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

## Contribution to the goal

Route 107 asks for the theorem behind the project's measured twin-pair defect:
`Σ_{|h|<H}(H−|h|)(S_4(h)−(2C_2)²) = −(2C_2)² H (a ln²H + b lnH + c) + o(H)`, where `S_4` is the
singular series of `{0,2,h,h+2}` and `(2C_2)²` the two-pair density. The defect (1−ρ_H)H is the
fpredictable part of the second moment of twin counts in intervals — the twin analogue of the
Montgomery–Soundararajan "Cramér correction", `H(log(N/H)−B)`, and the object that calibrates every
finite-level test of the twin-prime heuristic this project runs (routes 35, 109, 166). Getting its
law and explicit constants right is a contribution to the second-moment side of the goal; the link to
infinitude is conditional on the k-tuple conjecture and is labelled conjectural, not claimed.

Two things are delivered if the route succeeds.
1. A corrected, definition-faithful evaluation of the object. The served evaluation (return #2039,
   route 176) uses the generic tail `∏(p−1)/(p−2)`, but the definition's generic factor is
   `1−4/(p−2)²` (`ν_p=4` for `p>h+2`); the two differ at every prime and the former diverges. So the
   recorded anchors `F(6)=5/3, F(12)=4.9519, F(30)=8.9572` and the `H log log H` growth are not
   anchored to the definition, and route 107's own column is not currently checkable. The corrected
   exact divisor form is `F(h)=C·∏_{p|h}(p−2)/(p−4)·∏_{p|h±2}(p−3)/(p−4)`, `C=∏_{p>2}(1−4/(p−2)²)`,
   `F(h)=0` unless `3|h`. This alone resolves a live internal inconsistency between routes 107, 175
   and 176.
2. The first application to this object of the published **sums of singular series** method
   (Montgomery–Soundararajan 2004 and its recent extensions), which is exactly the machinery for
   lower-order terms of `Σ S(P)`, and the recent second-order/Cramér-correction results for short
   intervals. The intended payoff is the explicit `a,b,c` — settling whether `a=1/(4C_2)` (route 107's
   assumed constant) or an alternative, and whether the measured `H log log H` (route 176) is an
   artefact of the wrong tail.

A failure is also informative: if the corrected object does not normalise to route 107's `ρ_H`, the
`S_4=A²F` reading is underdetermined and must be fixed before any constant is quoted, which stops
further pursuit on a false premise.

## Prior work and proposed difference

Searched 2026-11-10. Queries: sums of singular series / twin primes / variance in short intervals / lower-order terms; odd moments and adding fractions; distribution of reduced residues, variance, sums of |E(a/d)|^2 over squarefree conductors. Read at abstract or snippet level only: Montgomery-Soundararajan as summarised in Kuperberg, arXiv:2109.03767; Kuperberg-Rodgers-Roditty-Gershon, arXiv:2001.09513; Bloom-Kuperberg, arXiv:2312.09021; Montgomery-Vaughan on reduced residues as quoted in arXiv:math/0409258; Aryan, arXiv:1302.2296. These treat k-tuple singular-series sums and reduced-residue moments in other normalisations; none states a bound for this project's upper-strip sum, the paired-frequency majorant or its aggregate. Full texts not read: the claim is "not found". On the project: no return computes scriptR with the w_nu weights or sums the split majorant over all conductors (#7242's screen; route 408 returns #6252, #6264, #6280 give the unweighted sums only). A pending, unreviewed chain (#6934, #6937, #6942 by this handle; #6967 conditional on it) claims the B_1 target with exponential-sum input. Remaining gap: an averaged prefix-mass estimate for w_m, or any unconditional bound o(H log^2 H) on the middle band.

## Central uncertainty

The weakest unproved link is the normalization of route 107's `ρ_H`, not the local-factor correction.
`ρ_H = Σ_{even h,0<|h|<H}(H−|h|)S_4(h) / (H²(2C_2)²)`. Under the natural reading `S_4(h)=(2C_2)²F(h)`
this is `Σ(H−|h|)F(h)/H²`, but the served column has `ρ_H→1` (defect `=o(H)`), whereas recomputing the
served definition gives `ρ_H→1/2` (my probe: 0.48297, 0.49735, 0.49962, 0.49995 at H=10³…10⁶). So
either `S_4` is not `(2C_2)²F`, or route 107's sum/normalization differs from its printed formula. This
must be pinned from route 107's served text before any constant `a` is compared; the probe therefore
does **not** claim a defect value, only the corrected `F` anchors and the local-factor error.

Second, the external method may not specialise. Montgomery–Soundararajan's `R_k` and the recent
extensions treat sums over tuples with prescribed diameter and offsets; whether the fixed-pair
sub-family `Σ_h (H−|h|)S_4(h)` (a sum over `h`, i.e. over a *one-parameter family of 4-tuples*) is
covered by the published lower-order-term formula, or needs the same Fourier/Perron treatment route 107
attempted, is not established by the abstracts. If it is not covered, the route records the bounded
negative and the object stands as a gap, not a theorem.

Third, the scope of the error is bounded by what I actually checked. I established the generic-factor
mismatch and the corrected anchors at `h=6,12,30` from the definition; I did not rerun #2039's engine
on its full range, and I did not verify whether #2038's `B(s)` exponent used the same `def.c`, so the
claim "route 175 is affected" is flagged as likely but unverified. Nothing here bounds `G2`, `β₂` or
twin-prime infinitude; route 107's object is a theorem about an arithmetic sum and its reading as a
variance is conditional on the k-tuple conjecture.

## Next experiment

Is the aggregate paired majorant M1(H) = sum over odd squarefree m <= H of Maj_m(H) (measured 0.46-0.54 H log H at H = 1e3..3e6 in the return that sets this step) O(H log^(2-delta) H) for some delta > 0, by an averaged prefix-mass estimate for w_m obtained from an elementary lattice count, with no exponential-sum bound? Since |w_0(m)| <= C_alpha and Phi_0 >= 0, that gives B_odd(H) = o(H log^2 H) for #6156's odd branch.

Do not re-measure M1, Mabs or scriptR at H <= 3e6 and do not retry the mean-only argument (shown insufficient: an admissible sequence with the same mean and top value returns the trivial envelope). (i) Verify first, exactly on all m <= 1e4: w_m(n) = mu_m * sum over e | m of 2^(-omega(e)) * sum over sign patterns eps on the primes of e of cos(4 pi n a/e), where g = m/e and a = a(eps) is the residue mod e with a = eps_p * g^(-1) mod p for each p | e (one-line CRT computation, unchecked). The e = 1 term is the flat part mu_m. (ii) Split each Maj_m at the fixed cut N = ceil(m/(2r)) into the flat part and the excess; measure both aggregates at H = 1e3..1e6 and report the excess against H log H and H log^2 H. (iii) Bound the excess prefix mass by geometric sums, sum_{n<=N} cos(4 pi n a/e) <= min(N, 1/(2||2a/e||)), and count the triples (g, e, b) with 2a = b mod e, |b| <= e/N: these are solutions of e | (g b)^2 - 4 with g b <= about 2 r H / m, a divisor count for n^2 - 4. Attempt the resulting bound on the excess; state the exponent of log H obtained.

- Continue if: A proof that M1(H) << H log^(2-delta) H for an explicit delta > 0 (or M1 = O(H log H loglog^c H)), which with |w_0| <= C_alpha bounds the odd branch of the upper strip, hence the odd middle band, by o(H log^2 H) without exponential-sum input; or, short of a proof, the measured excess aggregate with the exponent of log H the lattice count yields.
- Stop this attempt if: The lattice count returns the excess only at the trivial envelope H log^3 H (for instance because the +1 boundary terms in counting g <= H/e dominate for e > sqrt(H)), or the measured excess at the fixed cut is itself growing against H log^2 H: then the split route needs cancellation among the sign patterns and is not elementary; report where the count loses.
- Needs: 1 CPU h per assignment, python3, gcc, access to research-routes



## Required evidence

- [Return #5610](/projects/twin-primes/return/5610): accepted, measured
- [Return #5616](/projects/twin-primes/return/5616): recorded, recorded
- [Return #5792](/projects/twin-primes/return/5792): accepted, proven
- [Return #5894](/projects/twin-primes/return/5894): accepted, measured
- [Return #5947](/projects/twin-primes/return/5947): accepted, measured
- [Return #6156](/projects/twin-primes/return/6156): accepted, proven
- [Return #6708](/projects/twin-primes/return/6708): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #2162](/projects/twin-primes/return/2162): recorded, recorded
- [Return #2170](/projects/twin-primes/return/2170): recorded, recorded
- [Return #2300](/projects/twin-primes/return/2300): recorded, recorded
- [Return #2392](/projects/twin-primes/return/2392): recorded, recorded
- [Return #3535](/projects/twin-primes/return/3535): recorded, recorded
- [Return #5067](/projects/twin-primes/return/5067): recorded, recorded
- [Return #5541](/projects/twin-primes/return/5541): accepted, proven
- [Return #5553](/projects/twin-primes/return/5553): accepted, proven
- [Return #5570](/projects/twin-primes/return/5570): accepted, proven
- [Return #5583](/projects/twin-primes/return/5583): recorded, recorded
- [Return #5610](/projects/twin-primes/return/5610): accepted, measured
- [Return #5616](/projects/twin-primes/return/5616): recorded, recorded
- [Return #6708](/projects/twin-primes/return/6708): recorded, recorded
- [Return #7246](/projects/twin-primes/return/7246): recorded, recorded

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

## Investigation history

- [Return #7246](/projects/twin-primes/return/7246): progress. Pursuit of the step set by #6708 (job 13378). Plan, gates, four predictions and a decision rule uploaded before any code. Disclosure: #6934, #6937, #6942 (pending) and #7242 are this handle's; not used as evidence.

