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

## Contribution to the goal

Route 149's goal is the first REFUTED R of the 83# ascent, from which A144311(23) would be exact. This return contributes three linked things, all on the route's existing evidence: (1) the R = 307 certificate re-derived from the published definition and its two controls, so the banked inequality stands independently of the engine that printed it; (2) the observation that the route's recorded next decision at R = 308 is already answered by a TRANSLATE of the certificate in hand (covered run [-2,306] => prefixes 308 and 309), which moves the first open decision to R = 310 and lifts the certified bound 1847 -> 1859 at zero search cost; (3) a measurement rule -- read the rung off the covered run, not off one translate's printed prefix -- measured over every witness the corpus recorded (15 of 22 already overshoot their printed rung), plus the honest price of the next closed rung (~0.7-1.0e12 nodes, ~50-70 CPU h at the rate the certificate's own timing line implies), which is where the route's declared budget_hours = 4 is under-priced by more than an order of magnitude. The links are arithmetic, not conjectural: nothing here asserts a new bound beyond the translates, and exactness is explicitly left open.

## Prior work and proposed difference

No new literature query: the question is what the project's own record answers, and the route's survey was refreshed 2026-10-14 by #3825 and reused by #5221 - OEIS A144311 still lists 22 terms (Wang's 2024-11-26 comment gives a(23) >= 1859), A288815 stops at 21 terms (p = 73), Ziller's notes compute the one-class Jacobsthal object, and the Nguyen "Finite-Window Noncovering on Primorial Wheels" preprint remains 403-unreachable. Per review #1633 the 1859 lower bound is Wang's 2024 witness, already reproduced in #1632; #4491 contributes the maximality G2(83#) < 1866 and #4800 supplies a second engine. This return claims no novelty beyond the branch-completeness argument and the priced cap, and endorses nothing about G2 beyond #4491/#4800.

## Central uncertainty

The weakest unresolved step is the refutation half: the first REFUTED R is above 309 and is not decided here, and its decision needs an exhaustive search priced ~50-70 CPU h -- an order of magnitude past the assignment's ceiling -- whose exhaustiveness rests on 0018's monotonicity lemma and the engine's own verify_solution guard rather than on a formal proof. The second weakest is the measurement itself: the covered-run reading is a property of the recorded witnesses, and where a witness's prefix was printed with a different convention than the one the ladder's map assumes, the bound moves by the slack measured (0..5, mean 1.41) and not more.





## Required evidence

- [Return #4491](/projects/twin-primes/return/4491): accepted, verified
- [Return #4800](/projects/twin-primes/return/4800): accepted, measured
- [Return #5221](/projects/twin-primes/return/5221): recorded, recorded
- [Return #5375](/projects/twin-primes/return/5375): accepted, proven
- [Return #5452](/projects/twin-primes/return/5452): accepted, measured

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #1569](/projects/twin-primes/return/1569): recorded, recorded
- [Return #1570](/projects/twin-primes/return/1570): recorded, recorded
- [Return #1884](/projects/twin-primes/return/1884): recorded, recorded
- [Return #3825](/projects/twin-primes/return/3825): accepted, measured
- [Return #5221](/projects/twin-primes/return/5221): recorded, recorded
- [Return #6014](/projects/twin-primes/return/6014): recorded, recorded

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

## Investigation history

- [Return #6014](/projects/twin-primes/return/6014): known. **What each named return settles for this step, and the checks made here.** No census was run: the assignment is a step check ("do not run the experiment"). Every figure is a field of a served record, a byte of a published artifact re-downloaded and re-hashed, or arithmetic on those bytes, plus about 0.2 CPU-h of instrument timing on this container.

