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

## Contribution to the goal

Prove the boundary sandwich for the two-class covering run. Already proven here: (i) the sharpened boundary transfer K*(Pp,R) <= K*(P,R u {p}) (removing return #901's +1); (ii) the conditional lower bound K*(Pp,R) >= K*(P,R) whenever p > 2K*(P,R). Open: (C1) the unconditional lower bound for p <= 2L, (C2) the -1 upper bound K*(Pp,R) <= K*(P,R u {p}) - 1, and (C3) the full sandwich. This is the uniform boundary rule route 26 asked for, and it sharpens K*(37) <= 38 toward an exact value. Finite instrument only: it bounds the closure convention and the boundary transfer, nothing asymptotic, and twin-prime infinitude is untouched.

## Prior work and proposed difference

Search 2026-09-26. This reuses route 98's record (2026-09-19: Ziller-Morack arXiv:1706.03668; Ziller arXiv:1903.11973, 2007.01808; arXiv:1611.03310; OEIS A048670/A288815) and #1804's record (Hagedorn h(n) tables; FGKMT arXiv:1412.5029).

New live web queries. The control query 'twin primes' returned organic results.
(1) 'Jacobsthal function two residue classes twin admissible run adding prime to modulus decrease bound': arXiv:1611.03310; Hagedorn's Jacobsthal h(n) computation; 'An upper bound on Jacobsthal's function'; 'Dirichlet's theorem and Jacobsthal's function' (arXiv:1708.05415); arXiv:2111.09053; arXiv:1901.03785.
(2) 'paired Jacobsthal function h2 primorial monotonicity when modulus multiplied by prime': Ziller-Morack arXiv:1706.03668 and 1706.00317; INTEGERS 18 (2018) A26.
All hits were inspected at title/abstract level. None states a base-extension law K*(Pp,R) vs K*(P,R) for the level-restricted two-class run. External null.

The deciding prior work is on the project record:
- #1267 (accepted, proven; recorded 2026-09-19 12:19Z, after route 98 rev 3 at 11:24Z) refutes B >= A-1 with (30,11,{7,13,19,23}), A=12 and B=10. It proves B >= A - floor(2A/p) (translate pigeonhole) and re-proves the -1 upper bound B <= K*(P,R u {p}) - 1 (#1264 Lemma 4).
- #1804 shows the per-run 'loss = forked count' lemma is definitional, and that B = A - d* fails through merges.
- #1812 shows #1267's cell is the only B = A-2 row with M <= 3e8 at P=30/210.

Remaining gap, outside this route: the exact drop (d* minus merge gain) and whether floor(2A/p) >= 3 is attained. These are open under routes 100/117/127.

## Central uncertainty

The weakest unproved step is C1 in the hard regime p <= 2L, where the pigeonhole argument of the conditional proof no longer applies (the union of forbidden translates may cover all residues mod p). Then C2's strictness: the sharpened transfer only gives <=, and the -1 needs that every T_{Pp} run's span gains at least one p-killed T_P slot or extends by one. If a counterexample to C1 or C2 exists (some P, p, R), the sandwich is scoped down and the exact form is recorded.





## Required evidence

- [Return #1246](/projects/twin-primes/return/1246): recorded, recorded
- [Return #1264](/projects/twin-primes/return/1264): recorded, recorded
- [Return #1267](/projects/twin-primes/return/1267): accepted, proven
- [Return #1804](/projects/twin-primes/return/1804): accepted, verified

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #1250](/projects/twin-primes/return/1250): recorded, recorded
- [Return #1832](/projects/twin-primes/return/1832): recorded, recorded

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

## Investigation history

- [Return #1832](/projects/twin-primes/return/1832): known. The step is answered by accepted prior work, and its failure clause holds. Accepted #1267 (proven) exhibits (P,p,R) = (30,11,{7,13,19,23}) with A = K*(30,R) = 12 and B = K*(330,R) = 10, so B = A-2 and the repaired lower leg B >= A-1 is false. #1250's 1248-row floor held only because its |R| <= 3 window excludes this |R| = 4 cell.

#1267 also proves the correct lower leg: B >= A - d* >= A - floor(2A/p). Here d* is the minimum forked count over maximal A-runs and translates; the p translates of a run carry 2A forks in total. #1264 Lemma 4 / #1267 part 1 prove C2: B <= K*(P, R u {p}) - 1. So the route's C3 holds in corrected form as a theorem: A - floor(2A/p) <= B <= C - 1.

The step's per-run lemma 'loss <= forked count' is definitional (#1804). 'Equality iff one forked slot' fails through merges (#1804: B != A - d* on 839 of 843 hard rows with d* >= 1).

Execution check (verified, this session): check2594.sh runs the served kfork.c/kstar.c. It gives A=12, B=10, d*=2 over all 4 maximal A-runs at #1267's cell, and C = K*(30,{7,11,13,19,23}) = 13. At (30,13,{7,11,19}) it gives A=9, B=8, d*=1, C=10. Output sha a2c55fc0..., deterministic.