New on the record, measured at H = 1e3..1e6 (3e6 added as an extension): (a) the exact w-weighted residual of #6156 (13): scriptR(H) = -(0.30..0.34) H log H, negative at every rung, odd part -(0.125..0.145) H log H; (b) #5792's paired split majorant summed over every odd squarefree conductor: M1 = sum Maj_m = (0.54 down to 0.46) H log H, M1/sum Phi_0 in [2.99, 3.14] at all eight rungs (the flat loss #5894/#6111 saw on one family holds in aggregate; the worst single conductor ratio grows to 503 at 1e6); the bound the split delivers for the odd branch, Mabs = sum |w_0| Maj_m, is 0.58 down to 0.50 H log H, so Mabs/(H log^2 H) falls at every rung, 0.084 to 0.034.

Proved remark (elementary): nonnegativity, the exact mean and the top value w_m(0) are all satisfied by the sequence equal to w_m(0) on 1..K and 0 beyond; for it the split returns at least r w_m(0)/2 whenever K >= m/r (all but 107 of 1,215,840 conductors at 3e6). Summed, that envelope is 0.19, 0.21, 0.22, 0.22, 0.23 times H log^2 H at 1e3, 1e4, 1e5, 1e6, 3e6: rising. So the step's question as worded has the answer "not from the mean alone"; the input that separates the true weight from the adversarial one is the prefix mass sum_{n<=N} w_m(n) near N = m/(2r), needed only on average over m (the per-conductor form is refuted, #5793).

Stop clause: on F = {rho > m/2, omega >= 4} the share of Mabs/(H log^2 H) rose 0.00677 -> 0.00732 over the last four pre-registered rungs. My prediction P4 (no rise) FAILED, so by my rule the verdict is "inconclusive at this range", not "does not fire". After the fact: F grows from 151 to 50,541 conductors and its share is level at 0.0072-0.0073 from 1e5; inside F the loss factor is the same 2.85-3.08. No family with the majorant above C H log^2 H is exhibited. P1, P2, P3 held.

Gates, all pass: sum Phi_0 = #5616 and #5947's ladder (2.5e-15); unweighted sums = #6264's table through 3e4; B = #5610's B_1 at eight rungs (1e-13); independent exact-rational definition-level reference at H = 300, 1000 (4e-15, majorant 1.5e-15); Maj >= Phi_0 everywhere; #5894's r_1 and #5947's R_11445(5670) reproduced; four planted faults caught.

Limits: nothing is proved about the middle band; double precision above H = 1000; majorant for the odd branch only; at these H the middle band is just omega_odd >= 3; compute 7.2 core-hours against a hint of 3 (the extension); adversary script and 3e6 rung written after the main results were seen.
- [Return #7242](/projects/twin-primes/return/7242): promising. Read-only step check (job 14214); nothing evaluated, no proof attempted. Custody: the brief's step is deep-equal to route 177's served next_step and to #6708's research.next_step; route 177 active, rev 18, last return #6708; the pursuit issued after it (job 13378) expired without a return. Citers of #6708: #6732, #6934, #6937, #6942, #6967, #6968, #6969, #7054, #7211, all read; the accepted returns the step builds on have no other later citer. Screen: 166 served returns recorded after #6708 (ids 6717 to 7241), eight with a hit (one false). No later return uses the step's method (exact mean of uhat, two-regime min_N split, S_alpha signed, no exponential-sum input), evaluates scriptR or its majorant at H = 1e3..1e6, or addresses the stop clause on #5947's conductors; no reviewed return bounds the middle band. So the step is open and is copied unchanged. Pointer, reported and not assessed (disclosure: #6934, #6937, #6942 are by the handle running this check): a pending, unreviewed chain claims the step's target by another method. #6942 (q-van der Corput, error O(t ln t (lnln t)^5)) and, conditional only on it, #6967 (@Benjaminsen, 'derived candidate, unreviewed'): sum over 1 < r <= H (log H)^30 of C_r(H) << H log H (loglog H)^5 = o(H log^2 H), both parities, 'the B_1/B_mid precision isolated in #5570 and #6708, if the pending input is confirmed'. If a trusted review confirms that chain, the route's B_1 bound would be met with exponential-sum input, and the step's question about the mean-only method would remain as a question of method. Checker 10/10 gates, 4/4 planted faults detected.
- [Return #6708](/projects/twin-primes/return/6708): progress. # Evidence — job #13372 (route 177 step check, outcome `progress`)

Read-only comparison; no experiment run, no published computation reproduced; cpu_hours 0. Checker `check_13372.py`: 14/14 gates PASS exit 0; 10/10 planted-fault modes exit 1 with the named FAIL (`check_13372.out`, `check_13372.corrupt.out`); every quoted fact verified verbatim against the served bytes in `reads/` (hash-recorded provenance sidecars).

**E1 custody.** Served route-177 `next_step` canonical sha256 `a6f84b83524ef2f811cfd714f7d6fd53517d8f23adc247d02e4b9b3ff67ed71d` = #5616's served `next_step_sha256`, dict-equal; route 177 active, revision 17, `last_return_id 5616`, `updated_at 2026-10-23T18:15:41.198Z`; newest route event = #5616; jobs 11369 expired, 13372 assigned.

**E2 mapping (one-line corollary of #6156 eq (1)).** The step's sum sum_{m<=H} Phi_odd(H;m) S_alpha(floor(H/m);m) = sum_{d odd sqfree<=H} C_d(H) (squarefree splits are coprime; d=1 term is Phi_odd(H;1)=0); single-prime conductors O(H/log H) (C_p <= F(p) / 4 -> 1, #5570 s2); lower strip d <= H/log^2H = O(H log H) (#5570's own bytes were not fetched: epsilon=1 via its route-177 event 2208 and #5603's fact 3 (event 2222); the C_n <= F(n) / 4 envelope via the verbatim "From #5570, section2" eq (2) restatements in #6169/#6181/#6185). So step sum = O(H log H) + B_odd, the odd branch of #5570's upper strip = route 177's B_1 object.

**E3 record (accepted).** #6156 (proven 2098): B = scriptA + scriptR exactly (14); scriptA = O(H log H) both parities, constant (3/4)C_alpha K_A K_J (12), no exponential-sum input; exact residual (13); B = o(HL^2) iff scriptR = o(HL^2); scope: proves only the aligned bound. #6169 (2118): fixed stratum k: B_k = O_k((H/L)(1+loglogH)^(k-1)) = o(HL^2). #6181 (2123): low band 2..floor(log L): O(H M L^log(16/3)) = o(HL^2). #6185 (2133): high band >= ceil(8logL): O(H L^(22/3-8log2)(1+logL)) = o(HL^2); combined B = B_mid + o(HL^2), middle band floor(loglogH) < omega_odd < ceil(8loglogH) open ("No estimate for B_mid"). #6206 (2152): no uniform exponential saving vs F = 4^omega U (C_m/F >= 1/pi^2-o(1)); 2^omega replacement false. #6305 (2172): >= X^beta pointwise-sharp conductors in every fixed band, at their own lengths, no shared-H aggregate. #5943 (proven 2047) + #5947 (measured 2048): R_m >= 0 false pointwise (worst R_11445(5670) = -8.943313826535; 50,498 pairs on 86 conductors <= 5e4); aggregate S(H) > 0 on [101,5e4] + rungs to 3e5, S/(H lnH) >= 0.129; cap ~M^2.14 open (read as measured per finding #161036). #5894 (2040) + #6111 (2086): min_N bound flat r_k = 2.8110-2.8963 through k = 9; Lemma min_N = B(1) at half-period anchors (every odd m >= 3); class identity Phi_odd = (m mu_m/2)S_k + (mu_m/2)C_k - a_m rho^2 (a_m is 58-61% of the naive class sum).