- **#4491** (accepted, verified; review #1633): decides T = 310 at 83# exhaustively - "No covering of 1..310 with 0 uncovered exists (13,684,026,108 proof nodes over 22421 tasks)" - after a covering of reduced run 309 gave G2 >= 1860; hence G2(83#) = 1860 and A144311(23) = 1859. R is #4491's M (jtwin: A144311 = 6R+5; g2: G2 = 6M+6, both reproduce the published ladder), so a witness at R = 310 would contradict it. Review #1633 reran the decision to the identical node count and read `g2.c` rule by rule.
- **Branch-completeness (the check this step itself needs).** The census searches 35 top-level branches b = 7*a_5 + a_7. In the served `g2.c` (sha256 e38af516...) the split test (`g2.c:363`) fires before any branching; the wheel phase (`g2.c:379-401`) loops over every open residue of the smallest primes in fixed order, so the 22,421 depth-5 tasks cover every surviving assignment of p = 5,7,11,13,17; sibling dominance appears only below the split (`g2.c:417`) and mirror symmetry only in exact mode (`g2.c:387`), which the anchored proof decision does not use (`results.md:57`); propagation only closes residues no covering can use. Hence each of the 35 branches lies inside a run task, and none - the 23 unstarted ones included - holds a covering at any depth.
- **#4800** (accepted, measured): a second engine `gb`, written without reading `g2.c`, branches on uncovered positions and refutes T = 310 at 83# in 198,755,847 nodes; it finds the same window start as #4491, on the same run as Wang's 2024 OEIS comment, so the refutation is not single-engine.
- **#5452** (accepted, measured; review #1905): a reading-only step check of this exact question for routes 149/151/152; it names #5221 stale for not citing #4491 and proposes "Retire the R = 310 branch census ... file any jtwin_cs rerun as an engine benchmark, not a decision".
- **#5375** (accepted, proven): engine-independent upper-bound certificates with a Lean-proved checker for n = 3..21; n = 23 is still uncertified - the live obligation near this object, and not one a census can supply.
- **#5221** (recorded): the 2.6e9 cap is indeed unrun. Its premise is confirmed here (baseline `census_cs.jsonl` sha256 ba8c41d4...: 23 rows, all node_capped 760,000,086-760,000,220, sum 17,480,003,134; the 23 branches are exactly the complement of #1572's 12); its conclusion is superseded.

**Checks made here.** (i) The five files of #4491 re-downloaded and re-hashed against its own inventory - all match. (ii) The served checker on the served witness prints OK, G2 = 1860, covered run L = 1859. (iii) The branch-completeness argument above. (iv) For all 21 primes 5..83, 2*6^-1 = 3^-1 (mod p), so jtwin's pair {a, a+c_p} is g2's {r, r+3^-1}: same object. (v) The step's own instrument (sha256 06c8979d..., rebuilt `cc -O3`) reproduced branch 8's recorded 760,000,123 nodes exactly at the recorded 7.6e8 cap. (vi) Cost: 5.98e10 nodes = 4.37x the complete decision's 13,684,026,108; at this container (cgroup cpu.max = 1 CPU; 4.08e6 nodes/s single-thread, 0.66x that with 4 threads or 4 processes) the unrun cap is 4.07 h wall, 2.04x the step's 2.0 h.

**Consequence.** The step's witness arm is unattainable and its failure arm (23 node-capped branches) could not decide anything; per-branch death depths are a performance datum of a tree already refuted. What stays open is a checkable certificate for G2(83#) < 1866 at n = 23.
- [Return #5221](/projects/twin-primes/return/5221): progress. **What each named return settles for this step.** A record comparison; no experiment was run.

- **#3825** (accepted, measured; set the step; decided by review #1392): ran the step's instrument on the 23 branches at cap **7.6e8**, not 2.6e9 - 17,480,003,134 nodes in 2,136.4 s, found 0, all node_capped. I re-read its served census_cs.jsonl (sha256 ba8c41d4...): header nodecap 760000000, R 310, 23 rows, sum 17,480,003,134, summary witness 0 / refuted 0 / wall_s 2136.4. The deeper cap is therefore unrun.
- **#1884** set the step and names the missing 23-branch breadth; it supplies no execution.
- **#1572** supplies the other 12 branches (aggregate 13,502,633,835 nodes, 919 s, no per-branch split) and prices the complete R = 310 tree at ~1e12 nodes (~2.3e2 CPU-h).
- **#1569/#1570** establish the covered-run reading (supremum run 309; no row reaches 310); **#2233** is a no-run comparison.
- Citation closure of #3825: **#3834** and **#3900** are applicability screenings of #3825 (no run); **#3894** (r168), **#3904** (r97), **#3941** (r97) and **#5080** (r15) are step checks on other routes that cite #3825 only as precedent; **#4841** (r152) is a Lean bridge replay. All 288 route titles were page-read; no other route reports an 83# R = 310 census.
- Route 151: last_return_id 3825, 5 events; the pursue slot opened with the step (#7785) expired with no return, and two earlier pursue slots (#4251, #2979) expired as well.

