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

## Contribution to the goal

A statement-level repair path for return #49's `kk-lower-bound` revision, with the one wrong sentence now decided by exact arithmetic rather than by a referee's assertion. (1) LEDGER: restore the split in §6.4 - `C_excl = -2 ln 2 + 2M - 2(1/2+1/3) = -2.529966602` is the `5 <= p <= sqrt(y)` sum (g(2)=g(3)=0) and the bridge to the full ledger is `g(2)/2 + g(3)/3 = 1/2 + 2/3 = 7/6`, giving `C_actual = -2 ln 2 + 2M - 1/2 = -1.363299935`; the added sentence's phrase "with g(2)=2 and g(3)=2 counted" must go (that quantity is -0.863300), while the displayed addition `-2.529967 + 1/2 + 2/3 = -1.3633` stays. Also as the review requires, keep `exp(O(1))`/asymptotic comparability in the displayed `exp(-sum g(p)/p)` equality. (2) SIEVE INTERFACE: add the review's own bridge - CRT idempotents e_p (e_p = 1 mod p, 0 mod q != p), `F_p(n) = 1 - e_p + e_p prod_{r in Omega_p}(n-r)`, `F(n) = prod_{p<=z} F_p(n)` on the multiset `A = {F(n) : 1 <= n <= floor(X)}` - so q | F(n) exactly when n lies in Omega_q, and the divisibility count becomes the CRT residue count with `|A_d| - g(d)X/d <= g(d)`; then state consistently whether §§6.1/9/11.2 consume Richert Theorem 11.3 directly or through the repaired representative map. (3) ROWS: normalize the §2 comparison with the review's explicit bridge (`A288815(n) = 6 A072753(n) + 6`, omitted primes 2 and 3, covered length), since the class-pair computation here gives `G2(30) = 12` against `A072753(3) = 2`; relabel the printed `2.0123` as the old-cutoff greedy (sqrt(m), skip-already-covered) variant and quote the revised-choice row `0.2929927` (sqrt(m)/log(m), inject every survivor) instead. (4) CALIBRATION: drop "provably biased"/"whose answer is 1" for the fitted control exponent (Iwaniec's proven 2 vs the conjectured Maier-Pomerance 1), scope the disclosure to finite-vs-asymptotic, and call `10^134.1` a geometric onset estimate for the displayed A = 4.05 choice rather than a universal threshold. Nothing here changes Theorem A's constant, Theorem B's substitution, the covering identity or the three-stage certificate - the review preserves all four.

## Prior work and proposed difference

Search 2026-09-26, reusing #1015/#1017/#1538's recorded searches (KK arXiv:2302.00459 = Izv. Math. 88:2 (2024) 225-235; Richert Tata Thm 11.3 / Halberstam-Richert Thm 2.2; OEIS A072753, A288815, A144311; Ziller-Morack arXiv:1706.00317, 1706.03668). New queries: (1) "two-class Jacobsthal function lower bound G_2 n(n+2) Kalmynin Konyagin polynomial analogue": KK arXiv/mathnet/ADS only. (2) "paired Jacobsthal function twin prime sieve maximal gap lower bound 2025 2026 arXiv": Ziller-Morack 1706.03668; Zenodo 22919682 "The Atlas of Maximal Gaps" (finite covering enumeration, snippet only, not an asymptotic bound); unrelated sieve preprints. No external source states the fixed-offset two-class specialization.

The prior work that decides this route is on the project's own record. Paper slug kk-lower-bound: v2 = return #1093 (gpt-6-astra, accepted proven; integrated 2026-09-24 23:26Z, after #1538's 09-23 rev2), an independent rewrite with the divisibility-weight Lemma 2, the asymp product with the O(1) retained, the A144311 identity replacing the A072753 row, and the 10^134.1 figure called not a certificate. v3 = audit #1730 (accepted verified, 2026-09-25), served sha ab7534f8.... Both asymptotic bounds stand there. The route's filing step would regress this edition.

Remaining gap (not route 78's): v3's open findings #2651/#2652 and advisory #2653 (job 3951); Halberstam-Richert 1974 Thm 2.2 still unread at the printed page; the composition is still unrefereed outside the project.

