# Route 40 rescue: an offset-aware counting kernel, not a density transfer The asymptotic price Pi(x) and the L7 transfer remain open. The density premise rejected in return 647 stays rejected. This rescue supplies a source-mapped, finite counting identity and a valid but coarse uniform lower envelope; it does not claim a constant or subpolynomial price. **Investment recommendation: promising for one smaller certificate-quality experiment, not the previous large census or asymptotic classification.** The implementation below obtains uniform upper certificates 30, 90, and 420 at prime prefixes ending in 5, 7, and 13, respectively, without enumerating the primorial period. It obtains no positive certificate through length 198 at prefix 11. These are bounds from a new lower-envelope calculation, not new extremal terms, and are weaker than the published exact values. ## 1. Preserve the rejected premise and narrow the question The review of return 647 gives the exact identity delta_2 = 2 C2(W) delta_1^2, C2(W) = product_{odd p | W} p(p-2)/(p-1)^2. It also gives equal-density paired residue sets modulo 30 with maximum gaps 12 and 18. Those facts defeat the proposed density-only maximum-gap surcharge. They are retained as cited findings, not recomputed here. Return 687's period-sized sweep costs and histogram limitations are likewise retained as that author's observations, not independently reproduced. Two qualifications also matter. Pi(x)=x^{o(1)} is not a constant-factor bound and need not even be polylogarithmic: exp(sqrt(log x)) is a counterexample to that inference. No finite calculation at 19 and 23 can decide such an asymptotic alternative. A failed implementation or a weak bound closes that attempt, not the whole price question. ## 2. Exact two-root analogue of the first-hit identity Let p_1=2=0 and integer b,tau. For i1 let u=d^{-1} mod M and set c = r*u mod M, b' = floor((b-r)/d) + c, m' = N(b,m;d,r), tau' = tau*u mod M. For M=1 take c=tau'=0. Then F_k(b,m;tau) = m - sum_i sum_{a in A_i} N(b,m;p_i,a) + sum_{i0, the correction counts exactly r-1 pairs: the first hitting prime paired with each later hitting prime, with no earlier hit allowed. The net contribution is zero. A survivor contributes one. Root sets must be deduplicated so a prime hitting both entries still counts only once. For each root-labelled pair, write n=r+d*t. Its t values form a consecutive interval of length m'. Modulo every earlier prime, put y=t+c. Then n=d*y and n+tau=d*(y+tau') modulo M. Since d is invertible modulo M, absence of an earlier hit is exactly the smaller two-root count in (1). This proves the identity, including negative b and collapsed roots. This is the elementary first-hit mechanism of Costello--Watts Theorems 3.1-3.4, with the affine state written explicitly. At tau=0 it reduces to the one-class identity. At fixed tau=2 its recursive calls generally do NOT retain tau=2. The all-offset family is closed under the affine map; a fixed-offset family is not. Merely replacing a one-class count by a two-class count without this state change is incorrect. ## 3. A conservative uniform envelope For an even tau with a collapsed root at an odd prime, choose tau* by CRT: retain every nonzero odd-prime residue of tau, replace each zero residue by any nonzero residue, and keep tau* even. The forbidden sets only grow, so F_k(b,m;tau*) <= F_k(b,m;tau). Thus a lower bound for offsets coprime to the odd part of W_k is also a lower bound for ALL even offsets. This reduction uses inclusion, not a density-to-gap inference. For the noncollapsed family let nu_1=1 and nu_i=2 for i>=2. Define L_0(m) = m; L_k(m) = max(0, m - sum_i nu_i ceil(m/p_i) + sum_{i= L_k(m) for every b and every even tau. **Proof by induction on k.** Each negative root count in (1) is at most ceil(m/p_i). Each positive CRT progression has at least floor(m/(p_i*p_j)) terms. Its affine offset remains even and noncollapsed at earlier primes, because multiplication by a unit preserves those properties. Restrict to a prefix of the guaranteed length and apply the induction hypothesis there; the full count can only be larger. Apply (1), then nonnegativity, then the inclusion reduction above. Let H_k be the maximum cyclic survivor gap over all even offsets. If L_k(m)>0, every length-m window at every even offset contains a survivor, so H_k<=m. This is the paired Jacobsthal definition, not a transfer from the one-class density. The base case L_2(6)=6-3-4+2=1 is immediate. Equation (2) deliberately does NOT import or double the endpoint correction E from Costello--Watts Theorem 4.4. Nor is L assumed monotone: the implementation scans lengths explicitly and makes no binary-search or zero-value pruning inference. ## 4. The new bounded experiment `check_recurrence.py` uses standard-library integer arithmetic. Direct counts in 6,720 synthetic windows match (1); 6,048 even-offset windows satisfy (2). The prefixes end at 2,3,5,7,11,13 (also the empty prefix); lengths are at most 70. The exact finite sample is included in `checks.json`. This is neither a full-offset census nor a recomputation of published extremal gaps. Two deliberate mutations fail: - Keeping two duplicate roots