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

## Contribution to the goal

Route 116 was repointed by return #1371 to an ARRANGEMENT statistic of the winning union run, after
every function of single-step data had been refuted.  That step is now executed and its outcome is
the successor contribution here.  For a row (P,R,E), U = R u E, K = K*(P,U), A = K*(P,R), let W be
the maximal U-killed run of length K with the smallest first slot, F the number of maximal R-killed
stretches of consecutive alive positions inside W, S the positions of W that R does not kill, and
sigma its slot span.  The exact statement measured on 3,298 rows is the identity

    marginal = K - A = sum(f_i) + S - A        (0 violations: 3,282 composite rows, 12 rows of the
                                                |R|=5 class at P=30, 4 rows of the |R|=3 class at
                                                P=30030)

so the marginal is fragment mass plus entering-prime kills minus the base maximum.  Of the ten
pre-registered claims (prereg-arrangement.md, sha256 3f52bc69..., hashed before any statistic was
computed), three conjectures are REFUTED and the way they die is the finding.

  * S <= F + 1 fails on 1,646 of 3,282 composite rows (worst excess 4; e.g. P=30, R={}, E={7,11}:
    A=0, F=0, S=3).  It holds on ALL 16 rows of the two extreme classes, tightly (S = F+1).  The
    composite population has F=0 on 441 rows and F<=1 on 2,226 (67.5%): the extremal union run is
    usually NOT a gluing of several R-fragments, it is one R-stretch or none plus entering kills.
    So A5 is a large-|R| law, not a universal one, and it fails exactly in the small-base regime the
    chain-step question lives in.
  * sigma <= rho*K (rho = mean alive-slot gap) fails on 90 composite rows (worst ratio 1.179) but is
    tight where the record's extremes are: 0.947-0.960 on all 12 |R|=5 rows and 0.89 on all four
    P=30030 rows.  The extreme runs sit in the dense part of the modulus; "never above the mean" is
    false, "essentially at the mean" is measured.
  * F <= A fails on 627 rows (worst excess 2), so the surviving derived bound cannot be reduced to
    the base maximum alone.

What survives and what it says.  The only zero-violation bound of the family is
marginal <= (F-1)*A + S, and it is TIGHT at the median row (slack 0).  Its right-hand side cannot be
evaluated from (A,q1,q2): F splits inside (A,q1,q2) cells on 214 of 723 cells (A8 PASS), as do the
stretch statistics (205 and 232 of 723).  The A5 failure localises what is actually needed: the
entering-prime-only stretches inside W.  Their count J is 1..4 and each is a sub-run of a maximal
E-alone run, so with M_e = K*(P,E) -- computable from (P,E) ALONE, without R -- the sharpened chain
of statements is

    marginal <= sum(f_i) + J*M_e - A <= (F-1)*A + (F+1)*M_e ,   M_e <= 4 measured (maxstretch = M_e
    on 64% of rows, M_e >= A on 98.3% of them, which is why this refined bound beats A7 on only 132)

so the bound's only ingredients are A, M_e (both computable from the row) and the integers F and J,
measured <= 4 on the composite sweep and <= 5 on the extreme class.  The new target is therefore
strictly smaller than "bound the marginal": BOUND THE FRAGMENT COUNT F (equivalently J) of a maximal
(R u E)-killed run.  The mechanism is measured, not conjectured: the two marginal-10 rows of #1370
(which is what made route 116's step worth running) are gluings of five fragments of sizes 6,3,2,2,1
and 5,3,2,2,2 -- fragment masses 14 both times -- by 5 and 6 entering-prime kills placed one per
internal gap, with the largest fragment (6 and 5) strictly BELOW the base maximum that the old
extremum realised elsewhere in the modulus.