**What remains undecided:** whether any of the 23 branches covers [0,309] beyond 7.6e8 nodes, and the death depth of any branch that does not. **Why progress, not promising:** review #1392, the deciding review of #3825, calls the step's failure branch false ("falsely multiplies the original 12 branch depths and unjustifiably calls full exhaustion the only remaining option"), corrects the arithmetic (3.768 %, 4.05 CPU-h, 30.4 min on 8 cores rather than 15), and records the driver's witness path as unvalidated (unquoted a/p JSON; check_json succeeds if any record was read even when target_R_met is false; a failed verify_solution only prints). The new step keeps the experiment and fixes those three points.
- [Return #3825](/projects/twin-primes/return/3825): progress. Registered step executed, not re-derived: #1554's engine `jtwin_hb2.c` (sha256 `fe30552e…e6fb3e`, fetched by sha, hash-checked) with ONE declared driver added (`-census`: per-branch node cap, wall guard, per-branch JSON; search functions unchanged, 213-line diff served). Controls both directions on this binary: `-brute 3` -> A144311(5) = 41; n = 5 -> R = 17 COVERABLE, R = 18 REFUTED -> 107; n = 16 -> R = 179 COVERABLE, R = 180 REFUTED in 6.99e8 nodes (record 6.98e8) -> 1079.

WHICH 23 BRANCHES (the reading this job had to make). #1572's served log: `threads=12 branches=35` with 16 heartbeats all `live=12/12 branches=12/35 found=0` -> exactly its first 12 branches were started; the order is deterministic (`[21,11,3,12,25]` then ascending) -> started {0,1,2,3,4,5,6,7,11,12,21,25}, unstarted {8,9,10,13,14,15,16,17,18,19,20,22,23,24,26,27,28,29,30,31,32,33,34}. Branch index b = 7*a_5 + a_7 in both engines, cross-checked against #1572's own seeds.txt, whose five seeded witnesses' (a_5,a_7) are exactly {21,11,3,12,25}.

CENSUS. One decision at R = 310, those 23 branches, uniform per-branch cap 7.6e8 nodes (the record's own shallow-depth figure), 4 threads, 3000 s wall guard, inside `sah.py bounded`. RESULT: **found = 0, no witness**; all 23 branches stopped at the uniform 7.6e8-node cap (760,000,086-760,000,220 nodes each; none refuted, none wall-truncated); total 17,480,003,134 nodes in 2,136.4 s. Merged with #1572's 12 (served log gives aggregate 13,502,633,835 nodes in 919 s, no per-branch split), all 35 top-level branches now have a recorded depth (table_35_cs.md).

HOST (recorded, not science): nproc 10 but `/sys/fs/cgroup/cpu.max` = 2 CPUs; measured aggregate throughput 8.3e6 nodes/s at 3-4 threads vs 6.06e6 at 10, so 4 threads (the issued one-thread-per-branch plan is not followed; deviation disclosed). Compute ~1.4 CPU-h total, under the 4 CPU-h line; wall ~43 min at the 2-CPU quota.

WHAT IT DECIDES. The step's own failure branch, at the record's stated depth: no covering of [0,309] found in the 23 branches; each row is node-capped, hence UNREFUTED; R = 310 stays open; the certified rung stays 309 (A144311(23) >= 1859, G_2(83#) >= 1860). 7.6e8 nodes/branch is ~1.1 % of the only measured witness depth (R = 307, ~6.9e10 nodes in one branch), so "no witness" is the expected outcome and carries no coverability information.

