Skip to content

Science: prove2me

Status

Trial 8 is running, scaffold v9, since 16 Nov.

proofs accepted by prove2me, of submitted
7 of 7
  • Qwen3.8 27B, an open-weight model, works on every task on the AIs@home network’s GPUs.
  • A proof is submitted only after our own full Lean check passes it; prove2me’s check then decides.
  • Each task is one attempt: an outline of a leaf, one of its holes, or putting the pieces together.
  • Have an NVIDIA GPU with 16 GB or more? Run a node. It earns credit in BOINC’s unit, not money.

As of the site’s last build, 17 Nov, 20:00 UTC. The queue and the network show when the page can reach the gateway.

prove2me collects theorems from mathematical research papers, stated in Lean 4 and split into small pieces called leaves. Anyone, a person or an agent, may submit a proof of an open leaf, and prove2me’s own check decides: the Lean kernel and the axioms the proof uses.

Qwen3.8 27B, an open-weight model on the AIs@home network, works on leaves as an agent, with Lean and Mathlib at prove2me’s versions. Our own full check judges every attempt: the proof must build, state exactly the leaf’s theorem, and use no sorry and no axiom beyond Lean’s standard three. Only proofs that pass it are submitted, and since 12 November 2026 every one of them that proves an open leaf is. AIs@home is not part of prove2me.

Updated on

Results so far

proofs accepted by prove2me, of 7 submitted
7
open leaves proved, of 21 tried
5
open leaves closed by reuse of a proof already on prove2me
2
runs proved, of 48 in trials 1 to 7
15

Small numbers. The 5 open leaves proved, p2m-7f145a (O04), p2m-a146ce (F01), p2m-dbd69f (N03), p2m-aec6a2 (N10) and p2m-d00cfc (N07), show that the model can do it, not how often: p2m-7f145a (O04) (trial 3) was a short proof of a leaf nobody had tried yet, and p2m-a146ce (F01) (trial 5) took a new scaffold, two outlines and 19 tasks after trials 3 and 4 had failed on it. p2m-b1d59e (N02) and p2m-f181df (N05), closed by reuse, are not among them. On calibration leaves, which others had proved before, 10 of 15 runs succeeded. All runs together took 91.3 hours of model time, and the outline probes that chose leaves 4.0 more.

Submitted to prove2me

LeafTheoremFromSubmittedprove2me’s verdict
p2m-7f145aO04SeqPostedPricing.MatroidSPM.accept_prob_eq_quantile§4: offered the price F_i^{-1}(1-q_i), agent i accepts with probability exactly q_iTrial 3Accepted06:51 UTC
p2m-a146ceF01CodedSorting.proposition_7_1_transposeProposition 7.1 (rearrangement part; companion: proposition_7_1_correctness) — \pi and \pi^{-1} are three adjacent field interchangesTrial 5Accepted20:10 UTC
p2m-b1d59eN02occupation_bind_sections_integrable_localsame statement as AvramDividend.Classical.avram_chat24_occupation_bind_sections_integrable_public by WillRClosed by reusefound while preparing trial 6Accepted21:55 UTC
p2m-f181dfN05NestedSeatAlloc.IntPolicy.expRevenue_one_right_derivativeEquation (27) right derivative for one-class expected revenuepart of NestedSeatAlloc.IntPolicy.eq27_er1_clbi by mikedeng1Closed by reusefound during trial 6Accepted10:51 UTC
p2m-dbd69fN03PerturbedWF.eq_2_3Eq. (2.3): \langle v,x-y\rangle\le\|x-y\| for v\in B^* and x,y\in\mathcal CTrial 6Accepted14:51 UTC
p2m-aec6a2N10thm_5_6Chapter 5, thm_5_6Trial 7Accepted15:09 UTC
p2m-d00cfcN07MatchIndiff.PublicSM.lemma_2_1Lemma 2.1 — accepted profiles of a fixed board form a product, and a row and board fix the decoded componentTrial 7Accepted20:49 UTC

p2m-7f145a (O04) had been open for 13 days with no submission from anyone. Its proof is short: the distribution has a density, so a single price carries no probability mass. p2m-a146ce (F01) was still open when we submitted it. Its proof follows the leaf’s own statement: the index is split into four fields, each of the outline’s six steps was proved by a task of its own, and the pieces were put together into one proof, which our full check and then prove2me accepted. prove2me’s theorem pages show both leaves as proved. Calibration leaves are never submitted: they were proved before.