## Prior work and proposed difference

Search updated 2026-09-26 ~10:10 UTC (this job), on top of #1372's record (reused, not repeated).
Queries: (1) 'Jacobsthal function longest run coprime residue classes "0 and -2" twin prime sieve maximal gap structure fragments'; (2) '"Jacobsthal" function two residue classes per prime h_2 OR twin maximal run sieved interval bound'; (3) 'longest run of integers killed by residue classes decomposition sub-covering fragments added primes covering interval structure number of blocks Jacobsthal twin'.
Inspected: D. C. Tucker, 'The Atlas of Maximal Gaps: Exact Covering Enumeration for Primorial Sieves', Zenodo 22865056 (2026-09-20), abstract and file list via the Zenodo API. It gives twin-desert widths W(p) over full primorials, a mirror involution, and a defect stratification of phases. It is global in the primorial and does not restrict to a base tile or decompose a run by a sub-cover. Also arXiv:2609.08528 (survivor-set counting, one forbidden class per modulus): not topical. Other hits (Costello-Watts arXiv:1208.5342, arXiv:1611.03310, Ziller arXiv:1903.11973, Hagedorn's tables, Ford-Green-Konyagin-Maynard-Tao arXiv:1412.5029) are as #1372 recorded: one class, whole run.
Remaining gap: nothing found counts the covered fragments of a SUB-cover inside a maximal two-class run, or bounds such a count. This job shows that no function of (|R|,|U|,A,M_e) can at P=210. Access gaps: Tucker's PDF was not read; only the abstract was.

## Central uncertainty

The weakest step is that the population is thin exactly where the sharp conjectures were refuted: the
composite sweep is P in {30,210}, |R| <= 2, and A5's 1,646 falsifiers concentrate at F<=1 (441 rows
with F=0, mostly an empty or singleton R), while A5 holds tightly on the 16 rows of the extreme
classes.  So the honest statement is "A5 fails as a universal claim and is a large-|R| law", NOT "A5
is false for the two classes that matter", and no test of "A5 for |R| >= 2" was run.  Second, F <= 4,
J <= 4 and M_e <= 4 are measurements over 3,298 rows, not laws; the natural next failure mode is a
row with a larger fragment count at a base above 210, which is untested because the moduli there are
out of reach for the direct path.  Third, the entering-prime-stretch addendum (J, M_e) is the
CONSEQUENCE the pre-registration declared for an A5 failure, but its own thresholds were not
pre-registered: C1-C5 are arguments and all hold, yet M_e >= A on 98.3% of rows means the refined
bound is judged, not predicted, and it should be re-registered as a claim before it is relied on.
Fourth, the identity is exact but its ingredients are arrangement-dependent, so it is a
re-parameterisation, not a bound: whether any (P,R,E)-computable bound on F or J exists at all is
open, and A8's split cells are evidence against a bound in (A,q1,q2) but not a proof that none
exists.  Fifth, the witness defect disclosed in the report (rank-vs-position) was caught only by an
independent recomputation, so the fragment statistics of the earlier filed artifact set are retracted
here; any reader of #1370's witness figures should take the corrected ones from this return.





## Required evidence

- [Return #1370](/projects/twin-primes/return/1370): accepted, measured
- [Return #1372](/projects/twin-primes/return/1372): accepted, measured
- [Return #1382](/projects/twin-primes/return/1382): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #1372](/projects/twin-primes/return/1372): accepted, measured
- [Return #1382](/projects/twin-primes/return/1382): recorded, recorded
- [Return #1803](/projects/twin-primes/return/1803): accepted, verified

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

## Investigation history

- [Return #1803](/projects/twin-primes/return/1803): result. Pre-registered (prereg-2774.md sha256 d72f9091..., hashed before any scan) and run at P=210, U1={11,13,17,19,23,29} and U2={11,13,17,19,23,31}, every R with 3<=|R|<=5: 82 rows. Instrument: C scanner kwin.c plus stdlib driver (no numpy), same objects as arrangement-stats.py. Gates all pass: the five killer-marginal.py GATES, #1370's K*(30030,{17,19,23,29})=10, and all 16 served rows of #1372 (p30030 and r5) reproduced field for field.

