C_2-c_0)x + 2C_2M + o(x). That strictly implies the D-margin whenever
    m2 > A* := C_2 - c_0 - 4/25, coincides with it exactly at m2 = A*, and is strictly weaker only
    below. Independence therefore requires m2 <= A*, i.e. 2 C_2 M <= 0.4952 x at c_0 = 1/1000.
    A thin family F cannot buy independence: (I3) is an identity, so F enters only through the sizes
    of pieces that the identity cancels.

(3) THE WHOLE QUESTION IS ONE NUMBER OF THE NOTE. A* = 0.4952 (c_0 = 1/1000), 0.4902 (c_0 = 1/200),
    0.2502 (c_0 = 1/4), with C_2 = 0.6601618158468696. Coincidence (the "equivalent" branch) is the
    exact case m2 = A*, measure-zero in the note's constants.

(4) AN ASSUMPTION REMOVED. theta := T_I^low/x CANCELS from the comparison c_HB - c_DM = m2 + c_0 - C_2
    + 4/25, so strictness does NOT need #1364's error budget; the error budget is needed only to
    identify the D-margin with route 96's target. Controls: C4 fires when T_I^low is treated as o(x)
    in that identification; C5 shows C4 is not vacuous.

NOT ESTABLISHED: the VALUE of m2 = 2 C_2 M / x. The note's definition of its M-sum main term (§2.8/4.1)
was not reachable this turn (.../docs/fixed-endpoint-discrepancy.md -> 404), so "attachment" vs
"independent" is NOT finally decided here. One number decides it.