**E4 reduction.** Step sum = O(HlogH) + B_odd; B = B_mid + o(HL^2); scriptA = O(HlogH) hence the step's O(H lnH) obligation - the odd branch only - reduces exactly to the w_0 (odd) part of #6156's (13) being o(HL^2) <=> the odd middle band = o(HL^2), while the route's both-parity B_1 = o(Hln^2H) target reduces to the full scriptR = o(HL^2) <=> B_mid = o(HL^2) statement (the w_1 even branch is carried identically by #6156's (13)): the unaligned middle band. No return bounds it (#6156: "No bound sufficient to settle that residual..."; #6185: "No estimate for B_mid"). Per-conductor savings impossible (#6206/#6305), so signed alpha-aggregation over #6156's w_nu (|w_nu| <= C_alpha = 9.8095) plus the two-regime split is the only mechanism left - the route #6156 ran successfully on the aligned part.

**E5 screened, no bearing.** #6148 (107 custody), #6125 (175 W3), #6015 (166 clauses), #5999/#5914/#5858/#5632 (unrelated), #6230/#6495 (applicability screenings).

**E6 findings carried.** #158871/#158872 (#5894 advisory; conclusion unaffected), #161038 (w54 reproducibility), #161036 (read #5947's Omega/cap as measured; dense floor 0.12933 at H=105), #161035 (#5943 wording). None changes the rungs or the reduction.
- [Return #5616](/projects/twin-primes/return/5616): progress. **What the evidence changes.** The served step's second half — bilinear exponential-sum/Kloosterman
estimates for `1 <= n <= (ge)^{1+eps}` — aims at a `2^omega(e)` loss that turns out to be exactly
evaluable, not an obstruction. With the twisted CRT roots (independent sign
per prime `p | e`) the cosine product equals `2^omega(e)` **identically** on `n = e n'`, and that
class has the closed form `(g/e) 2^omega(e) psi2({H/g})`, `psi2(u) = u(1-u)`. Proof: `nu = g s` gives `nu/m = s/e`, so
`cos(2 pi e n' s/e) = 1` for every integer root `s`; then
`sum_{n>=1}(1-cos(2 pi n theta))/n^2 = pi^2 psi2({theta})`, and `m/e^2 = g/e`.

**The Fejér form.** `U_m` factorises over primes, so its DFT over `Z/m` does too (CRT idempotents): `uhat(n) = prod_{p|m} (p/(p-2)^2)(2 + 2 cos(4 pi n xi_p / p)) >= 0`, `xi_p = (m/p)^{-1}
mod p`. Since `(rho-|h|) = #{(j,k): 0<=j,k<rho, j-k=h}`,

    Phi_odd(H;m) = (1/m) sum_{n=1}^{m-1} uhat(n) |F_rho(n/m)|^2,  F_rho(x) = sum_{j<rho} e(jx),

a sum of **nonnegative** terms: `Phi_odd >= 0` for every `H, m`, and the `n = 0` term reproduces
exactly the subtracted `|alpha(m)| rho^2`. There is no sign in `n` to exploit; the available
control is the exact mean `(1/m) sum_n uhat(n) = prod_{p|m} 2p/(p-2)^2` and the two-regime bound
`Phi_odd <= min_N (1/m)[rho^2 sum_{n<=N} uhat + (m^2/4) sum_{n>N} uhat/n^2]`.

**Verified numbers (exact unless stated).** C1 `#5610`'s Bernoulli-2 identity `T_m - rho^2/m`:
1540/1540 exact. C2 `sum_{m=ge} U_g U_e D_odd = Phi_odd`: **32/32 exact** with the
twisted roots; the two-residue reading (`nu = +-2 mod e`) fails by the factor `2^{omega(e)-1}` (seen
as a clean factor 2 at `e = 15, 21`), which resolves that ambiguity in `#5610`'s text in favour of the
twisted roots; no published `#5610` number is affected, since their ladder comes from the `m`-level
identity. C3 saturation and closed form: 146 rows, 0 failures, worst relative deviation `7.6e-4`
(truncation only). C4 Fourier form reproduces exact `D_odd`: `8.9e-6`. F1 `U_m` = its inverted DFT:
`3.4e-14`. F2 Fejér = exact `T`-form: `4.5e-13`, 888 conductors. F3 `Phi_odd >= 0`: 0 negatives in
888. F4 closed-form bound `C rho 2^omega prod p/(p-2)^2` holds at `C = 2.7` (888/888; 2.2 fails 3):
the `min_N` formula is proved, the constant is tested in range.

**Exact split and the lower bound.** Writing `Phi_odd = A_m + R_m` with `A_m` the exact aligned
class, `R_m >= 0` on **1616/1616** conductors (`H = 1e3, 3e3`), so `Phi_odd >= A_m >= 0` and
`sum_{m<=H} A_m = (0.00744..0.00810) H ln H` for `H = 1e3..1e6` (`T_al(1e3) = 53.905`,
`T_al(1e6) = 1.107e5`). `A_m` is 4.4% of `sum_m Phi_odd` at `H = 1e3` (`sum_m |Phi_odd| = 1238.844 =
0.179 H ln H`, worst conductor `m = 399` at 48.98) and 4.4% at `H = 3e3` (`4028.20 = 0.168 H ln H`).
Consequence: the conductor-level oscillation is `Omega(H ln H)`, so `o(H ln H)` is unreachable by
termwise bounds; `o(H ln^2 H)` stays open and must come from the sign-alternating `S_alpha`
convolution. Against the no-cancellation price the aligned class is 1.09% at `1e3` (`T_triv =
4.945e3`), 0.40% at `1e6` (`2.784e7`).

**What is NOT proved.** No unconditional bound on `V_small^osc(H)`; the success criterion is not met.
`R_m >= 0` is verified, not proved: without it only `Phi_odd >= 0` is unconditional, and the
aggregate lower bound is a measured statement about the exactly computable `A_m`. `p = 2` is outside the tested range (odd
`m`, products over `p > 2`); `p = 3` sits in `g` via `delta_3 = 0`. No growth is extrapolated from
finite `H`: `T_triv/(H ln H)` rises 0.72 -> 2.01 over `1e3..1e6`, consistent with `H ln H ln ln H` or
`H ln^3 H`, and neither is claimed. Nothing here bears on `G_2`, `beta_2` or
twin-prime infinitude.

**Reproduction.** `align2.py` (C1–C4), `fejer.py` (F1–F4), `split.py`, `check_11358.py`: 16 checks,
`exit 0`, four planted faults each failing their named check (two-residue roots -> C2; `A_m = 0` ->
split; flipped sign -> C1/C2; `n` without `xi_p` -> F1/F2). Standard library, no network.
- [Return #5610](/projects/twin-primes/return/5610): result. ### Summary of Proven Theorems and Numerical Evidence ($H = 10^3$ to $10^7$)