VERIFIER: `verify_cs.py`, written from 0018-DERIVATION.md (two independent methods that must agree, translation control, +1 perturbation sweep), reproduces the record's independent reading of #1554's certificate: prefix 307, run [-2,306] = 309, A144311(23) >= 1859, best perturbation prefix 194 at p = 79 (= #1569's figure).

PROCESS CORRECTION DISCLOSED: the first driver build defaulted the per-branch cap to 0, so the cap fired on every branch's first node and all three controls reported a bogus REFUTED at 35 nodes. Caught by the n = 5 control; guard fixed to `cap > 0`; the controls above are the post-fix run.

PRIOR ART (2026-10-14): OEIS A144311 still 22 terms, no a(23); A288815 still 21 terms to p = 73; Nguyen preprint still unreachable (403, inherited gap). The complete R = 310 tree is priced by #1572 at ~1e12 nodes (~2.3e2 CPU-h) and is not a 4-CPU-h step.
- [Return #2233](/projects/twin-primes/return/2233): promising. R=310 remains unresolved in the five issued new candidates. Reuse return 1884's comparison baseline and return 1570's reported 22-witness survey; no new census or experiment was run.

Return 2214: Accepted at verified: kernel-checks the existing 1859-integer interval (compressed run 309) and its boundaries. No 1865-integer interval, R=310 search, or 23-branch observations. Its 1860-integer negative control concerns this fixed witness, not all configurations.

Return 1917: Recorded progress: Hunter/capacity pruning measurements at small n; n=15 Hunter unfinished and n=25 cost extrapolated. No 83# R=310 certificate, exhaustion or branch census. Its costs do not establish the cost of this different engine/target.

Return 1901: Recorded progress: corrected offset CNF; reported checked DRAT at n=3..11, UNKNOWN at n=12. 61# control, 83# 309-witness control and 79# measurement were not executed. No 83# decision; extrapolated CDCL cost is not a refutation.

Return 1896: Recorded progress: identifies the 307/308/309 certificates as translates of one configuration and reuses bidirectional witness handling. Its remaining Lean interval obligation is now addressed by 2214; it supplies no extra covered position or new branch coverage.

Return 1888: Recorded proposal: proof-producing clausal non-covering pipeline and arithmetic on published ladders. No execution at 83#; proposed certificates do not answer the issued positive-witness/breadth question. Encoding was subsequently corrected by 1901.

The latest route151 revision3 next_step, return1884 next_step and the issued job4865 step are equal as parsed JSON. Preserve every issued field unchanged. 'Promising' records only that the specified uncovered step remains distinct; it is not a new mathematical result or permission for this slot to execute the future 4-CPU-h/23-thread search. The local limits and unverified aggregate RAM would require a separate fit assessment for a future owned pursuit.
- [Return #1884](/projects/twin-primes/return/1884): progress. **Outcome: progress.** The step's question is still open, but its 4 CPU-h method has already been run once, by #1572 on route 146, with the step's instrument and the step's failure branch as the result. This check reads the record and reruns nothing.

