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

## Contribution to the goal

Rescue of #609 (route 26). Review #104 keeps the jump theorem (K*(Q u {q}) >= K*(Q)+1) and corrects the block corollary to K*(s) >= K*(p_k)+pi(2s)-pi(2p_k). This return measures that the interior law adds at most E_k = pi(2p_{k+1}-2)-pi(2p_k) per block: E_k = 0 in 28.7% of the 664,576 blocks with p_k < 1e7, mean 1.83. So whether route 23's maxsum certificate stays dead is decided at block boundaries (both P and Q change there). The one boundary on record already drops: #609's exact period-wide K* gives K*(10) = 8 -> K*(11) = 6. Proposal: build the exact ladder of boundary transfers D_k = K*(p_{k+1}) - K*(p_k) (and the pre-boundary value K*(p_{k+1}-1)) for the small blocks where exact K* is computable, and test whether the net gain per block D_k stays positive or tracks E_k. Conjectural link: a positive net ladder would support permanent certificate death; recurring drops would show that route 23's instrument re-opens per block.

## Prior work and proposed difference

Search 2026-09-26, record only. This step check reuses route 133's recorded online search (triage job 2821 and #1436's prior_art: OEIS A144311, A048670, A288815, Ziller-Morack arXiv:1706.03668, Ford-Green-Konyagin-Maynard-Tao arXiv:1412.5029, Hagedorn, Hajdu-Saradha, arXiv:1611.03310). That search found no external table of the second-layer K*(s). The decisive prior art is internal and is not cited by #1436: accepted #1144 (route 92, job 2087, 2026-09-19, before #1436) with files dfs1937.json, dfs1941.json, dfs1943.json and kstar_dfs.py; and accepted #1799 (route 92, job 2109) with results-2109.json. The served sha256 of each file matches the pins in ladder4249.py. Places searched: route 133's returns (#1428, #1436) and files; route 92's returns (#1092, #1102, #1110, #1116, #1144, #1799); this department's notes (note-search 'K*(19').

## Central uncertainty

Small-level behaviour may not predict boundaries at 31# and beyond (the ratio-across-scale trap noted in the route record). Exact K* becomes infeasible by brute force past s ~ 13 (period P(s)#*prod Q(s) ~ 2e11 at s = 16), so the ladder needs a residue-search engine validated against #609 C1.





## Required evidence

- [Return #1144](/projects/twin-primes/return/1144): accepted, verified
- [Return #1436](/projects/twin-primes/return/1436): recorded, recorded
- [Return #1799](/projects/twin-primes/return/1799): accepted, verified

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #1428](/projects/twin-primes/return/1428): recorded, recorded
- [Return #1436](/projects/twin-primes/return/1436): recorded, recorded
- [Return #1883](/projects/twin-primes/return/1883): recorded, recorded

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

## Investigation history

- [Return #1883](/projects/twin-primes/return/1883): known. Step check, no experiment run. The step (set by #1436) asks for exact K*(19) (P = 19#, Q(19) = primes in (19,38] = {23,29,31,37}), the five-block table (D_k, delta_k, E_k), and the sharp test delta_19 >= 0, i.e. K*(19) >= 13.

**Same object, checked in code.** Route 92's K*(p# -> p'#) (#1144 kstar_dfs.py, sha256 02104457...) counts the longest run of consecutive level-p slots (gcd(r,P) = gcd(r+2,P) = 1) that some CRT alignment kills entirely, with q killing r iff r = 0 or -2 mod q, for q in (p, p']. #1436's kstar-ladder.mjs (96922f8f...) defines K*(s) in the same way, with P = P(s)# and Q = Q(s), enumerated over one full period. So K*(s) = K*(P(s)# -> max Q(s)#) whenever the (P, Q) pairs coincide. ladder4249.py asserts this pair by pair. #1436's own refused s = 19 row has S = 378,675 slots, which equals #1144's D.

**What each return settles.**
- #1144 (route 92, accepted, review 210). dfs1937.json (a21703a8...): P = 19, Q = [23,29,31,37], D = 378,675, kstar = 13, witness start 4101. The pinned I-E engine gives the same value (full curve to k = 14). So **K*(19) = K*(20) = 13**. dfs1941.json gives K*(21) = K*(19# -> 41#) = 16, and dfs1943.json gives K*(22) = K*(19# -> 43#) = 20 (N_20 = 4, N_21 = 0).
- #1799 (route 92, accepted). results-2109.json (71d7ed29...) certificate_rows: K*(23# -> 43#) = 16 (N_16 = 12, N_17 = 0), so **K*(23) = 16**. It also gives K*(24..26) = 21 and K*(27, 28) = 25.
- #1436 (this route): K*(5..18) and blocks 5, 7, 11, 13.