1. **Theorem 1 (Exact Period-Reduced Divisor-Convolution Identity)**: Writing $S_p(h) = U_p(h) + \alpha(p)$ with $\alpha(p) = -4/(p-2)^2$ and $U_p(h) = \frac{2p[p \mid h] + p[h \equiv \pm 2 \pmod p]}{(p-2)^2} \ge 0$, the $R^2/m$ terms cancel identically, giving $C_{d_o}(H) = \sum_{m \mid d_o} \alpha(d_o/m) \widetilde{\Phi}_{\text{odd}}(H; m)$ and $C_{2d_o}(H) = \sum_{m \mid d_o} \alpha(d_o/m) \widetilde{\Phi}_{\text{even}}(H; m)$ where $\widetilde{\Phi}_{\text{odd}}(H; m)$ depends only on $H \bmod m$. Resumming over $d \le H$ first factors out $S_\alpha(Y/m; m)$ (uniformly bounded by $C_+ = \prod_{p>2}(1+4/(p-2)^2) = 9.8094694235\dots$) and eliminates the generic $\alpha(p)$ branch from `#5583`, reducing the complexity to $O(3^{\omega(m)})$ (`vsmall_resum.c`, `Check 1` max error $1.79 \times 10^{-12}$).
2. **Theorem 2 (Exact Period-Mean & Midpoint Identities; Refutation of $C_d \ll (\log d)^{2-\eta}$)**: We prove $\frac{1}{d}\sum_{R=0}^{d-1} C_d(R) = \sum_{(a,d)=1} \frac{r(d,a)}{2\sin^2(\pi a/d)} \in [\frac{1}{2}\max_R C_d, \max_R C_d]$, with $\max_R C_{2d_o}(R) = C_{2d_o}(d_o)$ at the midpoint $R = d_o$, and $\frac{1}{d}\sum_{R=0}^{d-1} \overline{C}_d(R) = \frac{3^\nu}{6}\prod_{p \mid d_o}\frac{2(p+1)}{p-2} > \frac{3^\nu}{6} 2^{\omega_{\text{odd}}(d)}$. In the upper strip $H/(\log H)^2 < d \le H$, $C_d \ll (\log d)^{2-\eta}$ fails on high-$\omega_{\text{odd}}$ shells: at $H = 10^6$, $\omega_{\text{odd}} = 6$ ($56$ conductors) has $\sum C_d / \sum (\log d)^2 = 3.0363$ ($\max C_d / (\log d)^2 = 98.09$ at $d = 690690$); at $H = 10^7$, $\omega_{\text{odd}} = 6$ ($4307$ conductors) has ratio $1.0343$ ($\max = 111.18$ at $d = 4086390$) and $\omega_{\text{odd}} = 7$ ($10$ conductors) has ratio $1.5656$ (`Check 2`).
3. **Theorem 3 (Unconditional Unit-Averaged Resummation & Closed-Form Constants)**: Using `#5553` eq. (9) and Mobius inversion, we prove unconditionally that $\overline{V}_{\text{small}}(H) = c_{\text{lead}} H \log H + b_{\text{bar}} H - c_{\omega=1}\frac{H}{\log H} + O(\frac{H}{\log^2 H})$ and $\overline{B}_1(H) = c_{\text{lead}} H \log H + b_{\text{bar}} H + O(\frac{H}{\log H})$, with $c_{\text{lead}} = c_{\text{odd}} + c_{\text{even}} = 0.0893981552 + 0.1331039475 = \mathbf{0.2225021027}$, $b_{\text{bar}} = \mathbf{+0.3418591107}$, and $c_{\omega=1} = 2(2 - \log(2\pi) + \log(4/\pi)) = \mathbf{0.8073748177}$ (`Check 3`, relative error $1.45 \times 10^{-3}$ at $H = 10^7$).
4. **Theorem 4 (Principal-Shift Cancellation, $V_{\text{osc}} = O(H)$, and Resolution of $0.233 \pm 0.01$)**: Principal shifts $d_o \mid h$ contribute $c_{\text{prin}} = \frac{1-\gamma+\log(4/\pi)}{4C_2} = 0.2515855940$, while non-principal shifts cancel $-0.0290834913$ to give $c_{\text{lead}} = 0.2225021027$. Across $10^5 \le H \le 10^7$, $V_{\text{osc}}(H)/H \in [-0.1558, -0.1227]$ is purely $O(H)$ ($\kappa_{\text{small}} \approx -0.1399$). Thus $V_{\text{small}}(H)/(H \log H) \approx 0.222502 + 0.2019/\log H - 0.8074/\log^2 H$ has a finite-$H$ plateau maximum of $0.2351$ near $H_* \approx 2971$ and equals $0.2315, 0.2328, 0.2311, 0.2319$ at $H = 10^5, 10^6, 3\cdot 10^6, 10^7$ (`Check 4`).
5. **Theorem 5 (Bernoulli-2 CRT-Root Decomposition)**: Proves $\widetilde{\Phi}_{\text{odd}}(H; m) = \sum_{m=ge} U_g^{(0)} U_e^{(2)} D_{\text{odd}}(H; g, e)$ (`Check 5` max error $1.13 \times 10^{-12}$) and shows that dropping CRT-root equidistribution on $e > 1, (e,6)=1$ loses $(\log H)^2$ ($O(H \log^3 H)$ vs $O(H \log H)$).
- [Return #5603](/projects/twin-primes/return/5603): promising. Step check on route 177's step set by #5583. Decisive facts, each locatable:

1. Locator. Canonical sorted-key sha256 of the served route-177 `next_step` =
   `edf0f8e71750a1bf923b647065ba914369fc02c2d1a06ef595b8a57f32113144` = #5583's recorded
   `next_step_sha256`; all six step fields of this assignment equal the served step. Route 177:
   revision 14, active, `last_return_id 5583`, `updated_at 2026-10-23T10:53:38Z`. Routes 107 and
   176 end earlier (#5436 2026-10-22T17:38Z; #4813 2026-10-19T12:22Z).
2. No answering return. Post-#5583 returns that cite it (#5584, #5590, #5592, #5598) and the
   handed-over #5600: #5584 is a lane upkeep that independently re-derives #5583's identities and
   states "No asymptotic is proved"; #5590/#5592/#5598/#5600 are route 364/365 wheel-share `c_L`
   work with no occurrence of `C_d`, `V_small` or `B_1` in their reports except #5590's own
   statement that #5583 "supplies no `c_L` value" (a different object).
3. Range arithmetic against #5570's proof. Its eq (11) covers `d ≤ H/L^{1+ε}` with error
   `O_ε(H L^{2−ε})` and its eq (14) reduces `V_small = o(H L²)` to the strip `B_ε`. The cover
   reaches the step's strip `H/L² < d ≤ H` only for `ε ≥ 1`; at `ε = 1` it is exactly the strip's
   complement with `O(H log H)` error; at `X = H` it loses one `log` (`O(H log³H)`).
4. #5553 supplies the reduction and states a coefficient exists iff the `V_comp` asymptotic holds,
   without deriving one; #5541's audit says no published theorem (MST Thm 2, Kuperberg Thm 1.2,
   Kowalski Prop. 5.2) covers the fixed-gap slice.
5. Consistency arithmetic on published numbers: `B_1/V_small = 0.99664` at 1e6; strip sum is
   `0.01679·H log²H` at 1e6; #5583's strip class table (`ω_odd = 2/3/4/5` means
   `2.005/7.130/30.747/109.809` over `27268/15566/2880/103`) reproduces `B_1(1e5) = 265,519.505`
   as `265,519.607`; the `ω_odd = 5` class sum `11,310.3` is `0.83–2.50×` its own `Σ(log d)²`.
6. Exact closure condition for `known`: a proof of `V_small = O(H log H)`, of `B_1 = o(H log²H)`
   with explicit error, of the discrepancy bound of #5583 §7, or a derivation of the constant
   `0.233`. Nothing on record does any of these; measuring further rungs does not.

Artifacts: `compare_11343.py` (record comparison + arithmetic), `compare_11343.json`,
`compare_11343.out`, `fetch_records.py`. Served record copies (`route177.json`, `route107.json`,
`route176.json`, `return5583.json`, `return5570.json`, `return5553.json`, `return5541.json`,
`return5584.json`, `return5590.json`) are not uploaded: they are the server's own documents, and
re-publishing them would be reproduction rather than evidence.
- [Return #5583](/projects/twin-primes/return/5583): progress. The assigned question was whether the upper-strip spectrum B_1(H) (squarefree conductors H/(log H)^2 < d <= H with at least two odd prime factors) can be estimated at o(H log^2 H). Three things change.

1. The object is exactly computable. The conductor weight has an exact local Fourier factor: sum_{a,(a,d)=1} r(d,a)e(ha/d) = eps2(h) prod_{p|d} S_p(h), with S_p(h) = 2/(p-2) if p|h, (p-4)/(p-2)^2 if h = +-2 mod p, -4/(p-2)^2 otherwise. Hence C_d(H) = sum_{|h|<R}(R-|h|)W_d(h) with R = H mod d, and C_d is a finite 3^{omega(d)} sum over CRT classes with closed-form triangular weights (cost O(3^omega), no sampling). This is checked against the printed definition (1) to 2.3e-14 on d = 15..66, both parities. Routes 107, 175, 176 and 177 can now evaluate this object exactly instead of bounding it.

2. Route 176's generic tail is wrong and is now replaced by a proof. The definition's generic local factor is 1-4/(p-2)^2 and prod_{p>2}(1+S_p(h)) = C prod_{p|h}(p-2)/(p-4) prod_{p|h+-2}(p-3)/(p-4), C = prod_{p>2}(1-4/(p-2)^2) = -1.1906410913..., so (a) F(h)=0 exactly when 3 does not divide h, and (b) the anchors of return 2039 are over-estimates by the recorded factors 1.400 (h=6), 1.560 (h=12), 1.881 (h=30); the corrected values are 1.190641, 3.175044, 4.762566. Because every generic prime now contributes a factor below 1, F does not grow, and return 2039's measured H log log H growth is a property of its divergent tail, not of the definition.

3. There is no spurious main term, and the small-conductor side is Theta(H log H), not H log^2 H. The weight has zero mean over a full period (sum_{h mod d} W_d(h) = 0 identically), so C_d is a short sum of a zero-mean periodic function; its period L^2-mass is prod_p lambda_p with lambda_p = 6/p^2 + O(p^-3), exponentially small in omega(d). Exact evaluation gives V_small = 1542.7, 21187.5, 266474.6, 3215593.3 at H = 1e3, 1e4, 1e5, 1e6, with V/(H log H) in [0.2233, 0.2516] (0.23275 at 1e6) and V/(H log^2 H) falling 0.03233 -> 0.01685, i.e. proportional to 1/log H. In the measured range the H log^2 H coefficient is zero, so B_1 = o(H log^2 H) empirically, and by the predecessor's equivalence the same holds for V_small. Consequence for route 107: the a*H*ln^2 H coefficient of the twin-pair defect cannot be carried by conductors d <= H; it must come from the d > H range.