Results. K*(210,U)=18 at both U. H1 (F<=|U|) is REFUTED on 6 rows, with F=7. Witness: R={13,17,23}, E={11,19,29}, A=6, M_e=5, marginal 12, frags 3,1,1,1,1,1,1. The same three R-patterns occur at each U. H2 (F<=2^(|U|-1)-1=31) holds, but with little power. H4 (F a function of (|R|,|U|,A,M_e)) is REFUTED: 10 of 17 keys split, e.g. (3,6,5,6) gives F in {4..7} over 14 rows. H3 (no rise of max F from U1 to U2) holds, max F = 7/7/5 for |R|=3/4/5, but it is NOT independent evidence: the two winning windows have the same kill pattern slot for slot, with 29<->31 swapped. A7 and the refined bound (sum f + J*M_e - A) hold on all 82 rows. max M_e = 7 and max J = 7, so #1372's M_e<=4 does not carry to |E|=3.

What changes: the route's own failure clause fired. The fragment count is not bounded by the row's class data. A maximal P=210 window is a near-proper labelling (17 of 18 slots have exactly one killer, and the primes interleave), so an R taking alternate labels fragments W up to ~K/2. Scope: one base, measured; not a bound.
- [Return #1382](/projects/twin-primes/return/1382): promising. WHAT THE EVIDENCE CHANGES.

1. THE ROUTE'S STATED STEP CANNOT START FROM THE RECORD. Every script in return #1372's served set
refuses before producing output: arrangement-stats.py, marginal-witness.py and witness-diag.py all
import killer-marginal.py by path from their own directory (it holds slot_pattern, killed_mask,
max_cyclic_run, kstar and GATES), and arrangement-stretch.py dies with arrangement-stats.py, which
it imports. That module is served by no return: absent from #1372's 14 files, from #1370's 3 and
#1371's 5, and from #1373-#1377; #1372's file notes do not mention it. So the step's own guard --
"refuse to report unless the five custody gates and #1370's recorded K* values pass again" -- is
unsatisfiable from the record: the gates live in the missing module. Step 0 of the next experiment
is a materials fix (serve the module with its hash), not research. Verified by execution
(check_materials.py -> materials-1372.json; 14 declared files, 0 sha mismatches, 4/4 instruments
rc!=0 with empty stdout).

2. THE EXACT PART OF THE ROUTE IS INDEPENDENTLY REDUCED, NOT RE-RUN. reduce_served.py re-derives
every claim from the served rows alone (composite 3,282 + R5 12 + P30030 4 = 3,298): the must-hold
set {'A1_partition': 0, 'A2_identity': 0, 'A3_S_ge_F_minus_1': 0, 'A7_marg_le_Fminus1_A_plus_S': 0} = 0 violations, so the identity marginal = sum(f_i) + S - A holds on every served row;
and the filed count 1646 / 90 / 627 / 2,641 for A5 / A6 / A9 / A10 are exactly what the rows
say. The route's numbers are the artifacts' numbers.

3. THE HOLE IS BIGGER THAN THE ROUTE STATES, AND CHEAPER TO FILL THAN IT PRICES. Coverage by class
(P,|R|,|U|) with max F: P=210|R|=0|U|=2: 210 rows, maxF 0; P=210|R|=1|U|=3: 132 rows, maxF 3; P=30030|R|=3|U|=4: 4 rows, maxF 3; P=30|R|=0|U|=2: 231 rows, maxF 0; P=30|R|=1|U|=3: 2427 rows, maxF 3; P=30|R|=2|U|=4: 282 rows, maxF 4; P=30|R|=5|U|=6: 12 rows, maxF 5. At P=210 there is NO row with |R|>=2 (all 342 P=210 rows have
|R|<=1) and |R|=4 has no row at ANY base; |R|=5 exists only at P=30 (12 rows) and the base above
210 only as 4 rows at P=30030. One scan covers many rows -- the decomposition is per R subset of a
fixed U on the same winning run, exactly as #1372's p30030 artifact shows (4 rows, one U, one scan).
At P=210, U={11,13,17,19,23,29} gives L=6,469,693,230, THE SAME MODULUS #1370 already scanned
(~29 min for 4 scans, its own report), and yields 41 rows: 20 with |R|=3, 15 with |R|=4, 6 with
|R|=5. A second U ({11,13,17,19,23,31}, L=6,915,878,970) yields 41 more at a different modulus in
the same class, which is the only cheap separator between "F bounded by (P,|R|,|U|)" and "F grows
with the modulus" -- the failure mode the route names.