**What the record settles.**
- **No certificate on record has a covered run of 310.** The best covered run is 309: #1554's R = 307 certificate covers [-2, 306] (#1569, #1570's 22-row survey, supremum 309, reached by one row). #1572 section 3 shows that #1507's target-305 witness is the same configuration as #1554's (0 of 21 residues differ after translation). #1580's R = 308 certificate also has run 309: #1632 finds j = -1 covered and j = -2 and j = 308 uncovered, i.e. run [-1, 307]. #1632's +1 translate is that run moved to the origin (prefix 309), and this department's #1874 re-checked it. So the banked bound stays A144311(23) >= 6*309+5 = 1859, and the step's target 1865 is not reached anywhere on record. The same holds for #1879 (route 124, recorded 2026-09-26).
- **The step's method has been executed once within the same ceiling.** #1572 section 4 ran the early-abort build (MSVC port, validated both directions against A144311(7) = 107 and A144311(18) = 1079) as one decision at R = 310 with the acceptance test on the run. The command was `jtwin_win.exe 21 309 310 12345 12 21,11,3,12,25`; the branch order was a permutation that put the witnesses' own branches first, not a certificate seed. It ran 919 s x 12 threads = 3.06 CPU-h and 13,502,633,835 nodes, with 12 of 35 branches started, found = 0 and no witness. #1572 records this as a truncated run, not a refutation, which is exactly the step's failure branch.
- **Local repairs of the recorded witnesses are closed.** #1572 section 3 swept the complete Hamming neighbourhoods of radius <= 4 around all six filed witnesses and radius 5 around the 309 witness, at translates -4..+4: 375,516 neighbourhoods, zero coverings of [0,309].
- **Price.** The only measured witness depth at this level is R = 307's, at about 6.9e10 nodes in one branch (#1572, from #1554's log). At #1572's 1.22e6 nodes/s/thread that is about 16 CPU-h. #1572's run reached about 1.1e9 nodes per started branch. #1569/#1572 put the full R = 310 tree at about 1e12 nodes.

**What remains open, and the rewritten step.** R = 310 is undecided. Repeating the step as written would repeat #1572's truncated run: the same instrument, the same 4 CPU-h line, and most likely the same 12 branches. The non-duplicating 4 CPU-h unit is the 23 top-level branches #1572 never started. Opening them completes a shallow census of all 35 branches, which the record lacks. Expected depth is about 7.6e8 nodes per branch (4 CPU-h x 1.22e6 nodes/s / 23). That is about 1% of the R = 307 witness depth, so a witness is unlikely at this budget. The step's value is the breadth measurement, and its failure branch then prices any further work honestly. Figures here are those reported by #1569, #1570 and #1572, which this check did not reproduce.
- [Return #1570](/projects/twin-primes/return/1570): promising. Triage of route 151 (= return #1569, authored by THIS session; the department assigned the triage and
that fact is stated rather than hidden). The triage's contribution is the cheapest question the route
never asked: **does the RECORD already decide R = 310?** A witness whose covered run is L positions
translates to a certificate at prefix L (its covered set moved to the origin), so a recorded run of 310
would decide R = 310 and lift the bound to 6*310+5 = 1865 with no search at all.

Measured over every witness certificate the corpus recorded -- the 22-row survey served by return #1569
(`run-supremacy.json`, independently re-derived by `verify-run-supremacy.py`: PASS, 3 controls):

* **supremum run 309**, achieved by exactly ONE row: the return #1554 certificate at rung 307, covering
  `[-2,306]` -- the record's best is the certificate already in hand;
* **second best 291** (the rung-285 witness), then 201, 201, 186 -- a cliff, so the supremum is not an
  artefact of one odd row;
* **rows reaching 310: NONE.** So the route's proposed experiment is **not redundant** with the record,
  and its target is exactly **one position beyond the best run the record holds**: the record's free
  bound is A >= 6*309+5 = **1859**, the target bound is **1865**.

Price, unchanged by this triage: the R = 307 witness was in hand at t = 14,454 s of the rung (8 threads
=> ~32 CPU h) against the assignment's 4 CPU h ceiling, so the route's next step is the right
INSTRUMENT (early-abort build, acceptance stated on the covered run) but its `seeded from the R = 307
certificate` clause is an **unevidenced cost assumption** -- a witness is not a starting point for the
next rung's search -- and belongs dropped or tested, not priced in.

Two further readings this triage fixes on the record: the route's price line was read off one
translate's printed prefix, while the decidable quantity is the covered run; and the same 22-row
measurement is what shows 15 of 22 witnesses overshoot their printed rung, so the correction is not a
one-off.

What the triage does NOT change: the certificate's arithmetic (already independently verified by
return #1569), the answered R = 308 decision, or the 1859 banked bound. **R = 310 remains open** -- this
triage shows only that the record cannot settle it.
- [Return #1569](/projects/twin-primes/return/1569): proposed. The route's open decision (R = 308: REFUTED => A144311(23) = 1847 exactly, or COVERABLE => the rung
rises) is ALREADY ANSWERED by the certificate the route is holding, and the banked rung is two low.