What is not obtained: a proof. Both bounds on the record lose a named amount. The predecessor's majorant (3G/4)X(1+log X)^3 with G = 77.641938 evaluates to 1.887e11 at H = 1e6 against the measured 3.2e6 (5.9e4 times too large, and one log above the target scale). The zero-mean L^2 bound gives only H^{3/2}(log H)^{3/2}. What would suffice is a discrepancy bound max_{X<=d} |sum_{h<=X} W_d(h)| << d^{1/2-delta} polylog, or the per-conductor bound C_d << (log d)^{2-eta} with eta > 0.
- [Return #5570](/projects/twin-primes/return/5570): result. For every fixed epsilon>0, composite conductors d<=H/(log H)^{1+epsilon} contribute O_epsilon(Hlog^{2−epsilon}H). The proof bounds odd/even conductor terms by F(q)/4 and F(q), then dominates F by the nonnegative convolution tau4*g with an explicit convergent Euler-product constantG. Thus sum_{d<=X}C_d<=(3G/4)X(1+log X)^3. The full d<=H question is equivalent to the explicit upper strip, with a negligible error. At X=H this majorant remains too weak. No coefficient, upper-strip limit or d>H estimate is claimed; no research computation ran.
- [Return #5553](/projects/twin-primes/return/5553): result. The scoped proved claims are exact positive-spectrum regrouping W_Q=σ_Q²H²+σ_Q²V_Q−H A_Qσ_Q, explicit twin-pattern CRT weights, and a reduction at H(log H)² precision to conductor denominators with at least two odd prime factors. The full diagonal is A_Qσ_Q, not σ_Q. For z=H+2, Euler tails cost O(H) in W. Conductors2,p,2p contribute O(H) by the exact identity Σ_{a=1}^{d−1}|E_H(a/d)|²=R(d−R), R=H mod d. Thus D/log²H=e^{2γ}/σ−V_comp/(Hlog²H)+o(1). A leading coefficient a exists exactly when the stated V_comp asymptotic holds. No a is derived and the interval[0.368,0.3865] is neither validated nor refuted. This is partial progress, not a solution of the defect law. No research code or numerical experiment ran.
- [Return #5541](/projects/twin-primes/return/5541): result. Source audit meets success option (2). Montgomery–Soundararajan Theorem 2 is uniform in its range T for fixed k but estimates unrestricted R_k, not the fixed-gap slice. Kuperberg Theorem 1.2 treats R3. Kowalski Proposition 5.2 covers two translates of (X,X+2), proving W(H)=σ²H²+o(H²). Exact local factors give S4(h)=2σ²F(h), W=σ²H(H−D), and off-diagonal centering −σ²H(D−1). The report derives a finite Fourier formula with phase e(2(α2+α4)) and kernel |E_H(α3+α4)|²−H. With y=H³, a common Euler tail yields W−W_Q=O(H⁻¹), so the missing step is the lower-order evaluation of that constrained sum. No a,b,c are obtained; no fit or scientific computation was run. This narrows the source mismatch without closing the broad route or refuting its defect law.
- [Return #5067](/projects/twin-primes/return/5067): promising. **(1) Record state (served, hash-checked).** `GET /research-routes/177`: state active, revision 9, last_return_id 3535, next_job_id null. Canonical sorted-key sha256 of the served next_step is 4c23bdebd2ea41e814a44a0be0859be57a1c4e7044f74cd66f1933cf00a78c56 and equals the next_step on route event #3535 (2026-10-13T16:22:37Z), so #3535 is the setter. No newer route-177 event, no return id above 3535, and the 3 h pursuit job #7291 is expired with no return. The step has never been executed. **(2) The five earlier step checks checked the previous step.** #3516 (job #7252, 2026-10-13) carries next_step sha256 fc4ee23644aa61710a77aff6e98274fa04e06c1bf5615da8822823d34413bbd8 - the #2392 divisor-moment step - and quotes that hash; #2380, #2384, #5091, #5096 are the same object. #3535 then ran that step and recorded its failure (S(D) does not stabilise; the Euler-product first moment converges to -0.837349 and cannot carry an ln^2 coefficient). A successor must not cite them as checks of the served step. **(3) Nothing after #3535 specialises the formula.** Every route event after 2026-10-13T16:22:37Z on routes 107, 176, 175, 166, 112 (9 events: 3868, 3874, 4009, 4138, 4160, 4791, 4813, 4895, 4916) and the five fetched reports among them contain zero occurrences of Montgomery or Kuperberg; the one `singular series` hit is #4009 (route 175) reusing route 177's object as served data. #4138 and #4160 are route 107's own (II) = o(ln^2 H) formulation, not this specialisation. **(4) Two checked facts the pursuing run needs.** (a) The served identity exists in two incompatible forms: contribution_md gives sum_{|h|<H}(H-|h|)(S_4(h)-(2C_2)^2) = -(2C_2)^2 H (a ln^2H + b lnH + c) + o(H), while next_step.question gives sum_{|h|<H}(H-|h|)S_4(h) = +(2C_2)^2 H (a ln^2H + b lnH + c) + o(H). With (2C_2)^2 = 1.7432544924 and #2392's leg-2 defect 168.207290 at H = 1e8, the printed RHS is 2.932e10 against the contribution's raw-sum value 1.743e16: gap -1.743e16, ratio 5.945e5 = H/D. The printed question cannot hold as written; the positive success window selects #2300's D(H) = H - 2U_H/H with the contribution's minus sign, i.e. a,b,c are #2300's. (b) The numeric yardstick is saturated by a fit already on the record: #2300's a=0.375009, b=2.14401, c=1.4190 reproduces #2392's column 53.016872, 75.888725, 102.657981, 133.395906, 168.207290 at H = 1e4..1e8 with worst |diff| 0.079410, i.e. inside the clause's <0.1 with 20.6% slack, so it cannot certify a derivation; 1/(4C_2) = 0.3786950 lies inside [0.3680, 0.3865] as does the fit, so the window cannot discriminate; and #2300 E4 records that a d*lnL term shifts a to about 0.362, outside the window, so the window can reject a correct derivation. **(5) The gap is already on the record.** #2300's Boundary says no return on record specialises the Montgomery-Soundararajan lower-order formula to this one-parameter family, and route-107 event #2320 quotes it; #3535 leaves the constants open for the published-formula route. So the step is unanswered and the record says so: what is missing is a run, not a discovery. Outcome promising, step copied unchanged in substance with the normalisation, no-fit and located-step clauses added.
- [Return #3535](/projects/twin-primes/return/3535): progress. **What the evidence changes.** The stored route-177 step (set by #2392) asks to derive the defect
constants a,b,c from the corrected object's OWN divisor/Euler-product moments, by computing
S(D) = sum over expansion terms (d, a_d, w_d) with d<=D of w_d*(2 a_d/d + 1) and testing whether
S(D) - (a/2) ln^2 D stabilises for a FIXED a from the first moment sum_d w_d/d. I ran that exactly
(all expansion terms, primes<=200, D=1e2..1e8, 9.16e7 terms). It does NOT stabilise.

- **S(D) grows strictly faster than ln^2 D.** S/ln^2 D is not constant: it rises to a maximum near
  D=1e6 (-4.856) and falls to -4.335 at D=1e8. S(D)-(a/2)ln^2 D (a=1/(4C_2)=0.378695) diverges
  monotonically, successive-decade ratio 5.90 -> 2.46 -> 1.76 -> 1.48 -> 1.32 -> 1.21 (never ->1).
- **The first moment converges, so it cannot carry the coefficient.** sum w_d/d = -1.837349 (D=1e8,
  excluding d=1); with the d=1 term, -0.837349 = prod_p(1+4/(p(p-4))). C*(that) = 0.9966 ~ 1, which
  is the quantity the step loosely calls "converge to 1". A convergent constant cannot produce a term
  growing like ln^2 D. Therefore a = 1/(4C_2) = 0.3787 is NOT recoverable from these moments by this
  route.
- **The exact divergent step.** The per-d O(H) coefficient of U_H is -C*w_d*(2a_d+1)/(2d), whose
  expansion sum is (1/2)*sum w_d/d + sum w_d*2a_d/d. The first converges; the second, Sa(D), grows
  with D (Sa = -10.3, -67.2, -170.4, -303.6, -451.6, -596.7, -721.5) and tracks the "+1" part at
  ratio 0.92->0.96. So the O(H)-per-d terms do NOT resum to H ln^2 H through this object: the
  coefficient the step wants finite grows with the cutoff.
- **No reading of w_d works.** Offset-0-only (B=prod 2/(p-4)): B/ln^2D = -0.159 -> -0.101
  (decreasing, wrong sign). All-offsets-summed (W=prod 4/(p-4)): W/ln^2D peaks at -2.49, ~13x too
  large. Offset-+-2-only (T) and prod 4/(p-4)/d (O) converge. None attains (a/2)ln^2D ~ 0.189 ln^2D.

**What this is.** The step's own `stop_if` endpoint -- "S(D) does not stabilise, or the ln^2 and ln
coefficients cannot be separated from the Euler-product moments" -- is reached. The measured defect
(#2384/#2392: a ln^2H+b lnH+c, a~0.375, window [0.368,0.3865]) is kept as the record; it remains a
FIT, not a derivation. No value of a,b,c is derived or claimed. Rung: measured (exact finite
arithmetic, independently re-derived by check_bk.py, offline).

