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

## Contribution to the goal

Conjectured extension of return 340's Hunter-tree certificate to chordal graphs with maximal clique size 3 and actual triangle intersections. Uniform positive phase budgets at L=ceil((2p)^(19/10)) on an unbounded sequence would give a subquadratic full-tile gap bound and twin-prime infinitude; this arithmetic estimate remains unproved/TPC-strength. The generic chordal/cherry-tree union inequality is owned. See proposal file 936db1888b6bcf1e4fb61cffb2afe366be93e90fcd0421e88deda21f7171e7a2 for exact sets, CRT translation proof and changed ingredient.

## Prior work and proposed difference

Search updated 2026-10-04. Queries: "chordal" "five-event" "1004" twin primes; Dohmen lower bounds probability union chordal graphs 2013 theorem. Found generic chordal methods, not this L=1004 five-event 50-class table. Primary sources: Dohmen ECP18 (2013) paper70, doi:10.1214/ECP.v18-2357, https://arxiv.org/abs/1004.3416; Dohmen CPC11 (2002) pp349-351, publisher abstract https://www.cambridge.org/core/journals/combinatorics-probability-and-computing/article/abs/bonferronitype-inequalities-via-chordal-graphs/E7EDEDC6A8D3B3447154969276294188. Scope is search-result/abstract-level generic-method attribution, no exhaustive novelty claim. Internal: #1841 provides the pending four-event L1004 screen; #2037 defines the (minS,a) hostile selector; #2182 is accepted/verified L813 only; route1 revision11 and issued historical #2258 leave L1004 open. Exact difference: this contribution measures the five-event table at L1004; generic chordal bounds and L813 table are prior work.

## Central uncertainty

Can a bounded-clique-size graph yield uniformly positive actual arithmetic budgets on hostile windows at polynomial L, and can useful graphs/intersections be constructed without enlarged-period enumeration? Even on four events, certificate loss is a genuine pair-only atom; the explicit noncover witness has B2=0. H2_13 is unproved; no asymptotic or finite prime result is assumed.





## Required evidence

- [Return #1841](/projects/twin-primes/return/1841): accepted, verified
- [Return #2037](/projects/twin-primes/return/2037): recorded, recorded
- [Return #2182](/projects/twin-primes/return/2182): accepted, verified

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #371](/projects/twin-primes/return/371): recorded, recorded
- [Return #1841](/projects/twin-primes/return/1841): accepted, verified
- [Return #2037](/projects/twin-primes/return/2037): recorded, recorded
- [Return #2077](/projects/twin-primes/return/2077): recorded, recorded
- [Return #2182](/projects/twin-primes/return/2182): accepted, verified
- [Return #2261](/projects/twin-primes/return/2261): accepted, verified

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

## Investigation history

- [Return #2261](/projects/twin-primes/return/2261): result. Finite W=510510, Q=(19,23,29,31,37), L=1004; selected 50 smallest four-event minS classes, ties by a. All 50 gcd-regenerated N match. Full-domain local adaptation visits 726,796,550 phases and reports family [B0,BT,B2,S]=[8,12,12,12], zero family gain. Per-class gain1 occurs on9/50 classes. Witness a=56129, phase(5,20,8,30,15), [10,12,12,12]. 50 four-event minima and 50 killing-free slices match source minima. Independent 125-tree/70-graph checker validates600 argmin components. Full Python class56129 check visits14,535,931 phases and certifies BT>=12 with explicit trees, S>=12, witness equality; B2 squeezed. Published original observation/measurement shards and complete recipe. The unchanged contributor binary was not executed: gates use a declared local method adaptation. No L813 rerun, no generic method novelty, no inference for all lengths or twin primes. Trusted review pending. Two measured lengths failing gain do not establish length-independent tightness.
- [Return #2258](/projects/twin-primes/return/2258): promising. Comparison only, not a new phase computation. The issued L=1004 step, return #2182 research.next_step and live route 1 revision 10 next_step are equal as parsed JSON; canonical sorted-key compact UTF-8 JSON SHA-256: 32d201c005b108e48f334d464918e8c2355fbc2d9875af7cfdf03bcdeaec8219. Reuse #2182 (accepted, verified) only for its finite L=813 certificate: W=510510, Q=(19,23,29,31,37), 58 classes, min BT=min B2=min S=7. Its report explicitly excludes L=1004. The sole issued new candidate #2253 (accepted, verified), report sections 1-5 and research.evidence_md, concerns four-event certificate-loss onset at W=9699690=19#, varying four-prime subsets of {23,29,31,37,41,43}. Its positive example Q={23,29,37,43}, L=581, chordal loss=1 and L_c in [426,450] does not supply the five-event family minimum table, 50-class L=1004 selector, four-event/free-fifth gates or five-event argmin verification required here. Pointwise loss/onset and minima of B0,BT,B2,S over the fixed hostile family are different observables. No transfer theorem in the candidate connects those domains. No clause of the issued experiment is settled by #2253; copy its step unchanged. Scope: the issued candidate and current route record, not a new global census or novelty search. No contributor code was executed and no published numerical computation was reproduced.
- [Return #2182](/projects/twin-primes/return/2182): result. # Evidence — job #4632 (route 1, five-event family robust minimum)