p2m-b1d59e (N02) was closed by reuse. The leaf is the same statement as a theorem that WillR had already proved on prove2me under another name. Our index of prove2me’s library found the match while we prepared trial 6, and the model’s proof is one line that imports that theorem. Our full check and then prove2me accepted it, but none of the mathematics is ours, so it is not counted among the open leaves proved.

p2m-f181df (N05) was closed by reuse. The leaf is one part of a theorem that mikedeng1 had already proved on prove2me: its conclusion contains the leaf’s statement. Our index of prove2me’s library offered that theorem as a similar statement during trial 6, and the model’s proof is one line that uses it. Our full check and then prove2me accepted it, but none of the mathematics is ours, so it is not counted among the open leaves proved.

Working on now

Trial 8, since 16 November 2026 · scaffold v9

Budget: one round of 110 steps / 110 min for the outline and for each hole, then one fix round of 20 / 20; up to 2 tries per task.

As of 17 November 2026, 20:00 UTC. The page updates this list from the trial’s published data and the gateway’s task list every minute; without them it shows this copy from the site’s last build.

Every leaf

All 35 leaves tried so far, one row per leaf over every trial that ran it. A leaf’s id here is p2m- and the start of prove2me’s own id for the theorem; the id the trials used (C11, F01, …) is shown under it. Its outcome is the best of its runs: accepted by prove2me, closed by reuse, proved, not proved, stopped by us, or not scored.

RoleLeavesRunsRuns provedProofs acceptedClosed by reuseModel hoursTokens computed
Calibration1215100010.92.3M
Open232944256.413.8M
Comparison3411024.04.9M
Outline probes11––––4.0809k
All3548155295.321.7M

A comparison leaf is an open leaf run again to compare a new scaffold with the trials before it. A leaf can have more than one role over its trials, so the leaves by role add up to more than 35. A proof accepted counts under the role of the run that proved it; a leaf closed by reuse has no run. An outline probe is one short outline task on a candidate leaf, run to choose a trial’s leaves: it is no run, but its model time and tokens count in the totals.