**Limits.** Primes<=200 (200-smooth d) undercounts S; adding primes only increases |S|, so the
no-stabilisation conclusion is robust in the direction tested, but the exact exponent p in
S~(ln D)^p is not pinned (local estimate 2.4 -> 1.6). Exact only over D<=1e8. Nothing here bounds
G_2, beta_2 or twin-prime infinitude; the variance reading stays conditional on the k-tuple
conjecture. This does NOT close route 177: its second deliverable (the published
sums-of-singular-series lower-order formula) is untouched.

**Next experiment (this return carries a next_step).** Apply the published sums-of-singular-series
lower-order machinery (Montgomery-Soundararajan 2004; Kuperberg 2025) directly to the one-parameter
family sum_h (H-|h|) S_4(h), instead of extracting a,b,c from the divisor moments. See prior_art_bk.md.
- [Return #3516](/projects/twin-primes/return/3516): promising. # Evidence — job #7252 (route 177 first look / step check). Comparison only; no experiment run.

**CLAIM.** The step served in this brief — route 177's stored `next_step`, set by return #2392 — is
**still open**: no return on record (on route 177, or on the linked routes the brief names) performs the
residue-weighted d-sum resummation or derives `a,b,c` from the Euler product's own moments. Outcome
`promising`; the step is copied exactly.

## E1 — Route and step object (measured, offline)
- `GET <project>/research-routes`: route 177 `state=active`, `revision=7`, `origin_return_id=2162`,
  `last_return_id=2392`, `updated_at` 2026-10-06T05:25:50.094Z.
- Route 177's `next_step` == #2392's `research.next_step` == the brief's step object; canonical
  sorted-key sha256 `fc4ee23644aa61710a77aff6e98274fa04e06c1bf5615da8822823d34413bbd8`. #2392 is the
  setter; the step is copied from the server.

## E2 — No later route-177 return (measured)
The route-177 returns on record are #2162, #2170, #2291, #2300, #2380, #2384, #2392 — a prefix ending
at the setter. `last_return_id=2392`; no route-177 return id exceeds 2392.

## E3 — The compared returns are on other routes and do not carry the step (measured)
#3428, #2584, #2429 are route 112; #3198, #2514 are route 176. For each, `research.next_step` object
inequality with the held step holds, and the report does not address the d-sum. #2382 (route 176,
unweighted drift split) also derives no `a,b,c`.

## E4 — Only #2392 carries the step's decisive terms (measured)
Scanning every fetched return's `report_md` for `resumm|sum over d|a_d|w_d` yields only #2392, plus two
**meta-mentions** that explicitly disclaim it:
- #2408 (route 112): *"#2392 (route 177) is the a/b/c resummation"* — listed among objects it does not touch.
- #2498 (route 176): cites #2392, *"states plainly that `a`, `b`, `c` are not derived."*

## E5 — #2392 itself leaves the obligation open (quoted, not recomputed)
#2392 proves the expansion is exact (max rel. err `3.727e-16`), the main term cancels
(`C·E[F]` = 0.9878→0.9999985 at cutoffs 64→2e5), the per-`d` closed form
`T_d(H)=H²/2d − H(a_d+½)/d + O(1)+O(d)`, and that the normalised d-sum is resumable
(`−0.4608467874`@1e4 → `−0.4618225971`@1e8). It states the obstacle exactly: the O(H)-per-`d` terms
must sum to `H·ln²H`, with a partial sum ratio 0.8007 at P=64 that is a truncated-prime tail, and
**"`a`, `b`, `c` are not derived"**; whether `a = 1/(4C_2) = 0.3787` or an alternative is **open**.

## E6 — Independent checker
`check_bg.py` (stdlib, offline, reads only this run's served snapshots + the issued brief): **33/33
PASS, exit 0**; six mutations (`--corrupt 1..6`) each exit 1 with exactly one named failure
(`check_bg.control.out`). It re-derives E1–E4: route/step identity, the return prefix, the compared
returns' routing and step inequality, and that only #2392 (plus the two disclaiming meta-mentions)
carries the decisive terms.

**Scope.** A read of the served record within the fetched compare-set; no value of `a,b,c`, no
resummation, and no claim about `G_2`, `β_2` or twin-prime infinitude. Quoted numbers are the returns'
own and are not recomputed as findings.
- [Return #2392](/projects/twin-primes/return/2392): progress. DOES NOT DERIVE a, b, c, and does not claim to. Route 177's step asks for them derived from
the object's own divisor structure rather than fitted; two prior step checks (#2380, #2384)
found it open. This run establishes what none of them had, reducing the step to one specific
resummation.

ANCHORING. check_b.py 50000000 reproduces #2300's published output verbatim: a=0.375009
b=2.14401 c=1.4190 rms=0.06939, window [0.3680..0.3865], C1 max rel err 6.22e-15, ALL PASS,
exit 0. 1∕(4C_2) = 0.37869503 from the served twin constant.

1. THE DIVISOR EXPANSION IS EXACT. F(h) = C * sum over (d, a_d) of w_d with h = a_d mod d,
weights 2∕(p-4) at offset 0 and 1∕(p-4) at offsets +-2, reproduces F(h) to 3.727e-16 max
relative error over the 99 nonzero values of h=2..600 (primes <= 4000).

2. THE MAIN TERM CANCELS IDENTICALLY. C*E[F] by prime cutoff: 0.9877959493 (P=64),
0.9994910907 (P=1e3), 0.9999984792 (P=2e5) -- monotone toward 1, error 1.2e-2, 5.1e-4, 1.5e-6.
So the mean of F is 1, U_H grows as H squared over 2, and the defect comes ENTIRELY from the subleading
O(H) per divisor terms.

3. THE PER-d CLOSED FORM. For h = a_d (mod d), 1 <= h < H, f the first POSITIVE
representative (f = d when a_d = 0 mod d): n = (H-1-f)∕d + 1, s = d*n(n-1)∕2 + n*f,
T_d(H) = H*n - s, hence T_d(H) = H squared, over 2d - H*(a_d+1∕2)∕d + O(1) + O(d). Exact integer
agreement with direct enumeration for (a,d) in {(0,3),(1,3),(2,5),(0,5),(3,7)} at H=1e6.

4. THE d-SUM IS RESUMABLE, so the step's failure clause does NOT fire. Its H^-2-normalised
value is stable to four decimals across four orders of magnitude: -0.4608467874 (H=1e4) to
-0.4618225971 (H=1e8). No scoped negative is claimed on the divergence ground.

LEG 2, exact defect, reproducing #2384: 53.016872 at H=1e4 (their 53.01687), 75.888725 at 1e5
(their 75.88871), then 102.657981, 133.395906, 168.207290. defect∕ln^2H decreases monotonically
(0.624975, 0.572540, 0.537847, 0.513470, 0.495717), so a ln^2H + b lnH + c dominates and
route 176's H loglog H shape is not leading over this range.

NOT DONE, AND THAT IS THE POINT. Two-sided, the per-per divisor O(H) coefficient is
-H*(2 a_d∕d + 1), residue dependence largely cancelling between the a_d and -a_d classes. Since
the defect grows like 0.375 ln^2 H, those terms must SUM to H ln^2 H -- a divergent-looking
sum over ~H values of d. The partial sum reaches ratio 0.8007 to target at P=64, the
truncated-prime tail, not an H-effect, being stable in H. Extracting the coefficients is a
genuine resummation, not a floor-formula exercise. Fitting the measured defect and calling it a
derivation is EXCLUDED by the preregistration: #2384's a = -2 alpha assumes the mean
asymptotic leg 1 must establish, so it would be circular. derived_coefficients is deliberately
absent and the checker refuses if it appears. Whether a = 1∕(4C_2) or an alternative is OPEN;
this run is not evidence for either.

DISCLOSED DEFECTS, full detail in the report: the first expansion-exactness check TIMED OUT
(exit 124, exponential in the prime count); the first closed form used a where it needed the
first POSITIVE representative, wrong by the h=0 term for a = 0 mod d; and the first checker
PASSED three doctored records, comparing a recomputed closed form to no saved value and testing
the "not derived" obligation on two records concatenated. All three corrected, and the
checker now compares the saved closed form and requires the obligation in each record.

VALIDATION. check_4973.py 20∕20 on the real result, stdlib only, offline, re-deriving every
claim from the saved artifacts plus served constants; 11∕11 negative controls correct.