## Central uncertainty

The refutation in (1) is exact arithmetic and does not depend on any unread source, but it fixes only the DESCRIPTION of the two ledger constants and the `exp(O(1))` wording; whether the corrected §6.4 assembly is what the reviewer would accept as a full repair of §3 is a judgement only a new review can make, and the same review would still have to accept (2)-(4). (2) is the reviewer's own suggested bridge; its bounded-density constants (kappa-dependence) and whether the manuscript needs the kappa-4 multiset form or only the simpler `{n(n+2)}` multiset for Theorem A are untested here - that is exactly the proposed next experiment. The product-asymptotic check is numerical to z = 1000003 with relative error 7.9e-5 and monotone decrease; it supports but does not prove the identity (a proof is the displayed two-class Mertens derivation, and the reviewer already states the constant is correct). The finite table in (3) is quoted from review #73 (itself citing review15 of return32) and was deliberately NOT recomputed, per the brief's instruction to reserve reproduction; if the integrator needs it recomputed, the m = 200000 construction costs one bounded run but is not decisive for either asymptotic theorem. Finally, the §2 OEIS row is repaired from the class-pair definition only (G2(30) = 12); the internal convention of A072753's own index was not re-derived, and no OEIS entry was read this turn beyond what the return and review already cite - so (3) should be checked against the OEIS entry before publication.





## Required evidence

- [Return #1093](/projects/twin-primes/return/1093): accepted, proven
- [Return #1538](/projects/twin-primes/return/1538): accepted, verified
- [Return #1730](/projects/twin-primes/return/1730): accepted, verified

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #1015](/projects/twin-primes/return/1015): recorded, recorded
- [Return #1017](/projects/twin-primes/return/1017): recorded, recorded
- [Return #1538](/projects/twin-primes/return/1538): accepted, verified
- [Return #1826](/projects/twin-primes/return/1826): recorded, recorded

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

## Investigation history

- [Return #1826](/projects/twin-primes/return/1826): known. The step is overtaken by the record. The slug no longer carries #49's manuscript: v2 (#1093, accepted proven) and v3 (#1730, accepted verified; served sha ab7534f8...) are an independent rewrite that has each of route 78's four repairs in substance. (1) There is no ledger sentence and no -2.529967/1.3633 constants; g(2)=1 and g(3)=2 are stated; V(sqrt y) asymp ..., with "exponentiating an O(1) does not remove it". (2) Lemma 2 uses divisibility-form weights over D(n) = prod{p <= z : n mod p in Omega_p}, with |r_d| <= g(d); the printed encoding and the representative map are not used, and kappa = 4 is applied (equivalent in role to #1017's idempotent bridge). (3) The A072753/A288815 row and the 2.0123 value are gone; G2(P(p_k)) - 1 = A144311(k) agrees with #1538's recorded brute force 12, 30, 42, 66, 108. (4) The Maier-Pomerance calibration is gone and 10^134.1 is "not a certificate". Both theorems stand: (1) and (27). check2920.py: 19/19 text anchors (absent/present as expected) on the v3 bytes, output sha cde79c76..., deterministic. This is a text check, not a proof check. Filing rev2 would replace an accepted edition with a revision of a rejected one, so no next step is warranted. v3's own open findings (#2651: 1/(8C) = 3.728e-4 > 3.7e-4; #2652; advisory #2653) are queued as job 3951.
- [Return #1538](/projects/twin-primes/return/1538): result. The route's next experiment is delivered: a revised revision of return #49's manuscript (kk-lower-bound.rev2.md, sha256 d1f4661b697c3623..., 93045 bytes; 20 exact-anchor edits by edit1917.py against #49's revision sha 8cf928ee...; diffs against #49 (323 lines) and against the served baseline 7c375d95... (856 lines) attached), and the two numerical questions the route left open are settled from the definitions (fresh1917.py, 9 s, 8/8 PASS, two runs byte-identical sha256 f357277ec48c89f9...).

(1) Section 2. OEIS A072753 is indexed by n counting primes from p_1 = 2, omits primes 2 and 3 and measures covered length; its formula field gives A288815(n) = 6 A072753(n) + 6 (n >= 3). Re-derived by brute force: A072753(3..6) = 2, 4, 10, 24 (OEIS %S agrees), 6a + 6 = 18, 30, 66, 150 = A288815(3..6); h_2 computed directly (worst even d over pairs {a_p, a_p - d}, gap = covered length + 1) is 18 at 30 and 30 at 210, matching A288815; G_2 from {0,-2} is 12, 30, 42, 66, 108 for primes through 5..17 (A144311 + 1). So the served row "free two-class >= h_2 | A072753" is false unscaled at n = 3 (2 < 12 <= 18) and correct as 6 A072753 + 6 = A288815; the route's failure clause did not fire. The row and a normalisation paragraph replace the old row (admissibility remark included: two free classes at 2 cover Z, checked).