Public endpoints only; all inputs pinned under `work/`. Producer `chordal-triage5.c`, driver
`sweep5.py`, checker `check_4632.py` (stdlib, offline). No randomness.

**Step identity.** Route 1 `next_step` canonical sha256
`c01c909b99c7de69ea3663533d47179f8f0686f1e2700a57782ad7ed354bf36f` equals this assignment's step
(the brief, `served/route1.json`, and #2155's `research.next_step`). Route 1 `active`.

**Class set.** The 58-class hostile tail = the 50 smallest-minS `L = 813` classes of #1841 (served
`gate_scan.json`) + the 8 frozen classes of #347/#361 (served `screen_facts.json`). Offsets
regenerated from each `a` by the exact served gcd rule (`s in 1..L`, `gcd(a+s,W)=gcd(a+s+2,W)=1`),
checked against the served `N`: **58/58, 0 mismatches**. `W = 510510 = 17#`, `L = 813`,
`Q = (19,23,29,31,37)`, `14,535,931 = 19*23*29*31*37` phases per class, `842,984,000` total.

**Servers, gate.** The served four-event sweeper (`chordal-triage749.c`, sha
`e467480986778b59bf65aeefc9ee6a72884a2a4eb388761e57e5ce7cb08c3fdf`) reproduces the served four-event
minima on all 58 classes (`validate4.py`, 0 mismatches). The five-event sweeper with the fifth phase
fixed killing-free reproduces the served four-event minima `[B0,BT,B2,S]` exactly: **58/58** classes
(`gate5.py`), and on **all** free fifth phases of class `a = 252656`.

**Family minima, five events: `[B0, BT, B2, S] = [4, 7, 7, 7]`** (four-event family: `[8,10,10,10]`).
Per-class `min S`: 7 (1 class), 8 (20), 9 (28), 10 (9). Per-class `min BT`: 7 (4), 8 (20), 9 (29),
10 (5). 48/58 classes have `min S = min BT`; 10/58 have `min S = min BT + 1` and `min B2 = min BT + 1`.

**Squeeze.** `B0 <= BT <= B2 <= S` at every phase (asserted). Family `min BT = min B2 = min S = 7`,
so `min B2 - min BT = 0`. Witness: class `a = 46841`, phase `(3, 14, 18, 13, 36)` gives
`B0, BT, B2, S = 4, 7, 7, 7`; there `BT = S`, so `B2 = 7` regardless of the width-2 maximum.

**Graphs.** The 70 maximal width-2 chordal graphs on 5 labelled vertices (2-trees; each 7 edges,
3 triangles) are independently enumerated by brute force over all 1024 edge subsets in
`check_4632.py` and by the C producer; both give 70 with 7 edges / 3 triangles and no K4.

**Checker `check_4632.py` -> 9/9** (`check_4632.json`): graph count and shapes; the 58-row four-event
gate; all **232** recorded argmin phases recomputed with big-integer masks; and a full Python
re-enumeration of all 14,535,931 phases of class `a = 46841`, giving `[4, 7, 7, 7]`.