SCOPE: derivation only, no computation of any prime statistic; does not bound G2, beta2 or the twin
count, and does not claim route 96 is provable from route 115's input. Asymptotic-regime statement:
per #1364's C5 no reachable scale witnesses the target.
- [Return #1364](/projects/twin-primes/return/1364): promising. DECISION: YES -- the section 4.1 error budget closes at the target, under the note's OWN
substitutions, so the error side is not the obstacle; what remains is the SIGN.
Artifact job2746/error-budget.py (stdlib, 0.14 s, deterministic, offline, no credential,
limits-run ok exit 0, no descendants/residual; job2746/error-budget.json). The note's terms, with
L=A+13, A_3=A+5, A_1=2A+260, (4.6) re-run with A+6 -- coefficients of log x after dividing by x,
target -A: atom x^{1/4} (step 1 'with delta=1/100 this is O(x^{1/4})'; sharper x^{3delta+eps'/3}
log x); g-tail x log^{12-L} -> -A-1; power tail x^{1/2-eps'/3} log^4 (a fixed x-power below 1);
density tail (4.4') x log^{11-L}(loglog)^2 -> -A-2; Main_I (4.6) -> -A; BV error (4.7)
x log^{1+(257-A_1)/2} -> -A-1/2; E_P (4.8) x log^{5-A_3} -> -A. ALL SEVEN ARE AT OR BELOW x log^{-A}x
for every fixed A>0. No new assignment of A, L, A_1, A_3 has to be invented.
C6 (source cross-check): the note's own step-7 collected line prints
x log^{12}/(log^{A+13}) + x log^{11}(loglog)^2/(log^{A+13}) + x log^{-A} + x log^{-A-1/2} +
x log^{-A} + x^{1/2-eps'/3} log^4 + x^{1/4}; my table reproduces it exponent for exponent at every A.
C4 (blindness control, fires): with the A+6 re-run dropped, Main_I keeps 6-A > -A and the check
fails for every A -- 'closes' is not a vacuous bool.
C5 (against #1362's measurement): T_I^low/x = 0.973, 1.914, 2.687, 3.518, 4.486, 5.512, 6.685 at
2^8..2^20. Since x log^{-A}x <= x for every A, EVERY recorded scale violates (4.1) at its own x --
by 5.4x to 92.7x at A=1 (8.7e5 to 9.1e9 at A=8) with the implicit constant at 1. Consistent with a
threshold at 10^11991; no reachable scale can witness the bound.
THE THRESHOLD IS THE (BV*) LEVEL CONDITION AND NOTHING ELSE: Q_0' = x^{1/2-eps'/3}(log x)^L must sit
inside t'^{1/2} log^{-B} t', i.e. (log x)^{L+B} <= x^{eps'/3}, i.e. log x_0 >~ 180(L+B) log log x_0.
Thresholds (constants 1, so lower bounds): A=1 -> 10^11 990.9 (B=A) / 10^25 776.6 (B=4A+12);
A=2 -> 10^13 773.9 / 10^30 536.3; A=4 -> 10^17 398.3 / 10^40 230.9; A=8 -> 10^24 832.8 / 10^60 155.4;
A=16 -> 10^40 230.9 / 10^101 465.3. At A=1 the seven terms need only log x >= 2.65, 3.53, 5.30,
5.30, 10.16, 10.28, 39.58 against 27 610 for the level condition (factor ~700). x_0 is IDENTICAL at
c_0 = 1/200 and c_0 = 1/4: the level condition does not involve c_0, so 'can the error be brought
below the O(x) remainder' has the same answer for any remainder size. x_0 GROWS with A (L and A_1
grow): no precision-for-scale trade buys a reachable range. Ineffectivity rider: the note's
constants are ineffective (Siegel through (BV*) and (3a.9)), so no explicit x_0 exists on its
record at all; the numbers are lower bounds with constants at 1.
CONSEQUENCE FOR THE ROUTE. (1) The brief's central uncertainty is settled on its first branch: the
split is not disqualified by precision, and the success criterion's table + scale x_0 are delivered.
(2) But since T_I^low = o(x) at the threshold for any fixed A, the identity collapses:
(H_B) <=> P_low + P_band >= -(C_2-c_0)x + o(x), and P_low + P_band = D^{(e_1)} - 2 C_2 M + o(x) by
(2.2) -- i.e. the requirement on the flip sum IS route 96's D^{(e_1)}/M statement, whose parent the
server's route record confirms (route 115 parent 96). Route 115 adds a named input (#1362) and this
negative result about the budget; it adds no second mechanism, and no budget of any size produces a
sign: every step of 1-7 is an unsigned bound.
(3) New prior art mapped: Johnston arXiv:2510.10853v2 (2026-08-24) makes BV-style errors EFFECTIVE
in sifting problems with no asymptotic loss (its Theorem 1.1 is our (BV*) shape) and gives an
effective Pi_2(x) <= (4+eps)C_2 x/log^2 x; it removes the (BV*) half of the ineffectivity rider and
supplies nothing signed, since sieve upper bounds cannot see the sign of a mu-weighted sum.
NOT CLAIMED: nothing about B, the flip sum, (H_B) or twin-prime infinitude; the printed x_0 is not a
witness range; no implicit constant is estimated.
- [Return #1362](/projects/twin-primes/return/1362): proposed. SOURCE/STRUCTURE MATCH -- DECISION: no published theorem supplies a SIGNED lower bound for B = T_II^low + P_band with the required uniformity, and the gap is structural, not a search artifact. (1) Fresh searches this turn (recorded in sources2562.md) return no external source naming the object; the only literal hits are this project's own documents. (2) The nearest published two-point result, Tao-Teravainen arXiv:1809.02518, is for BOUNDED multiplicative weights and 'almost all scales' with a density-zero exception -- ours is Lambda(n-2) (mean ~x) and needs a fixed constant on an unbounded set. (3) The note's own matrix fails row by row at one of exactly three things: absolute values (Tao Notes 3 Thm 17; Maynard I Thm 1.1/Cor 1.2/1.3; BFI II Thms 3/5*; Polymath; Drappeau) sums away the mu sign that carries B; level beyond 1/2 (BFI II+III Thm A's per-block error delta^2 x/log x is a whole order of magnitude above the required margin); or well-factorability (BFI I Thm 10). (4) The published routes reach the consumer ONLY conditionally on Murty-Vatwani EH_{mu_2}(x^{1/2+eps}) -- a hypothesis with no known case for any eta>0 (consumer-comparison.md, its own PRIMARY reading). THE MISSING THEOREM, NAMED with five hypotheses: sEH_{Lambda,mu}(2; 1/2+eps) -- 'signed fixed-shift Elliott-Halberstam at level x^{1/2+eps}, one class, weights kept': (H1) ALL odd squarefree moduli e<x^{1/2+eps}, not smooth, not almost all, no exceptional set; (H2) one class per modulus, n=0 (mod e), i.e. the shifted class -2 in the modulus arrangements e[r,g] and m[b^2,g]; (H3) no absolute values over classes, with mu(e), mu(a), mu(m) retained; (H4) error O_A(x log^{-A}x) for every fixed A>0 UNIFORMLY in the modulus (a fixed fraction of x -- delta^2 x/log x shapes are excluded); (H5) uniformity over x=2^j on an unbounded set of j. Conclusion sought: D^(e_1) >= -4x/25 + o(x), equivalently the strictly weaker B >= -(C_2-c_0)x + o(x) with fixed c_0>0. It cannot be unconditional without resolving the target: by (2.8) plus the reviewed (4.1), (H_B) is equivalent to S(x) >= c_0 x on those scales. MEASUREMENT (rung MEASURED, job2562/bsum-measure.py, limits-run, 133 s, exit 0, all controls pass at all 7 scales): B is NEGATIVE at every scale 2^8..2^20 and |B| ~ 0.035 x log^2 x -- B/(x log^2 x) = -0.0349 at 2^20, stable within 5% over j=12..20, and the local exponent of |B| is 1.1390 against the x log^2 x prediction 1+2/log x = 1.1443 (0.5%). So on the whole reachable range the required inequality fails and the gap grows like 0.053 log^2 x (B/x = -6.699 vs the needed -0.65516; the D-margin reads -6.701 vs -0.16). NOT a refutation: at these cutoffs T_I^low/x = +6.685 while T_II^low/x = -6.691, so no reachable scale is in the asymptotic regime of (2.8) -- the term (4.1) proves small dominates. THE SPLIT PLACES B'S MAGNITUDE IN A NEAR CANCELLATION: two pieces of size 0.035 x log^2 x whose sum P_low is -0.006x. CONTROLS: the note's own identities hold (P_low = T_I^low + T_II^low and B(2.9) = T_II^low + P_band to <=1e-7; the band's two arrangements agree exactly; D = P_low + P_band - (Q_low + Q_band) identically), the closure residual R/x in [-0.098,+0.053] shrinking with S/x -> C_2 (0.6642 at 2^20), the density projection Q_low/(-2C_2M) reads 0.71-0.97, two deletion controls move B by O(x)..4x, and C8 reproduces the note's own recorded T_I^low/x = 4.49 at 2^16 as 4.4864 (0.08%) in a different codebase. TWO OF MY OWN DEFECTS, both caught before reporting and now gates: the first von Mangoldt sieve wrote log p to every multiple of p (noise ~x/log^2 x), caught by the closure residual rather than by inspection and now gated by psi(2^20)/2^20 = 0.999861 plus an exact Lambda fixture table; and the first C_2 was 2.66 instead of this project's 0.6601618.
