Research route #22 · investment progress is separate from claim verification.
Job #1286 — Triage of route #22 (exact U-set LP certificates on the zone)
Attempt [REDACTED] · type explore · purpose discovery · lane none.
Author rung of this return: heuristic (an investment decision on a proposal, not a claim).
Decision
Promising — one bounded next experiment is justified, and only for a narrow purpose.
The route's data claim is already on the record and was not re-derived here: at 11 levels
p = 23..67 the level-X LP certificate stays in 17–25 while zone twins grow 29 → 121, levels
≤ X^0.75 certify 0 from p = 31, and levels X^1.25 / X^1.5 recover 85% / 100%
(uset_level_lp2.out sha256 662c2200…, uset_level_lp.out sha256 1b905ac0…, return #559).
What triage had to decide is whether one more measurement is worth its cost. It is, because
the 17–25 column is flat over a factor-~3 in p and the sign of its trend is the whole question;
and it is bounded (0.5 CPU-h, 4 GB, 1 GB disk, ~1 h of agent time).
What this changes
- The LP framing is prior art, cleanly. The closest located source is the MathOverflow
question “Best possible sieves for the jacobsthal problem, linear programming, and the prime
2” (asked 2011-06-16, 1k views), where the “linearized Jacobsthal function”
j_lin(n)is defined exactly as a sieve LP over real weights indexed by the moduliD | n, with the explicit dualmin Σ|c_d| / max(Σ c_d/d, 0), and is computed to n = 75. So neither “sieve as an LP”, nor “LP over modulus-indexed data”, nor “the dual is a weighted-discrepancy quotient” is new. This strengthens the route rather than weakening it: the object is standard enough to be optimisable with off-the-shelf LP, and the thread even records the working computational observation that optimal solutions carry only polynomially many nonzero variables at n = 75. That observation is the best available support for the proposed restricted-atom relaxation, which is otherwise justified only by the arithmetic argumentstrikers | n(n+2) < X². - The parity caution does not transfer, and that is now sourced. The parity obstruction
cited in the route (Polymath8b arXiv:1407.4897 §8; Tao, 254A Notes 4, dual sieve problem) is
consistently stated for approximate discrepancy data (
q ≤ x^{1-ε}). Route #22's LP uses exact anchored counts in a single zone window, which is precisely why a positive bound at level D = X is not a contradiction. The route already says this; triage confirms the separation is the standard one and is not an artefact of the small p range. - The remaining gap is narrow and unstated in the literature. No located source asks how the LP optimum grows with the truncation level D relative to the window length X when the modulus-level data is exact and anchored. That is the route's uncovered step. Absence of a source is not established novelty — the Hailperin/Prékopa line is only partly read (see access gaps) and no LP-sieve literature sweep beyond these queries was done.
Weakest assumption, mapped onto the borrowed method
The proposed experiment restricts atoms to strike types with prod(T) ≤ X². Two distinct claims
hide there, and triage should not merge them:
- Validity — every real position's strikers divide
n(n+2) < X², so the dropped atoms are arithmetically irrelevant. This is an argument, and it is sound as stated. - Optimum preservation — dropping atoms removes constraints, so the restricted LP is a relaxation and its optimum is an upper bound on the exact optimum, not automatically equal to it. Equality is an assumption. The route already names the check (compare restricted against unrestricted where both exist, p ≤ 53); that check must run before the p = 71..151 sweep means anything, because if the two disagree at p ≤ 53 the larger-p numbers are not comparable to the recorded column.
This is the weakest link, and it is already inside the plan, so the plan is scoped correctly.
Two smaller assumptions, both addressed: HiGHS returns floating-point optima and the column is nearly flat, so the exact-rational re-solve at 3 levels is load-bearing rather than decorative; and the success threshold (≥2× from p = 67 to 151, or >15 % of zone twins) is pre-declared, so the experiment can fail cleanly.
Scope and what not to conclude
The route itself is explicit that a finite certificate cannot yield an asymptotic proof strategy and that the zone twin count is already known to 1e11 by direct enumeration. Triage agrees and adds no optimism beyond that: the value of this experiment is a measured instance of parity in U-set coordinates (does exact level-X data stay bounded/zero, or does it grow?), and nothing about infinitude. If the certificate stays at 17–25 while twins keep growing, that is a clean negative and closes the tangent; it is not a proof of anything larger. Do not close the broader route on one failed proof attempt.
Execution and evidence
All landing/infrastructure work and the prior-art retrieval for this assignment happen locally and verifiably; the recipe below reproduces the only concrete outputs this return reports (the department tool self-tests and the transcript scrub). Nothing here reproduces published numbers, per the triage instruction, and no local computation was spent on the route's own arithmetic.
Transcript. Attached, scrubbed as data (per-line JSON parse, redaction inside decoded string
values, re-serialised) from this harness's own session log for this assignment. Removed: the
account token (prefix sah_…, 9 occurrences), provider userId / userEmail on every line,
installation identifiers, absolute home-directory paths (116), and the private session id. No
per-turn token usage is attached: this harness records only contextTokenCount and
stepCreditsUsed, so usage is unknown and is omitted rather than estimated.
Caution for the person. The brief reports that 9 of this handle's own returns are queued and cannot be reviewed by a deepseek-v4-flash agent, because a model never reviews its own kind; they wait for a tier-3 agent on another model. This return therefore also stacks unreviewed. Reporting so the user can route a reviewer.
Declared premises of this result: return #559 (recorded).