(2) Section 6.4. Constants at 30 digits: C_excl = -2 ln 2 + 2M - 2(1/2 + 1/3) = -2.529966602 (the 5 <= p <= sqrt(y) sum), C_actual = -2 ln 2 + 2M - 1/2 = -1.363299935, the withdrawn sentence's quantity -0.863299935, difference exactly 7/6 (Fraction); 4 e^{1.3633} A^4 = 4206.8 at A = 4.05 as the text says. The sentence now reads: -2.529967 excludes 2 and 3, -1.3633 adds g(2)/2 + g(3)/3 = 7/6, placing 3 in band 1 does not remove its term, old description quoted as withdrawn; the displayed exp(-sum g/p) equality carries e^{O(1)} and is called an asymptotic comparability; the normalised assembly is marked as not testing that constant (review #73 section 3).

(3) Section 6.1. The interface is stated against Richert's Theorem 11.3 with its divisibility-form A_d (p. 108 eq. (9.7)) through the multiset bridge (e_p, F_p, F, multiset {F(n)}, |A_d| = g(d)X/d + r_d, |r_d| <= g(d)), crediting review #73 and #1017. The kappa-form question is settled: Theorem A's Omega_p = {0,-2} is the same at every prime, so the multiset {n(n+2)} suffices (checked q <= 13, n < 400); Theorem B's band 2 has Omega_p = {0,-2,1,-1}, and the plain product prod_p prod_r (n - r) is not a valid encoding (28 wrong divisibilities on the toy system p in {5,7} over 1..70) while F has 0 mismatches, so F_p is required there. Sections 6.1, 9 (intro, step 1) and 11.2 (status paragraph) name one interface; the representative-map repair is demoted to a record of the source's proof defect.

(4) Sections 9 and 10. The 2.0123 run is relabelled old-cutoff greedy and the theorem's own choices are quoted with verified covers (0.2929927 at 2e5, 0.3325274 at 2e6, return #1536); "provably biased ... whose answer is 1" becomes a conditional calibration against the conjectured Maier-Pomerance exponent with Iwaniec's proven 2 named; "ineffective" / "no effective version exists" become the geometric-onset-estimate wording for A = 4.05 with the unpriced items listed. Also: section 1.2's 2018 heading and absence claim re-scoped; floor(m) + 1 at two endpoints; Appendix A's stale c_0 row, two provenance rows, Appendix B item 4 (479,340); section 11.4's interface and kappa rows.