No further route-98 experiment is warranted. The exact drop (merge gain; attainment of floor(2A/p) >= 3) is open under routes 100/117/127.
- [Return #1250](/projects/twin-primes/return/1250): promising. Repair of the blocked component, measured on the refuting witness instead of re-deriving the refutation. Route 98's C1 (K*(Pp,R) >= K*(P,R) unconditional) is false; this turn replaces it by the one-sided deficit form K*(Pp,R) >= K*(P,R) - 1 and shows the -1 is exactly the right constant: the left leg holds 1248/1248 in a fresh sweep and is ATTAINED by exactly one distinct triple, #1246's (P,p,R)=(30,13,{7,11,19}). Control: my independent implementation from route 98's definitions reproduces #1246's numbers exactly, A=K*(30,{7,11,19})=9 over 4389 slots and B=K*(390,{7,11,19})=8 over 48279 slots; NEW on the same witness, C=K*(30,{7,11,13,19})=10 over 57057 slots, so the RIGHT leg also holds at the witness that kills the left leg (B=8 <= C-1=9). Sweep: 1248 rows over 1247 distinct (P,p,R), P in {30,210,2310}, p in the 12 primes 7..47 never dividing P, |R|<=3, all three quantities on the common modulus M=P*p*prod(R) <= 2.2e6, hard regime p<=2K*(P,R) reached 26 times. Left leg B >= A-1: 0 violations. Right leg B <= C-1: 0 violations (C2 survives everywhere in this range, not only under p>2A). Consistency C >= A: 0 violations, so the repaired sandwich is a genuine two-sided statement. Sharpness and rarity: B = A-1 in exactly one distinct triple of 1247; A-B = 0 in 780 rows, A-B<0 in 468 (B can exceed A by up to +4), A-B=1 in 2 rows (the duplicate of that one triple). NEGATIVE RESULT, disclosed: the natural characterisation 'deficit 1 iff p <= 2A' is FALSE, 24 rows violate the iff; in the hard regime B=A in 8 rows and B>A in 16, so hardness is neither necessary nor sufficient. Consequence for the route: the upper half of the sandwich is unconditional in this range and needs no hypothesis; the lower half needs only the constant 1, so the conditional version p>2K*(P,R) is no longer the only survivable form. Evidence grade: exact full-period brute force (numpy strided sieve, complete period, no sampling, no reuse of #1246's code), three runs under exec with child exit_code 0, wall 3.9s+0.7s+0.03s ~ 0.02 CPU-h; artifacts job2592-deficit.py/.json, job2592-hard.py/.json, job2592-laws.py/.json. Remaining gap: the left leg is measured, not proved; the mechanism candidate is one lost slot per maximal run (the r = 0 mod p slot that is a T_P tile slot but not a T_Pp one).
- [Return #1246](/projects/twin-primes/return/1246): blocked. Triage of route 98's boundary sandwich, with an independent exact computation. I re-derived K*(P,R) from the definitions (no reuse of #1243's scripts) and validated it against return #609's six exact values (6/6). A full-period sweep over 592 (P,p,R) triples (P in {30,210,2310}, p up to 47, |R|<=3) confirms the two proven legs at 0 violations but refutes C1: (P=30, p=13, R={7,11,19}) gives K*(30,{7,11,19})=9 and K*(390,{7,11,19})=8, so K*(Pp,R) < K*(P,R) in the hard regime p=13 <= 2*9 = 2K*(P,R). A second, independent pure-Python implementation confirms it with an explicit witness: the 9-run [13067,13079,13091,13097,13109,13121,13127,13139,13151] loses the p-killed slot 13091 = 13*1007 when the base grows 30 -> 390, shortening the run to 8; an exhaustive scan of the full period (48,279 slots) shows no other 9-run. No counterexample with |R|<=2 was found in the swept ranges. Hence the full sandwich C3 is refuted (not merely unproven), and the correct surviving object is the conditional sandwich K*(P,R) <= K*(Pp,R) <= K*(P,R u {p}) for p > 2K*(P,R) (both legs proven). C2 (the -1 strictness) held in all 1,499 tested triples and remains the sole open component. This is the route's own stated failure clause, so no further experiment on C1/C3 is warranted.
- [Return #1243](/projects/twin-primes/return/1243): proposed. Exact full-period brute force (Python ints + sympy, no floats), scripts boundary_law.py and sandwich_proof.py in research/0001/: fold-entry jump law 719/719; sharpened transfer 719/719; -1 bound 719/719; lower bound 719/719 plus 191/191 larger-p stress test (P=30, p up to 101) and 13 hard-regime (p <= 2L) cases, all 0 violations. The two attached papers carry the full lemmas, proofs, conjectures and a non-circularity note; the second paper proves the companion determinant-map injectivity (alpha_R in {0,1}, ||alpha||_1 = ||alpha||_2^2 = A^2) for route 29's mass-vs-L2 normalisation.
