/teal-sea
teal-sea / zeta-labstate of record · compiled 28 Sep 2026 · revision e4945c4 · source

Library · hunts/prime_pair_error/frontier/2026-09-06/certificate_lp_frontier/RUNS.md

Runs

1,206 words · 88 lines · source

All on the operator's laptop (16 GB M4) under a foreign load average of 60 to 140 from another session's agent fleet, so wall times are inflated several-fold. Memory stayed under 300 MB per job except the one noted.

scriptNywallnote
lp_frontier.py (dense primal)10^32..1000< 1 s each
lp_frontier.py10^410..20000.1 s .. 276 sy=3162 killed by me at 2.9 GB
lp_frontier.py (first dual form)10^4, 10^5anystalled > 10 minmonotone-chain dual is degenerate; abandoned
lp_allcells_cg.py10^4100, 3163 s, 242 smatches dense to all digits
lp_allcells_cg.py10^5100, 31614 s, 709 s
lp_allcells_cg.py10^51000queued
lp_allcells_cg.py10^61000, 3162queuedexpected ~10^4 excess at y=1000 by the N^{3/4} law
lp_prime_cells.py10^440..1000< 1 s each
lp_prime_cells.py10^5316..31625 s .. 17 s
lp_prime_cells.py10^61000135 sy > 1193 reports unbounded (float artifact, value is 0)
lp_dictionaries.py --lift 1510^4, 10^531, 1000.4 s .. 15 s
lp_dictionaries.py --lift 1510^51000, 3000queued
lp_dictionaries.py --selberg10^43159 sy = 100, 316 queued

Estimate before the larger runs: constraint generation converges in 3 to 6 rounds; each round is a dense (cells x y) dual simplex, cells growing to about 4y plus the violated set. At N = 10^6, y = 1000 that is roughly 10^4 x 10^3 per round, tens of seconds unloaded. Nothing here needed CI; anything at N = 10^7 or y = 10^4 should go there.

CI batch, 2026-09-07 (.github/workflows/certificate-lp-barrier.yml)

Launched after the three queued local rows were OOM-killed under a foreign load of 200+. Estimates, from the (10^5, 316) round profile (4364 cells, 3 rounds, ~150 s unloaded):

jobcells expectedper-round LProundswall estimateprediction
all-cells 10^6 / 316~5k1 min410 min0.32/sqrt(y): 18k; drift: 3.8k
all-cells 10^6 / 1000~15k5 min530 min0.32/sqrt(y): 10.1k; drift: 4.4k
all-cells 10^6 / 3162~30k30 min53 h0.32/sqrt(y): 5.7k
all-cells 10^7 / 1000~20k10 min61 h0.32/sqrt(y): 101k; drift: 44k
all-cells 10^7 / 3162~50k1 h+6may time out0.32/sqrt(y): 57k; drift: 41k
prime-cells 10^73.6kseconds15 min~0.5 sqrt(N) at y = sqrt N, 0 at N^0.53
lifted 10^5 / 1000, 3000; 10^6 / 30005k .. 30k1 .. 30 min40.3 .. 3 hthe hunt's family floor at its own seed support
selberg 10^4 / 3162.5k (x2 rows)minutes430 min~0.7 x 32.5 if the constant-factor pattern holds

Every round's value-psi is the optimum of a relaxation and therefore a rigorous lower bound on the final floor; a timed-out job still reports one. Owner: this session (watches the run and folds the numbers into RESULTS.md).

Outcome (run 34123341782): 4 rows succeeded with artifacts (prime-cells 10^7, lifted 10^5/1000, all-cells 10^6/316, all-cells 10^6/1000); 6 rows hit their time limit (GitHub reports "cancelled"), and their artifact steps did not run. Two converged results were recovered from logs: (10^6, 1000) again (10699.906, identical) and (10^7, 1000) = 111891.518 in 3210 s. Cause of the waste: lp_allcells_cg.py appended y = sqrt(N) to every job through its default --alpha 0.5, so the y = 1000 jobs went on to attempt y = 3162 after finishing. Fixed: no default alpha. The (10^7, 3162) row itself is out of reach of plain constraint generation in 350 minutes (9.2 million violated cells after round 0); it needs a warm start or a smarter initial cell set before it is worth another runner-day.

Local runs, 2026-09-07, continuation (dual witnesses; one worker, all thread env vars = 1)

script(N, y)wallmemorynote
dual_witness.py(10^3, 31)12 ssmallexact rational solves for 168 primes, 61 rows; enclosures at 60 digits
dual_witness.py(10^4, 100)37 ssmall1229 primes, 198 rows, 100 x 100 exact system
prefix_witness.py(10^3, 31), (10^4, 100)0.2 s, 14 ssmallclosed-form rates checked exactly against the basis solve
fake_mass_witness.py(10^3, 31), (10^4, 100)30 s, 100 ssmallConstructions A and B; pair scan is 17 x 30 and 35 x 98 two-variable LPs per greedy step
fold_family.py(10^3, 31), (10^4, 100), (3600, 60) holdout0.1 s eachsmallthe prescribed halving-fold family and the rough-shift family; no LP solve
compensated_fold.py(10^3, 31) with the bundle 76:1,200:2,333:1; (10^4, 100); (3600, 60)1 s, 12 s, 3 ssmallbundle verification, Lemma 8 checks, Rules C0 and C; the coordinator bundle + credit diagnosis at 10^4; no LP solve
signed_fold.py(10^4, 100), witness from results/fake_mass_N10000_y100.json2 ssmalldescending fold recurrence, reconstruction check, twin-exchange scan; no LP solve
grouped_signed.py(10^4, 100), same witness + Codex's beta3 ssmallh-decomposition identity, excess bound, confluence-triple scan; no LP solve
escalator.py(10^4, 100), source 232 -> {102, 123, 126}; review controls a=102/{204}, a=103/{204} zero weights, empty D3 ssmallcontrol J, telescoping (Lemma 11), S(a) = -z_a, source-exhaustion (Lemma 12, corrected coordinate and max({0} u ...) bound); exact moments/coordinates, point-valued mpmath gains (no enclosure); no LP solve

No CI, no background jobs, nothing at 10^5 or beyond.

Local runs, 2026-09-07 (barrier work)

script(N, y)wallnote
barrier_lemmas.py(10^3, 31), Y up to Nsecondsrestricted dual at Y = N reproduces the LP floor
barrier_lemmas.py(10^4, 100), Y up to 3000~2 min
barrier_lemmas.py(10^5, 316), Y up to 9480~5 min
barrier_lemmas.py(10^6, 1000), Y up to 10^4~15 min, ~700 MBthe rigorous 7162
barrier_lemmas.py(10^5, 30), (10^6, 60)~1 min eachstaircase vs restricted dual
lp_allcells_cg.py(10^5, 30), (10^5, 60), (10^6, 60)0.1 s, 0.2 s, 3.7 sfloors at y = N^{0.3}
lp_frontier.py (dense)(2 10^4, 141) and the elevation/variance probes~1 min