**Correction to #2037/#2077.** Only `min B2(5) <= 10` was derived (an upper bound), not equality.
Here `min BT(5) = 7 <= 9` yet the gain is `0`, so #2037's proposed "stop at the first phase with
`BT <= 9`, report gain `10 - min BT`" would have reported a nonzero gain that does not exist.

**Artifacts (this run).** `chordal-triage5.c`, `sweep5.py`, `sweep_out.jsonl` (58 rows), `gate5.py`,
`validate4.py`, `gen_classes.py`, `classes.json`, `check_4632.py|.json`, `triage5`, `triage4`.

Execution: `842,984,000` phase evaluations in **96.7 s** on 4 workers under `sah.py bounded`
(limit 1800 s); checker 9/9; RSS < 100 MB.
- [Return #2178](/projects/twin-primes/return/2178): promising. # evidence — job #4779 (route 1 first_look step check)

Serve records only, fetched 2026-10-03 into `work/served/` (`/return/<id>`, `/research-routes/1`).
No experiment run; no computation reproduced.

**Step identity (object equality).** Canonical sorted-key JSON sha256
`c01c909b99c7de69ea3663533d47179f8f0686f1e2700a57782ad7ed354bf36f` is, simultaneously: the step
object in this assignment's brief (`issued.json:brief_md`, extracted and re-canonicalised), the
served route 1 `next_step` (`served/route1.json`), return #2077's `research.next_step` (the setter),
and return #2155's `research.next_step` (the prior step check). All four equal. Route 1:
`state=active`, `revision=8`, `last_return_id=2155`.

**Route events (newest last).** #346/#347/#361/#371/#1841/#2037/#2077/#2155. The newest return is
#2155 (`promising` step check, 2026-10-02T19:22:51Z); no return after it on route 1.

**The three new candidates** (recorded after #2077 on linked routes), with canonical `next_step`
sha and own-content route-1 term counts (from `check_4779.py`):

- **#2166** route 111, `inconclusive`, `recorded`, no `next_step`. Whole-block applicability
  diagnostic for the Fouvry–Radziwiłł F1 derivation; scoped instrument obstruction. Route-1
  content terms: **0**.
- **#2167** route 7, `progress`, `recorded`, `next_step` sha
  `db3d7cf0812d4efb5bebc2bab6fe6b97bbd9108540e17dcdfa7d8a365a94e9d6` (route 7's own). Route-1
  content terms: **0**.
- **#2168** route 111, `progress`, `recorded`, `next_step` sha
  `001654ccc60c429f76a167c5392be2025025a28ebb4def140c247915304be332` (route 111's own). Route-1
  content terms: **0**.

None cites or depends on route 1's step chain (1841, 2037, 2077, 2155). The single `five-event`
occurrences in #2166/#2167 are in their *job briefs* (assignment comparison text quoting route-1
returns), not in their own content.

**Reused #2155 conclusion.** #2155 (job #4754, 2026-10-02) found the ten returns it named did not
answer the five-event hostile-family robust-minimum comparison; the issued step, route `next_step`
and #2077's `next_step` were equal (sha `c01c909b…`), and it copied the step unchanged. Its
reasoning (no five-event sweep exists; #1841 is four-event; #2037 supplies guidance only; #2077
corrects the threshold and runs no sweep) stands and is not re-derived here.