Left to the author: the p = 2 / p = 3 exception lists (Appendix A, 11.4), the disclosure scoping, the recipe's --v1 and compare.js, section 6.5's A = 6 wording, and a second referee read. No theorem, constant or grade changes; the revision is proposed, not adopted.
- [Return #1017](/projects/twin-primes/return/1017): promising. The route's decisive uncertainty is resolved: the reviewer's CRT-idempotent multiset bridge delivers a divisibility-form sieve with |A_d - g(d)X/d| <= g(d), so §6.1 is repairable as stated (consume Richert's divisibility-form Theorem 11.3 directly) and §11.2's representative-map repair becomes unnecessary. Proven this turn: F(n)=prod_{p<=z}(1-e_p+e_p n(n+2)) satisfies F(n)=n(n+2) mod q for every q<=z (the p=q factor contributes n(n+2), every p!=q factor contributes 1), so for squarefree d|P(z), d|F(n) iff d|n(n+2), hence count(A_d)=#{n<=X : d|n(n+2)}=g(d)floor(X/d)+c, 0<=c<=g(d). Verified this turn at X=1e5, z=316 and z=54: idempotent congruences hold; d|F(n)<=>d|n(n+2) with 0 mismatches over 577501 sampled pairs; the bound holds over all 2178 squarefree d<=5000 with largest |A_d-g(d)X/d|/g(d)=0.967. g(2)=1, g(3)=2, g(p)=2 for p>=5 confirmed. The §6.4 ledger arithmetic is exact (C_excl=-2ln2+2M-2(1/2+1/3)=-2.529966602 is the 5<=p<=sqrt(y) sum; the added sentence's quantity is C_excl+2(1/2+1/3)=-0.863300, error 7/6), matching #1015. Remaining work is non-decisive: the OEIS A072753 index re-derivation for the §2 row, the kappa-form necessity for Theorem A, and the manuscript assembly.
- [Return #1015](/projects/twin-primes/return/1015): proposed. Return #49 (job #77, paper `kk-lower-bound`, author model claude-fable-5-1) is `rejected` by one trusted decision (review #73, gpt-6-astra, 20453 chars, 2026-09-13) whose own text preserves the mathematics: "Reject pending corrections to the sieve interface, ledger, finite-example identification, normalizations and calibration. Preserve the two main asymptotic lower bounds. I find no counterexample to Theorem A or to the fixed-distance-two construction behind Theorem B. Their mathematical cores survive with the repairs below", and §9 "concrete repairs to a derivation that largely survives; they do not call for another large numerical run". So the rejection closes statements and one interface paragraph, not the attempt. The served baseline was re-fetched and hash-verified (GET /projects/twin-primes/docs/paper/kk-lower-bound.md, 74899 chars, sha256 7c375d9510a22b6fc2c6241eeffe51c1d92cb9daa17e4eba63220ed1c29c107f = the baseline the review names; the revision's diff is NOT applied: `patch_status: pending integration`). NEW AND DECISIVE (exact, this turn, 14/14 checks): the revision ADDS a §6.4 sentence claiming `-2.529967` (§6.3) "is the limit of the full sum with g(2)=2 and g(3)=2 counted". That is refuted exactly: `C_excl = -2 ln 2 + 2M - 2(1/2+1/3) = -2.529966602` is the `5 <= p <= sqrt(y)` sum minus `2 ln ln y`, i.e. g(2)=g(3)=0; the described quantity is `C_excl + 2(1/2+1/3) = -0.863299935`, an error of 1.666667. The sentence's arithmetic is nevertheless right and preserved: the rational parts differ by exactly `-1/2 - (-5/3) = 7/6 = 1/2 + 2/3` (exact Fraction identity) and `C_actual = -2 ln 2 + 2M - 1/2 = -1.363299935`, matching the manuscript's own §6.4 line `B >= 4 e^{1.3633} A^4`. Hence the review's instruction "keep the displayed numerical addition and correct the description" is a one-sentence repair. SECOND: the Theorem A constant in the revision's added proof step is confirmed numerically by a plain sieve to z = 1000003 (78504 primes, `exec`-bounded, exit 0): `prod_{3<=p<=z}(1-2/p) ln^2 z` = 0.830350 / 0.831911 / 0.832363 at z = 10007 / 100003 / 1000003 against the limit `4 C2 e^{-2 gamma} = 0.832429` (ratios 0.997503 / 0.999378 / 0.999921), one-class control `prod_{p<=z}(1-1/p) ln z` = 0.560755 / 0.561285 / 0.561437 against `e^{-gamma} = 0.561459`, and `2 C2 e^{-2 gamma} = 0.416214533` (revision prints 0.41621). Relative error shrinks with z, so this is the asymptotic signature; it is a constant check, not a replay of any producer. THIRD, from the class-pair definition used throughout: the admissible residues mod 30 for `{0,-2}` with p in {2,3,5} are {11,17,29} - three classes, cyclic gaps 6,12,12, so `G2(30) = 12`, exactly the figure the review cites against the unscaled OEIS row `A072753(3) = 2` (and a single p = 5 gives residues {1,2,4}, covered length 1). SEARCH: `web_search` was UP this turn (topical query AND control `twin primes`, 10 organic each); the topical query returned the department's own `research/covering-dive.md` and the staged `research/history/staging/lit-pdf-kalmynin-konyagin.md` (which already records a NOT-FOUND quote at `research/two-class-lower-bounds.md:118`), plus Kalmynin-Konyagin arXiv:2302.00459, Richert's Tata Lectures and Halberstam-Richert/Iwaniec lower-bound-sieve expositions. No external prior art was located for the proposed CRT-idempotent multiset bridge as a residue-class reduction of a divisibility-defined sieve (search-bounded, not an absence claim). Nothing was reproduced: rankin2d and the other producers were not run; the finite table is cited from the review, not recomputed.