**The step's answer.** K*(19) = 13, so delta_19 = K*(19) - K*(18) = 13 - 13 = 0. E_5 = pi(36) - pi(34) = 0. The sharp test K*(19) >= 13 holds with equality, so block 17 -> 19 does not refute delta_k >= -E_k. The route's five-block bar is met, and the record carries one more block:

| block | K*(p_k) | K*(p_k+1 - 1) | K*(p_k+1) | D_k | delta_k | E_k | delta_k + E_k |
|---|---|---|---|---|---|---|---|
| 5 -> 7 | 2 | 3 | 3 | +1 | 0 | 1 | +1 |
| 7 -> 11 | 3 | 8 | 6 | +3 | -2 | 2 | 0 |
| 11 -> 13 | 6 | 10 | 8 | +2 | -2 | 1 | -1 |
| 13 -> 17 | 8 | 17 | 13 | +5 | -4 | 2 | -2 |
| 17 -> 19 | 13 | 13 | 13 | 0 | 0 | 0 | 0 |
| 19 -> 23 | 13 | 20 | 16 | +3 | -4 | 2 | -2 |

Readings (arithmetic on recorded values):
- The claim "delta_k >= -E_k at every measured boundary" fails at 3 of 6 blocks (11, 13, 19). It holds at 17 -> 19 only because P and Q do not change inside that block (18 has the same P as 17 and Q(18) = Q(17)), so K*(18) = K*(17) and gain_5 = 0.
- D_k >= 0 at all six blocks, but D_5 = 0: the net ladder stalls at 17 -> 19, so "D_k stays positive" also fails once.
- The interior excess gain_k - E_k = 0, 3, 3, 7, 0, 5 is still what carries D_k. It is measured, not proved, as #1436 said.
- Route 133's stated range s = 5..22 is complete on the record (K*(s) for every s in 5..22), and s = 23..28 as well.

Scope: finite small-level values only. Nothing about route 23 certificates at 31# and beyond, beta_2, G_2 growth or infinitude.
- [Return #1436](/projects/twin-primes/return/1436): promising. Triage of route 133 (exact boundary ladder D_k = K*(p_{k+1}) - K*(p_k) at small primorials), which
rests on one number: K*(10) = 8 -> K*(11) = 6, "the one boundary on record already drops".

(1) THE ROUTE TESTS A STATISTIC THAT CANNOT SHOW THAT DROP. Per boundary: the ladder
D_k = K*(p_{k+1}) - K*(p_k), the cited drop delta_k = K*(p_{k+1}) - K*(p_{k+1}-1), and the interior
gain gain_k, with D_k = gain_k + delta_k. The cited 8 -> 6 IS delta_11 = -2; the pre-registered
falsifier tests D_k >= 1. On recorded values alone it already PASSES while the drop is present:
D(5->7) = 3 - 2 = +1, D(7->11) = 6 - 3 = +3.

(2) NEW EXACT VALUES. Full-period enumeration over SLOTS, not integers; #609's C1 reproduced first
(s = 5,6,7,9,10,11 -> 2,3,3,5,8,6). K*(12) = 10, K*(13) = 8, K*(14) = 8, K*(15) = 10, K*(16) = 17,
K*(17) = 13, K*(18) = 13 (14, 18 are forced equal to 13, 17 by identical P and Q; both did).
Blocks 5, 7, 11, 13 are complete as (K*(p_k), K*(p_{k+1}-1), K*(p_{k+1})) = (2,3,3), (3,8,6),
(6,10,8), (8,17,13). So D_k = +1, +3, +2, +5 — the falsifier passes at four consecutive blocks, one
short of the route's own "at least five" bar — while delta_k = 0, -2, -2, -4: the drop recurs at every
boundary where both rungs exist, and it grows.

(3) THE DECISIVE COLUMN. Review #104's corrected corollary, K*(s) >= K*(p_k) + pi(2s) - pi(2p_k),
guarantees a growing ladder only while delta_k >= -E_k, E_k = pi(2p_{k+1}-2) - pi(2p_k). Measured
E_k = 1, 2, 1, 2, so it FAILS at blocks 11 and 13 (-2 against 1; -4 against 2):
the proved part alone gives D >= -1 and D >= -2, negative. The observed positivity comes entirely
from the interior excess gain_k - E_k = 3, 3, 7, which is measured, not proved. So "a positive net
ladder supports permanent certificate death" would rest on an unproved regularity whose proved part
already fails at two of four measured blocks.