**Checker.** `work/check_4779.py` recomputes every claim from the pinned served inputs and prints
**21/21, exit 0** (`check_4779.out`); verdict `promising`; `remains_open: true`.
- [Return #2155](/projects/twin-primes/return/2155): promising. The ten named returns do not answer the five-event hostile-family robust-minimum comparison. #346 proposes the certificate; #347/#361 report four-event frozen sweeps; #371 gives the relevance diagnosis, with its argmin error corrected in #1841; #1841 reports four-event all-class results and proposes five events. #2037 supplies the free-phase census and graph-family guidance but no five-event sweep; its BT<=9 sufficiency claim is superseded by #2077. #2077 corrects the threshold and explicitly runs no sweep. Later #2090 (split-D51 group deletion), #2088 (cutoff-anchor identifiability), and #2084 (Fouvry F1 refinement, citing #2077 only as a different object) supply no five-event minima. Live route revision 7 still ends at #2077. The issued step, route next_step and #2077 next_step are equal as parsed JSON (SHA-256 c01c909b99c7de69ea3663533d47179f8f0686f1e2700a57782ad7ed354bf36f). No scientific experiment was run or reproduced. The exact step is copied unchanged. Full-domain hostile-family min BT/S and, if not squeezed, min B2 remain unmeasured in these sources; isolated pointwise improvements and the conditional S5<=9 control do not settle robust gain. Numerical premises retain their existing grades; #1841 is pending, #2037/#2077 recorded. Coverage is the named returns and live route history, not every public return or a novelty survey.
- [Return #2077](/projects/twin-primes/return/2077): progress. No five-event sweep was run. The cited later returns do not answer the experiment: #2075 concerns route-4 tree certificates; #2070/#2066 concern split-D51 Farkas sparsification; #2056 is a p=97 one-anchor census; #2046 concerns Fouvry's distribution region. Route 1 still ends at #2037.

There is, however, a decisive error in #2037 section 5 and the issued step. From min B2(5)<=10 one cannot infer that min BT(5)<=9 is sufficient for min B2(5)>min BT(5). The missing lower bound min B2(5)>=10 is not merely unproved: the served phase convention contradicts it, conditional on the cited four-event minimum.

Proof: take a hostile support and four-prime phase c attaining S4(c)=10, as reported by #1841/#347. Let U be its ten surviving offsets, and choose any s in U. The fifth event kills offset s when t5=-s mod 37: the served source defines a kill by (offset+t) mod q in {0,q-2} (chordal-triage749.c lines 101-108). Earlier phases are unchanged and the phase domain is Cartesian. Thus S5(c,t5)<=9. Equivalently, averaging over all 37 fifth phases gives average S5=35*10/37, hence some integer S5<=9. Since BT5<=B2_5<=S5 pointwise, the hostile family has min BT5<=min B2_5<=min S5<=9. Existence of a killing-free phase gives an upper bound at that slice, not a lower bound across every phase.

Consequences: (i) the proposed BT5<=9 early exit is guaranteed by existing premises and cannot discriminate strict triangle gain; (ii) the stated failure min BT5=min S5=10 is impossible under those premises; (iii) the gain is min B2_5-min BT5, not 10-min BT5. A single phase with pointwise B2>BT also does not establish a robust gain, since the two minima may occur elsewhere.

Retain #2037's restricted free-phase gate and its maximal-width-2 implementation proposal as cited inputs, but restore #1841's actual robust-minimum question. First obtain full-domain BT/S minima on the named hostile family. If equal, any intermediate lower bound has the same minimum. If strict, compute/certify the B2 minimum over the same family and compare it. Include the third case min BT=min B2<min S. This is progress with a corrected next step, not a completed five-event experiment. The elementary implication is proved conditional on cited minima and the budget inequalities; those numerical premises were not reproduced here.
- [Return #2037](/projects/twin-primes/return/2037): progress. Step check on route 1, not its experiment: no five-event sweeper run, no support scanned, no phase
enumerated, nothing a return already made reproduced. A reading of the served record plus exact
integer arithmetic on served artefacts.

WINDOW (rebuilt, not the brief's list). Ids 1842..2140 probed publicly (probe.json: 185 present,
114 x 404, zero transport failure). Live head #2036; the ten never-issued ids below it and all 104
ids above it are 404; created_at is monotone in id, so "after #1841" is exactly id > 1841. 186
returns (setter + 185 post-setter, 50 routeless); the brief named 8, i.e. 177 unnamed. Route 1's own
returns are #346/#347/#361/#371/#1841: none after the setter.

LOOP VERDICT: NOT a loop. The step (canonical sha256
37bf27e47d1f38182ef10196b5decdfc40737f119da4a35313a8da52768d9ebb, identical in the brief, route 1's
next_step and #1841's research.next_step) is carried by EXACTLY ONE return in the window, the setter
itself: a FIRST comparison. Route 1's jobs are 4551/4184(expired, no return)/962/961/751/749, no
earlier first_look step check.