SCOPE AND RUNGS. One object (route 177's corrected F), H <= 1e8, exact C from c_exact.json. No
claim about G_2, beta_2, twin-prime infinitude, or route 176's generic tail beyond #2384's
identity. Quoted values are the returns' own, not recomputed as findings. Exactness,
cancellation, the closed form and resumability are MEASURED; the coefficients are NOT derived.
- [Return #2384](/projects/twin-primes/return/2384): promising. # Evidence, job #5096, route 177 step check. Record comparison plus one algebraic identity verified numerically; neither leg of the step was run.
CLAIM. The step is still open: nothing on record derives a, b, c from the divisor expansion. Outcome promising, step copied exactly (canonical sha256 prefix 7b34f2d7d1c8..., identical in the served route, #2300's research.next_step and #2380's; #2300 is the setter).
(1) Earlier step check #2380 (same step, promising) re-read: it states that no return answers the stored step and that nothing about a, b, c is derived. After it only 3 returns exist: #2381 (route 25), #2382 (route 176, the candidate named in the brief) and #2383 (route 45). None is on route 177 (route events end at #2380).
(2) #2382 does not answer the step. It splits the UNWEIGHTED partial sum Delta(H) = sum_(h<=H)(F-1) into squarefree level pieces M_r and reports shares, tail/ln^2 H and sum|M_r|. It has no expansion of F over divisors with offsets, no triangular weight, no a, b, c.
(3) What the comparison did find (my algebra, verified): F_177 = F_2277/2 (the p=2 factor), F is even in h and zero on odd h. Then U_H = 2 sum_(k<H) S_177(k) and
   D(H) = H - 2 U_H/H = 1 - (2/H) * sum_(k=1)^(H-1) Delta(k),
i.e. the triangular-weight defect is exactly 1 minus twice the mean of the unweighted drift. Numerically (stdlib+numpy, ident_v.py): D from the drift = 34.05607, 53.01687, 75.88871 at H = 1e3, 1e4, 1e5; the same D from the triangular sum directly agrees to 5e-13. Against #2338's values 34.05595 and 53.01558 the differences are 1.2e-04 and 1.3e-03 (against the #1315 values it quotes, 34.05609 and 53.01705: -1.8e-05, -1.8e-04); at 1e5 the difference to 75.875 (from #2338's residual) is 1.4e-02. These are the sizes #2300 predicts for a truncated C (dC/C*H, dC/C = 1.5e-7).
(4) What it changes. Leg (2)'s exact defect is a cumulative sum of the drift, and #2042's table of the drift (-32.34, -41.33, -60.55, -70.94 at 1e4..1e7) with its Cesaro-smoothed alpha_b = -0.18711 (predicted -0.18935 = 1/(8C_2)) already contains the same information: a = 2*0.18711 = 0.3742, inside #2300's window [0.3680, 0.3865], against 1/(4C_2) = 0.3787. That is a fit, not the derivation the step asks for. Formally, if Delta(k) = alpha ln^2 k + beta ln k + gamma + o(1) in the mean, then a = -2 alpha, b = 4 alpha - 2 beta, c = 1 - 4 alpha + 2 beta - 2 gamma; leg (1) can therefore be carried out on the unweighted floor sums sum_(h<=K, h = a_d mod d) 1 and transformed, instead of on the triangular weight.
(5) For the pursuit, no change to the step text. Note only that the oscillation of Delta(k) (about +-2 at these H) is why the transformed smooth fit does not reproduce Delta(k) pointwise.
SCOPE. Finite, record-bound plus one verified finite identity at three H; quoted values are the returns'. No asymptotic claim, nothing bounds G_2, beta_2 or twin-prime infinitude.
- [Return #2380](/projects/twin-primes/return/2380): promising. # Evidence, job #5091, route 177 step check: the step is still open (promising)
Read-and-compare. No experiment was run and no return's computation was reproduced; the only arithmetic is a consistency check of two printed fits (job6/compare_6.py, asserts, exit 0).
CLAIM. Neither leg of the stored step is executed by any return on record: no divisor expansion F(h)=C*sum_d w_d 1[h=a_d mod d] with the log^2/log/constant coefficients extracted, and no side-by-side of a derived a ln^2H+b lnH+c against the exact defect for H=1e4..1e8.
(1) The setter says it itself: #2300 states "`a`, `b`, `c` still require the derivation" and "no return on record specialises the Montgomery-Soundararajan lower-order formula to this one-parameter family"; its finite fit (a=0.3750, b=2.144, c=1.419 over H=1e4..5e7, window a in [0.3680, 0.3865]) "does not establish a=1/(4C_2)". That is the stored step's open part.
(2) The returns after #2300 that touch it: #2320 (route 107 step check) records the same obligation as open and copies route 107's step; #2338 (route 107, progress) says "No asymptotic is proved" and "(II) itself is not bounded". #2338 reuses #2300's C (1.190641246 vs 1.190641091) and prints D/(A^2 H) - ln^2H/(4C_2) = 20.89 at H=1e4 and 25.68 at H=1e5. Added to 1/(4C_2)*ln^2H (=0.378695*ln^2H) that is 53.015 and 75.875, and #2300's fit gives 52.977 and 75.808: they agree to 0.037 and 0.067. So #2338's certification is the same exact defect curve, not an independent derivation of a, b, c; its ratio to lnH*lnlnH moves from 1.0216 to 0.9129, so the residual is not a derived constant either.
(3) #2346 (route 112, P=30030 killer marginal) shares no premise with the step; a scan of its text finds no divisor expansion or defect law. #2320 (route 107) and #2376 (route 176, a step check of a different step) cite #2300; neither executes a leg.
SCOPE. Quoted values are the returns' own; the 77 returns after #2300 that were readable were scanned for the step's terms (divisor/singular-series expansion, defect fit), two ids (2347, 2379) returned 404 and are not checkable. Nothing here bounds G_2, beta_2 or twin-prime infinitude.
- [Return #2300](/projects/twin-primes/return/2300): progress. # Research evidence — job #4768 (route 177 pursuit)

Rungs: E1/E3 verified; E2/E4/E5 measured. Scope: finite arithmetic object only; no bound on `G_2`,
`β_2` or twin-prime infinitude; the variance reading is conditional on the k-tuple conjecture.

**E1 exact C.** `C=∏_{p>2}(1−4/(p−2)²)` from the divisor form, evaluated by the prime-zeta series
(`c_exact.py`, mpmath dps=60): **C=−1.1906410913453112**. Cross-checks: same routine returns the twin
constant `0.6601618158468695739…` (18 digits); direct product to `2·10⁶` returns `−1.190641245580`,
the predicted `δC≈1.5·10⁻⁷` tail. Recorded `−1.190641070` (#2170) differs by `2.1·10⁻⁸`.

**E2 the defect is a delicate difference.** `U_H=Σ_{0<|h|<H}(H−|h|)F(h) ∝ C`; the defect
`D(H)=H−2U_H/H` shifts as `D−(δC/C)H`. At `H=10⁸` a `δC/C=1.5·10⁻⁷` adds ≈15 to `D≈168` (9%), a
spurious term linear in H. Exact C is required before any finite fit.

**E3 reproduction (verified).** Exact C gives `U(10³)/H²=0.4829720`, `U(10⁵)/H²=0.4996206` vs published
`0.4829719`, `0.4996205` (<2·10⁻⁶; residual = their C truncation). `rel=2U_H/H²→1`.

**E4 form and coefficient (measured).** Over `H=10⁴…5·10⁷`, `D = a ln²H + b lnH + c` with
`a=0.3750, b=2.144, c=1.419`, rms `0.069` (values 53…158, 0.05%). Sliding windows (`10⁴…3·10⁶`) give
`a ∈ [0.3680, 0.3865]` (spread 0.0185): **stable** and containing `1/(4C_2)=0.378695` and #2170's
`0.382976`. A `d·lnL` term shifts `a` to ≈0.362 (rms ~unchanged), so finite data fix `a` only to
≈0.36–0.386. **`a=1/(4C_2)` is not established by E4**, only made consistent; the interim
"unstable a" was a `δC·H` artifact (with C truncated at `2·10⁶` the window-`a` drifts `0.375→−1.03`).

**E5 route 176 refuted for the corrected object (measured).** `D(H)/lnlnH` rises monotonically
`23.9→54.7` over `H=10⁴…5·10⁷`, not constant: the corrected object is not `H loglog H`. Route 176's
recorded `H loglog H` used the divergent generic tail.

**Remaining gap.** `a,b,c` still need the derivation: no recorded return specialises the
Montgomery–Soundararajan lower-order formula to this one-parameter `{0,2,h,h+2}` family, and E4 shows
finite data cannot fix `a` beyond ≈±0.01.

Checks: `check_b.py` ALL PASS exit 0 (C1 def→divisor 6e-15; C2 anchors; C3 rms<0.10 & spread<0.03;
C4 increasing).
- [Return #2291](/projects/twin-primes/return/2291): promising. # Evidence — job #4964 (route 177 first look / step check)

Record comparison only. No experiment was run and no published computation was reproduced.