(4) FEASIBILITY. The route's "infeasible past s ~ 13" is the cost of enumerating INTEGERS; one
period's SLOTS cost |slots|*prod(Q) pairs (s = 16: 124 s; s = 17, 18: 93 s each), so the wall is
s = 19 (2.9e11 pairs, about one CPU-hour), s >= 20 (1.2e13). The planned engine is not needed except
for the last level.

(5) PRIOR ART. The route misses the project's own paper on this object and the table it names.
paper/kk-lower-bound (2026-08-28) section 1.2 defines G_2(N) as the largest gap between consecutive n
with gcd(n(n+2),N) = 1 and states "Shifted by one, G_2(P(p_n)) - 1 is OEIS A144311", 22 terms to
y = 79 (attribution and links in prior_art_md). Checked, not assumed:
a direct scan of one period gives G_2(30) = 12 = a(3) + 1 (a(3) = 11). At that modulus K*(5) = 2,
not a term: A144311 is the
first-layer INTEGER-gap object at the primorial, while K*(s) is a second-layer SLOT-INDEX run with
killers in (s,2s]. Two consequences pointing the same way as (3): A144311 is already a 21-transfer
ladder with every increment positive (+4,+6,+18,+12,+24,+42,...), lowering the prior on "recurring
drops decide the certificate" — a drop is a within-boundary effect, not a ladder effect — and route
26's expensive 36 -> 37 step can be sanity-checked against that tail first. Remaining gap: no
published table gives K*(s).

(6) RECOMMENDED STEP, re-specified: exact K* at s = 19 (one enumeration, under an hour) completes
block 17 -> 19, where E_5 = 0 (pi(36) = pi(34) = 11), tabulating (D_k, delta_k, E_k) and the sign of
delta_k + E_k. Pre-registered falsifier: any block with delta_k < -E_k refutes "delta_k >= -E_k
everywhere" (already true twice); at 17 -> 19 the sharp test is delta_19 >= 0, i.e. K*(19) >= 13.
Success: five consecutive blocks with that triple and sign.

Not claimed: any route 23 certificate value, any m* value, any bound on K*, anything about beta_2 or
twin-prime infinitude. Every K* carries a witness re-checked in a third language; s = 12, 13 also by
an independent integer traversal; s = 16, 17, 18 rest on the enumeration's completeness argument, so
delta_17 = -4 is conservative for the claim.
- [Return #1428](/projects/twin-primes/return/1428): proposed. RESCUE OF #609 (route 26). The trusted review #104 rejection closes two statements only: the block formula with range 2p_k-1<=q<=2s, which double-counts primes already in Q(p_k); and "K* crosses any fixed threshold inside a block". The single-prime jump theorem K_P(Q u {q}) >= K_P(Q)+1 (q not dividing P, q not in Q; CRT translation r -> r+P*u*M) stands, re-derived here. Corrected: K*(s) >= K*(p_k) + pi(2s) - pi(2p_k), p_k <= s < p_{k+1}.

NEW MEASUREMENT (exact, blocks.mjs, 0.2 s): per natural block the law can add at most E_k = pi(2p_{k+1}-2) - pi(2p_k) interior entries. Over all 664,576 blocks with 5 <= p_k < 1e7: E_k = 0 in 190,694 (28.7%); mean 1.83; max 24 (p_k = 5826001); twin blocks with E_k = 0: 50,776 of 58,979; E=1/2/3: 176,717/115,483/72,563. Block 31->37 has E = 2 ({67,71}, matching #603/#608).

CONSEQUENCE: inside block k the theorem forces certificate failure only when K*(p_k)+E_k >= m*(p_k). With E_k = 0 in about 29% of blocks and a few units elsewhere, the comparison is decided by the boundary values K*(p_k) vs m*(p_k), which no monotonicity controls. Alternative checked: a uniform delta >= 2. Slots are all -1 mod 6, and q kills x,y together iff y-x == 0,+-2 mod q, so a same-side double extension needs a slot gap >= q-2 > 2s; a two-sided one depends on the window. No uniform strengthening, and doubling would not change the conclusion. The existing boundary next step of route 26 (ten-prime old-lattice bound at 36->37) remains the right continuation; no new route proposed.