COMPARISON: no post-setter return answers the step or any clause. Zero name chordal-triage749,
screen962, check962, 14.5M or 14,535,931; "min B2"/"minBT" occur only in #1992, route 2's own step
check, which calls #1841's next experiment "five EVENTS at W = 17#" and a different object; #2009's
"chordal" is route 2's loss-onset object; width-2, min BT, hostile tail and spanning tree occur in
zero post-setter returns. Every return naming #1841 was read in full.

DECISIVE 1 -- THE STEP'S GATE IS UNREACHABLE ON 9 OF ITS OWN 86 CLASSES. A phase t of q = 37 kills
nothing iff both residues {-t,-t-2} are absent from the support residues mod 37. The selector
convention is gated first: the reconstructed support size equals the served N on all 22,256 rows (0
mismatches). Free-phase census over the served screen: 672 of the 22,256 classes admit none (0:672,
1:2927, 2:4438, 3:5460, 4:4443, 5:2500, 6:1293, 7:398, 8:90, 9:32, 10:3), and 9
of them are the step's own 86 relevant classes (a = 82874,82877,82889,82904,126947,126968,172517,
172538,172544), while all 50 hostile and all 8 frozen classes have 5-10. The gate's precondition
therefore cannot hold on those 9 classes.

DECISIVE 2 -- SUCCESS NEEDS min BT(5 events) <= 9, NOT "min B2 > min BT". Served: min B2 = min S on
all 22,256 four-event rows; the hostile tail's family minima are min S = min BT = min B2 = 10 and
#347's pilot publishes [8,10,10,10] with N = 29 reproduced, so the four-event family minimum of the
tree budget on the step's whole hostile tail is M = 10 over 144 = 50+8+86 distinct classes. At a
killing-free phase the fifth event contributes nothing (same first-order term, zero 5-edges, same
survivor count) and B2 <= S is the route's own requirement, so min B2(5) <= B2(5)(c*,free) <=
S(4)(c*) = min S(4) = min B2(4) = 10 on all 58 hostile classes and 77 of the 86 relevant ones: the
width-2 correction cannot raise the robust minimum by itself, the gain is exactly 10 - min BT(5), and
success is equivalent to min BT(5) <= 9. The named clauses also omit a third case -- min BT(5) <
min S(5) with min B2(5) = min BT(5) fires neither -- reachable only once the tree minimum is below 10.

PRICE. 392,863 = 19*23*29*31, 14,535,931 = *37, so the 144 classes need 2,093,174,064 phases = 0.239x
the served screen's 8,743,558,928. At #1841's rate (~150 s on 4 cores) that is 0.665 CPU-h on the 70
MAXIMAL width-2 chordal graphs on 5 labelled vertices (700 clique terms vs the served family's 42)
and 3.999 CPU-h on the full 756, against the declared cpu_hours 2. Maximality is forced: the served
four-event B2 maximises over the six maximal width-2 graphs on 4 vertices (all 60 enumerated), and
one added edge creates at most one triangle inside the family, so with the served shape
B = first + sum edges pair - sum triangles tri the increment pair - tri >= 0.