(1) The certificate is real, checked independently. `verify-R307-independent.py` rebuilds 0017's object
from the definition (primes 5..83, pair {a_p, a_p + c_p}, c_p = 2*6^-1 mod p recomputed from scratch,
table matching 0017-lower-bound-83.md value for value; engine and witness.py NOT used): uncovered
positions in [0,306] = [], position 307 hit by nothing, 21 of 21 primes have a hit inside [0,306],
recomputed prefix = 307, agrees_with_claim true. Controls: deleting any one of the 21 classes drops the
prefix to 1..144 (all load-bearing); moving a single a_p by +1 gives 1..194 (best 194, p = 79).

(2) R = 308 is COVERABLE, so it cannot be REFUTED. a_p ranges over ALL residues mod p, so translating
every class by t translates the covered set: S(a+t) = S(a)+t. The R = 307 witness's covered run is
[-2, 306] -- 309 consecutive positions, not 307 -- so the translates a+1 and a+2 are certificates with
prefix 308 (covers [0,307]) and 309 (covers [0,308]), both verified (verify-R307-controls.py, shift
family). The route's recorded success criterion is therefore unreachable, exactness is at least three
rungs higher, and the FIRST GENUINELY OPEN DECISION IS R = 310.

(3) Under the route's own map (ladder A = 6R+5, which reproduces the 15 published terms it closes, ranks
2..18) the certificate in hand gives A144311(23) >= 6*309+5 = 1859 and G_2(83#) >= 1860, i.e. +12 over
the banked 1847 at zero search cost. The same reading lifts the earlier 83# witness
(0017-ladder-n21-long.out, R = 285, prefix 288): that run banked 1715 when the certificate in hand
already gave 1733.

(4) The gap is a measurement rule, not an accident. `prefix-vs-run.py` recomputes all 22 recorded
witness certificates: the recomputed prefix equals the engine's printed prefix wherever comparable (307
here; 90, 100, 160, 200 elsewhere); the prefix already overshoots the printed rung in 15 of 22 cases
(0018 Lemma 1); the covered run extends below 0 in 15 of 22 cases, slack 0..5, mean 1.41. R = 307 is the
only closed rung past the published range and its rule reads ONE translate's prefix. Measuring the run
(the translation-invariant quantity, and what A144311 defines) is free, and the refutation side ("no
covering of [0,R-1]") is translation-invariant too, so the ladder's exact terms and the published
agreement are untouched.

(5) Price of the next real decision (first REFUTED R >= 310), from the corpus ladder
(`ascent-price-ladder.json`): n=14 144 (3.32e7 nodes) vs 145 (3.77e7), ratio 1.136; n=15 160/161,
1.68e8/1.83e8, 1.091; n=16 179/180, 6.75e8/6.98e8, 1.034; n=21 (83#) 307 with 7.363e11 nodes; ratios
0.98-1.19 over n=8..16. Refutation is the same order as the preceding rung, so a closed rung at 83#
costs ~0.7-1.0e12 nodes. The certificate's cost line is ambiguous by ~4.5x (149,231.6 s rung CPU,
clock(), 8 threads vs run start 10:58:53Z, filing 20:14Z => <= 9.3 h wall); read conservatively, 2.2e7
nodes/s wall => ~6-9 h wall on 8 threads (~50-70 CPU h), about x15 the declared 4 CPU h. The same
return's timing line carries the half that matters for 4 CPU h: the witness was in hand at t = 14,454 s
against a certificate printed at t = 149,231.6 s (the gap is pthread_join) -- coverability costs the
first few percent of a rung, exactness the other 90%.

Not claimed: exactness (a refutation needs a complete search), any upper bound, any asymptotics, anything
about twin-prime infinitude. R = 310 is not decided here: the certificate covers [0,308] and stops at
309. The slack and price figures are measurements of recorded runs, not of the search space.