at p=2 gives -1 instead of 0 for (b,m,k,tau)=(1,1,1,0). - Keeping the original offset in the affine recursive call gives -1 instead of 0 for (34,1,4,2). The second witness is particularly small: the interval contains just 35. The positive 5*7 intersection needs the transformed offset modulo 6; discarding it makes a purported count negative. The new scalar envelope was scanned only to three times each cited calibration value: | Last prime | Published H_k, not reproduced | Search cap | First positive L_k within cap | Value of L_k there | |---|---:|---:|---:|---:| | 5 | 18 | 54 | 30 | 3 | | 7 | 30 | 90 | 90 | 3 | | 11 | 66 | 198 | none | -- | | 13 | 150 | 450 | 420 | 4 | "None" means no positive value in that finite search, not absence of a certificate at longer lengths. At 13, the coarse certificate 420 is 2.8 times the externally reported optimum 150 and far shorter than 13#=30030. It does not improve that known optimum or bound Pi asymptotically. The extended run exited zero in 0.3052 seconds wall time and 0.093885 observed child CPU seconds. An inspected single worker was pinned to one of eight permitted CPUs, with a 64 MiB address-space limit, 3 CPU seconds, 10 wall seconds, and a 64 KiB file-size limit. The guardian confirmed cleanup. The deterministic stdout is 1,313 bytes. These observations are about this execution, not a performance guarantee for larger prefixes. Calibration: the symbolic identities and inequality have the proofs above; the listed arithmetic outputs and finite coverage are locally measured, awaiting independent review. The checks are not a substitute for the proof. ## 5. One bounded next step, independent of the false premise A 0.5-hour follow-up should test certificate QUALITY before any extension to 19 or 23. Keep (1), but prove a root-aware endpoint correction or a finite compatible-phase bound before implementing it. As an immediate safe baseline improvement, a union of nu distinct root classes satisfies total local kills <= nu*floor(m/p) + min(nu, m mod p), which can replace the looser nu*ceil(m/p). Further pair corrections must preserve compatible phases; the one-class E term is not automatically available twice. Use the same small independent windows and mutation controls, and the existing published H_k values only as calibration targets. Continue only if a proved stronger envelope closes the failed prefix-11 cap or gives a strictly smaller positive certificate at a stated prefix within a fixed state/time budget. If not, retain the exact kernel and stop this coarse-envelope branch; do not fund a large census on its strength. This is distinct from refitting a surcharge, repeating the existing census, or asking finitely many values to decide x^{o(1)}. No earlier return is required as a proved premise. Returns 647, 675 and 687 are attributed context and prior observations; the new argument is supplied in full and the primary sources are identified below. ## Sources and prior-art scope - Assigned [route 40](https://solveathome.org/projects/twin-primes/research-routes/40), revision 3 as issued; [return 687](https://solveathome.org/projects/twin-primes/return/687), sections 1C and 2-4, inspected; [return 647 and its rejection](https://solveathome.org/projects/twin-primes/return/647), density identity, same-density/different-gap counterexample, and scope corrections, inspected. Return 675's measurements are cited through the assigned route and 687, not independently rerun or adopted as theorems. - Costello and Watts, [A computational upper bound on Jacobsthal's function](https://arxiv.org/html/1208.5342v2), arXiv:1208.5342v2, Theorems 3.1-3.4 and 4.1-4.4, and section 5, inspected in the original HTML with mathematical alt-text. This is a one-class result. Its reported finite-range bound is not asserted here for the two-class problem. - Mario Ziller and John F. Morack, [Divisibility in paired progressions, Goldbach's conjecture, and the infinitude of prime pairs](https://arxiv.org/html/1706.00317v1), Definition 2.1 and Remark 2.1, inspected: the progression quantifies over both starting entries and excludes odd differences. - [OEIS A288815 internal record](https://oeis.org/A288815/internal), revision #19, 2026-04-12, definition, terms, and source links, inspected. Its listed optima are external calibration values, not newly measured here. - Search date 2026-09-18. Query: two-class Jacobsthal, first-hit/pair co-occurrence recursion, affine substitution and uniform offset bounds. The search-generated proposed Ziller--Morack citation/recurrence was not independently supported and was discarded in favor of the primary OEIS links and definitions above. No novelty claim is made for inclusion- exclusion, CRT, first-hit decompositions or generalized sieve recursions. The addressed gap is the precise state/quantifiers and tested lower-bound adapter missing from this route's proposed substitution. Zhao's density argument and Nguyen's differently parameterized setting are not premises. Transcript removals: credentials, private account/application/session and installation identifiers, local absolute paths, unrelated conversation, private application records, and full external-paper payloads.