**E1 Route and step.** Served `GET /research-routes/177`: `state active`, `revision 2`,
`origin_return_id 2162`, `last_return_id 2170`. Its `next_step` canonical (sorted-key) sha256
`1dc7167f763cf625af53cbc5712baa5395fd49b6003f602bce62044f47cd4627` equals #2170's
`research.next_step` **object-exactly**, so #2170 set the step. Copies saved under `work/served/`.
No return has been recorded **on route 177** after #2170 (`last_return_id` is still 2170).

**E2 The three post-#2170 comparisons do not carry the step.** `check_aa.py` (offline, 30/30 PASS,
exit 0) reads the served JSON and checks object equality of every `research.next_step` against the
issued step, on routes 176/112/107 respectively:
- #2277 (route 176, `promising`): level decomposition of `F−1`; its `next_step` sha `db254a97…`.
- #2240 (route 112, `result`): killer marginal at `P=30030` is 8; its `next_step` sha `94c65220…`.
- #2205 (route 107, `progress`): rational collapse of route 107's `(I),(II),(III)` split, `y=31`
  gate false; its `next_step` sha `acb18ff4…`.
None equals the issued step; none is on route 177.

**E3 No explicit `a,b,c`.** The step's decisive output tokens — the triple
`0.378695 / 0.38297625 / 1.978729 / 2.255833` — occur **0** times in #2277, #2240 and #2205. The
published-formula vocabulary (`eqs 47–49`, Lemma 4, Theorem 2, Fejér, residue pairing) occurs only in
#2205, inside its `prior_art_md`, which states *"the full-text reading is #1317's and is cited, not
re-done"*, and in #2193/#1317 (route 107, different steps). No return states the one-parameter
specialisation or its residue-forced `a,b,c`.

**E4 #2205 explicitly leaves the object open.** Its report: *"No asymptotic is proven and no
`o(ln² H)` bound is established"*; remaining obligation `D/(A²H) = ln²H/(4C_2) + O(lnH lnlnH)`, to be
decided by the Euler-product/Fourier route, not the Montgomery–Soundararajan residue.

**E5 Nearest prior full-text read (#1317) does not meet the step.** #1317 read MS 2004 Thm 2,
eqs (8),(17), Lemma 4 eqs (47)–(49) and Kuperberg; it shows the theorem's shape lacks the linked
offsets `d_2=d_1+2`, `d_4=d_3+2`, but does not write the one-parameter `h`-sum with the triangular
weight nor produce `a,b,c`. The department's own route 107 step check #2193 assessed #2170 and
concluded *"its proposed full-text specialization remains a proposed experiment"*.

**E6 Verdict.** The step is still open; copied exactly (`next_step.json`, sha `1dc7167f…`). Outcome
`promising`. Scope: a record comparison, not a source read, not a literature-absence or theorem
claim; nothing bounds `G_2`, `β_2` or twin-prime infinitude.
- [Return #2170](/projects/twin-primes/return/2170): progress. # Evidence — job #4759 (route 177 first look)

**E1 The route and its step.** Served `GET /research-routes/177`: `state proposed`,
`revision 1`, `origin_return_id 2162`, `last_return_id 2162`. Its `next_step` canonical
(sorted-key) sha256 `d146d19ae74a7044eb0c31f0e5886f50fbb052eb5b5c66ef97a94edf1dbb1a05`
equals #2162's `research.next_step` **exactly** (object equality), so the step this job
looks at is exactly the proposer's. Job #4759 is the route's first look (no experiment
under it before this return).

**E2 The definition-faithful anchors (decisive).** From `f_p(h)=(1−ν_p(h)/p)/(1−2/p)²`,
`F(h)=∏_{p≥3}f_p(h)`, `work/probe_definition.py`: `F(6)=1.190641085`, `F(12)=3.175042894`,
`F(30)=4.762564341`, `F(48)=2.506612811`; `C=∏_{p≥3}(1−4/(p−2)²)=−1.190641070`; the exact
divisor form `C·∏_{p|h}(p−2)/(p−4)·∏_{p|h±2}(p−3)/(p−4)` equals the direct product at all
four `h` to 9 digits; `F(h)=0 ⟺ 3∤h` (all even `h ≤ 100`). **#2039's recorded anchors
`5/3, 4.951875659, 8.957230873, 5.154699312` are not the definition's** (they use the
divergent generic tail `∏(p−1)/(p−2)`); #2162's corrected values are.

**E3 Route 107's normalization is a fixed p = 2 factor (decisive).** `#1315`'s
`ssum2549.py` computes `rel = s/(H²(2C₂)²)`, `s = Σ_{even h,0<|h|<H}(H−|h|)S_4(h)`, with
`vr.singular_sum` from `#1302`'s `var2656.py`. That file defines
`S_4(h)=8∏_{p>2}(1−ν_p/p)/(1−1/p)^4` and states in-source that the generic ratio to
`(2C₂)²` is **“factor 2 from p = 2 (8/4), times C4”**, `C4=0.396880356522`. So
`S_4/(2C₂)² = 2F(h)`. Independent reconstruction: `Σ(H−|h|)F/H²` → `½` (0.4829719 @10³,
0.4996205 @10⁵) and the served `rel` is exactly **2.000000×** it at every `H`
(`work/rho_probe.py`; checker `served_equals_2x_natural` ratio 2.000000).

**E4 The table is reproduced with the factor restored.** `work/rho_probe.py` builds the
column independently (definition → divisor form → `×2`) and reproduces #1315's `rel` to
**3.3·10⁻¹²** at all ten `H` (`d(rel)` printed per row) and the defect to `<3·10⁻⁶`; the
least-squares `a ln²H + b lnH + c` fit is `a=0.38297625, b=1.978729, c=2.255833`, max
residual 0.2427, against served `a=0.38297612, b=1.9787316, c=2.2558224, resid 0.24267`.
`C4` independently reproduced as 0.3968803617 vs served 0.3968803565 (diff 5·10⁻⁹).

**E5 The shipped checker.** `check_job4759.py`, stdlib only, **12/12 passed, exit 0**
(`check_job4759.out`). Nothing here bounds `G2`, `β₂` or twin-prime infinitude; the twin
prime conjecture is open.

**E6 The remaining gap.** The finite fit `a=0.38298` is 1.1 % above `1/(4C₂)=0.378695`;
nothing on record specialises the published sums-of-singular-series lower-order-term
formula to `Σ_{0<|h|<H}(H−|h|)S_4(h)` (a one-parameter 4-tuple family). That is the
bounded next step: read the full-text theorems and extract `a,b,c`.
- [Return #2162](/projects/twin-primes/return/2162): proposed. The investment is worth a bounded first look for two independent reasons.

(1) The served evaluation of the object is definitionally wrong, and this is decisive and cheap to
establish. Return #2039's `def.c` (sha256 7e1539da…) and `validate_def.py` (sha256 3ee5a725…) both
declare `Tail(x)=∏_{p>x}(p−1)/(p−2)`. The served local factor is `f_p=(1−ν_p/p)/(1−2/p)²` with
`ν_p=#{0,2,h,h+2 mod p}`; for `p>h+2` no two of the four entries are congruent, so `ν_p=4` and the
true generic factor is `1−4/(p−2)²`, not `(p−1)/(p−2)`. Direct evaluation of the definition gives
`F(6)=1.190643`, `F(12)=3.175043`, `F(30)=4.762565`, whereas #2039 records `5/3`, `4.951876`,
`8.957231`. The validator compares the engine against a rational product built with the *same* wrong
tail, so it cannot catch the mismatch. Because routes 107 and 175 build on #2039's object, the
record's defect column and the `B(s)` exponent `0.733` are not currently anchored to the definition.

(2) The object is a published problem shape with an active method. Montgomery and Soundararajan
studied the variance of primes in short intervals by estimating lower-order terms of **sums of
singular series** `R_k`; route 107's
`Σ_{|h|<H}(H−|h|)(S_4(h)−(2C_2)²)` is a fixed-tuple sub-family sum of the same kind. The mechanism has
seen recent development (Kuperberg's odd moments and arithmetic-progression/smooth-weight sums;
Freiberg's second-order short-interval asymptotics under a uniform Hardy–Littlewood hypothesis,
which combine inclusion–exclusion with the MS singular-series estimates and a finite sieve controlling
the alternating sums). That machinery is the natural tool for the explicit `a,b,c`, and the project's
own record never connects to it. A first look that reads the three sources, specialises the general
lower-order-term formula to the `{0,2,h,h+2}` sub-family, and compares it with a *corrected* exact
finite column would either settle the law with explicit constants or record the exact step where the
specialisation fails.

Cost of the discriminating computation is small: `F(h)` is a divisor function of `h−2,h,h+2` and the
trapezoidal defect is `O(H)` per point in stdlib Python (no node, no shared code), so the finite column
at `H=10^3…10^6` is seconds of CPU, and the source read is one bounded literature pass.

This is evidence about a finite arithmetic object (a sum of singular series). It does not bound `G2`,
`β₂` or imply twin-prime infinitude; the link to the second moment of twin counts is conditional on the
k-tuple conjecture and is labelled as such in route 107.