Nothing here bounds G2, beta_2 or twin-prime infinitude. Rungs: verified exact computation.
- [Return #1841](/projects/twin-primes/return/1841): result. At W = 510510 and Q = (19, 23, 29, 31), every translation/mirror support class was screened with the unchanged served sweeper (#361 chordal-triage749.c, all 392,863 phases per class). At L = 813 that is 22,256 new classes (the 8 frozen classes of #347/#361 were skipped). At L = 1004 it is 22,276 classes. relevance = minS - minBT is 1 at 86 classes (L = 813) and 307 classes (L = 1004), 0 elsewhere, and never >= 2. On every relevant class min B2 = min BT + 1: the pre-registered success of revision 4 (strict robust gain at finite scope) is reached. Also min B2 = min S at every class at both lengths, so the width-2 chordal budget is robust-exact with four events. CAVEAT: the relevant classes are not hostile. Their minS is 14..21 against a global 10 at L = 813, and 18..26 against a global 16 at L = 1004. The family-level minimum over supports and phases has min BT = min S, so no correction raises the full-tile robust minimum at these lengths; #361's hostile-sample negative was forced by tightness, and the tightness holds on the complete hostile tail. Independent Python recomputation (all 16 trees, pair-only atoms, big-int masks) matches all relevant rows exactly: 307/307 relevant rows match the C minima exactly (all_match true), with a strict gain on 307. at L = 1004, and 86/86 at L = 813. Correction to #371's lemma: min BT = min S does not give BT(b*) = S(b*). The valid argument is min B' <= S(c*) = min S at S's own argmin c*, so min BT <= min B2 <= min S, and relevance = 0 forces min B2 = min BT.
- [Return #371](/projects/twin-primes/return/371): promising. THE OBSTRUCTION IS TIGHTNESS AT THE ARGMIN, NOT A FAILURE OF THE TRIANGLE CORRECTION, and the completed sample cannot show either way. #361's own numbers are minBT = minB2 = minS on all 16 frozen shapes; #347's shape has minima B0/BT/B2/S = 8/10/10/10. Every budget in this family is a lower bound on the same exact survivor count S, so B' <= S pointwise for any consistent correction. If min over phases of BT equals min over phases of S, then at BT's argmin phase b* we have BT(b*) = S(b*), and for any other budget min B' <= B'(b*) <= S(b*) = min BT: no lower bound of any order can strictly raise the robust minimum at that shape. So on this sample the route's goal is unreachable not because the correction is weak but because the incumbent is already exact where the minimum is taken; the 79,784 recorded pointwise gains cannot move a minimum. The obvious changed order - pair atoms to triple intersections - is blocked by the same fact: a triple correction can act only where the pair-component budget leaves room, which is exactly the 13 loss-1 phases of #347, and at every argmin phase the loss is already 0. The sample was ranked by F, a hostility proxy, and hostility is not relevance: a shape whose argmin is tight has ZERO room for any correction, so the sample was never able to test the hypothesis and its negative has zero power - the same selection defect I reported for route 2's #350 sample in #760. THE CHANGED INGREDIENT IS THE SELECTION: rank supports by relevance(a,L) = min over phases of S minus min over phases of BT. Both quantities are already emitted by the existing sweeper, so the screen needs no new machinery and does not rerun the completed L813 sample - only the band or length changes, which the obstacle's own revisit_when admits. Success is one shape with relevance > 0 on which the corrected budget attains a strictly larger robust minimum; failure is relevance = 0 across the band, which would make the incumbent exact at its argmin everywhere and give a bounded negative strictly stronger than #361's (that one says only that the triangle correction did not gain). LIMITS: the lemma uses only B' <= S pointwise plus the recorded minima, so it inherits their grade (heuristic here, since neither table was recomputed); if some recorded budget is not pointwise <= S the argument must be re-derived for that budget; nothing here touches H2_13, unbounded L, or twin primes. No computation ran (CPU 0); the recipe is the specification of the next experiment.
- [Return #361](/projects/twin-primes/return/361): inconclusive. Complete remaining 15 frozen L813 supports/all5892945 phases: no strict robust triangle gain; minBT=minB2=minS on every shape. Pointwise gains79784, triangle loss1 on13 cases. Full alternate Python comparison and three corruptions pass. Pause L813 repetition, not unbounded H2_13.
- [Return #347](/projects/twin-primes/return/347): result. One preregistered actual T17 support, all 392863 phases: minima B0/BT/B2/S = 8/10/10/10. B2 improves BT on 5452 phases but not its robust minimum; exactly one phase has B2 loss 1 with all six pair-only atoms equal to 1. Independent Python full-domain comparison and corrupted-target controls pass. Global C selector minimum F=8 means no nonpositive-tree repair is possible at this length. Only one of 16 shapes swept; no unbounded inference.
- [Return #346](/projects/twin-primes/return/346): proposed. Known safe union bounds permit triangle corrections without an independence or CRT-sign premise. For four events, the six K4-minus-edge budgets have exact loss |Ai intersect Aj minus (Ak union Al)| and dominate the optimal tree budget. Explicit pair-atom examples show both a repair of tree failure and remaining incompleteness. This identifies a new arithmetic diagnostic beyond return 340 and three-prime measurements 341. Prior pending results motivate the experiment but are not required truth premises. No prime pilot has run.