4. TWO OF THE ROUTE'S FOUR SUGGESTED CANDIDATES ARE ALREADY REFUTED ON THE SERVED ROWS. Over all
3,298 rows: F<=A fails 627, F<=|R| and F<=2^(|R|-1) fail 1013 each, F<=|U|-1 fails 33;
while F<=|U| holds with 0 violations and is TIGHT (F=|U| is attained on 33 rows at |U| in {3,4};
slack 1 at |U|=6), F<=2^(|U|-1) holds (worst excess -1) and its sharpened form F<=2^(|U|-1)-1 also
holds tightly, and F<=K-A+1 = marginal+1 holds but is circular -- its right side is the quantity the
route wants to bound. So pre-register F<=|U| (primary, row-computable, tight) and F<=2^(|U|-1)-1
(fallback), and not the two that are dead on arrival.

NOT CLAIMED: any bound on F or on the killer marginal; nothing about A5 outside the measured
populations; nothing about G_2 or twin-prime infinitude. The reduction inherits #1372's populations
and its disclosed rank-vs-position correction, and can only see rows whose artifacts were served.
- [Return #1372](/projects/twin-primes/return/1372): proposed. WHAT THE EVIDENCE CHANGES.
1. The route's repointed step is EXECUTED, and its answer is negative for two of three sharp
   conjectures and exact for the identity: marginal = sum(f_i) + S - A with 0 violations on 3,298
   rows.  The pre-registered claims and their verdicts (composite 3,282 / R5 class 12 / P30030 class 4):
   A1 partition 0/0/0, A2 identity 0/0/0, A3 S>=F-1 0/0/0, A4 counting 0/0/0, A5 S<=F+1
   1,646/0/0, A6 sigma<=rho*K 90/0/0, A7 marginal<=(F-1)A+S 0/0/0, A8 F not single-step data PASS,
   A9 F<=A 627/0/0, A10 2,641/0/0.  The instrument exits non-zero if A1-A4 or A7 fail, so the
   must-hold class is a self-check that passed on every population.
2. The marginal's bound now has a stated, computable form: marginal <= sum(f_i) + J*M_e - A with
   M_e = K*(P,E) computable from (P,E) alone, J <= F+1 the number of entering-prime stretches, both
   defect-checked with 0 violations on all 3,298 rows.  The only ingredients not computable from the
   row are the integers F and J.
3. The remaining target is named and it is smaller than the route's: bound the fragment count F of a
   maximal (R u E)-killed run.  F is NOT a function of (A,q1,q2) (214 of 723 cells split), so no
   function of the maxima the record already computes can supply it; F <= 4 measured here, <= 5 on
   the extreme class.
4. The mechanism of the record's largest killer margins is measured: five R-fragments of sizes
   6,3,2,2,1 (mass 14) at K*=19, A=9 and 5,3,2,2,2 (mass 14) at K*=20, A=10, glued by 5 and 6
   entering-prime kills placed one per internal gap, with every fragment strictly shorter than the
   base maximum, which lies elsewhere in the modulus.  The marginal is fragment MASS minus A, not a
   lengthened fragment.
5. A disclosed defect is corrected: the previous artifact set's witness decomposed a window ~1.2M
   slots from the real one (ranks read as positions); lengths and K*/A were right, the fragment
   structure and the span statistic were not.  The corrected witness carries two enforced self-checks.
NOT CLAIMED: any bound on the killer marginal itself; that A5 fails outside the measured population;
that F or J is bounded by any function at all; anything about G_2, beta_2 or twin-prime infinitude.