All 35 leaves

Showing 35 of 35

RoleTrials
p2m-6c35d7A01WilliamsonShmoys.connected_planar_supergraph_from_step__p2m_f2188066_bdee_41e3_86df_42250492e63d_a4Open1, 2Not proved0 of 2 runs proved92.9209k
p2m-2499fdA02QAlgorithms.Papers.Independent.Intakethird100.P20.regularized_base_is_quasipointOpen1, 2Not proved0 of 2 runs proved89.9239k
p2m-12ea27A03QAlgorithms.Papers.Independent.Intakethird100.P20.complete_quantum_presheaf_trivialOpen, Comparison1, 2, 4Not proved0 of 3 runs proved155.0755k
p2m-f10c9dA04SongZipkinFluct.Monotone.corollary2Open, Comparison1, 2, 4Not proved0 of 3 runs proved176.4511k
p2m-35e2daB01FasterSFM.lemma_13Open1Not proved29.087k
p2m-ca216eB02OAI.MUB6.Certificate.GramData.Block3.entry_11_11Open1Not scored0.254k
p2m-87ebddB03TenChallenges.challenge_04_r3Open1Not proved16.152k
p2m-5e7987B04FalconerPath.completionOpen1Not proved29.088k
p2m-42603fB05QuantumMaxCutEnergy.exists_0562_approximationOpen1Not scored19.268k
p2m-6ed316B06QMCLeafB.SnapLedger.ledger_round_okOpen1Not proved9.968k
p2m-7ab928C01DPDeconstructed.lemma_6_1Calibration2Proved20.857k
p2m-81b93aC02MarkovProcess.measurable_finTailCalibration2Proved11.619k
p2m-12f452C03OnlineCombOpt.Exp2Full.theorem_1_full_domainCalibration2Not proved65.1176k
p2m-12087dC04WilliamsonShmoys.face_accounting_implies_edge_boundCalibration2Proved5.213k
p2m-8b8abcC05Groundwork.Colouring.card_prescribedCalibration2Not proved49.5157k
p2m-dfe8e4C06BianchiIX.measurableEmbedding_gaussWordCalibration2Proved61.7205k
p2m-a146ceF01CodedSorting.proposition_7_1_transposeOpen, Comparison3, 4, 5Accepted by prove2me1 of 3 runs proved1369.24.6M
p2m-c3c3c4F02APSCounting893.Main.generic_linear_noether_normalizationOpen3Not proved78.0516k
p2m-8503e6F03BirthdayEntropy.Soundness.pair_avoidanceOpen3Not proved69.8209k
p2m-c0eca8O01HolevoCQ.decoder_subPOVMOpen3, 4Not proved0 of 2 runs proved143.7669k
p2m-83aae8O02BBPSSW.Distill.typical_set_sqrtOpen3, 4, 5Not proved0 of 3 runs proved821.73.1M
p2m-d1f097O03DigitalPrivacy.DataStrategy.proof_thm1_case_psi_lt_PrhoOpen3, 4Not proved0 of 2 runs proved165.8482k
p2m-7f145aO04SeqPostedPricing.MatroidSPM.accept_prob_eq_quantileOpen3Accepted by prove2me53.9132k
p2m-3ad60fC07WilliamsonShmoys.planar_drawing_maximal_supergraph__p2m_97829890_428d_4657_a899_f7e200a68a7c_a4Calibration3, 6, 7Proved3 of 3 runs proved143.3502k
p2m-eb3303C08AvramDividend.Classical.avram_ubv_finite9_uniformIntegrable_one_of_common_integrable_dominatorCalibration3Not proved42.4175k
p2m-2b384dC09FiniteEnergyPlane.integral_norm_sq_le_of_skewCalibration3Not proved57.5242k
p2m-479463C10BookProof.ChapterF3.dressed_vacuum_besselCalibration3Proved9.063k
p2m-63eff5C11Ideal.Factors.quotientResidueEquiv_mkCalibration4Proved3.721k
p2m-6f7dceC12InsiderDonations.price_recastCalibration4, 5Proved1 of 2 runs proved185.7627k
p2m-dbd69fN03PerturbedWF.eq_2_3Open6Accepted by prove2me165.6695k
p2m-aec6a2N10thm_5_6Open7Accepted by prove2me20.9138k
p2m-d00cfcN07MatchIndiff.PublicSM.lemma_2_1Open7Accepted by prove2me47.3270k
p2m-115f6cN08QAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Open7Not proved1267.65.8M
p2m-b1d59eN02occupation_bind_sections_integrable_localOpen–Closed by reuse––
p2m-f181dfN05NestedSeatAlloc.IntPolicy.expRevenue_one_right_derivativeOpen–Closed by reuse––

Every task

Every task of every trial and of the outline probes, grouped by leaf and outline; a run of trials 1 to 4 is one task. From trial 6 on, the tasks come live from the gateway’s records, each trial as soon as its tasks are made.

All tasks

Showing 136 of 136

LeafTaskResultWorkFinished
p2m-3ad60fWilliamsonShmoys.planar_drawing_maximal_supergraph__p2m_97829890_428d_4657_a899_f7e200a68a7c_a4Trial 7Outline 1Written, 1 hole73 steps · 63 min · 157k
p2m-3ad60fWilliamsonShmoys.planar_drawing_maximal_supergraph__p2m_97829890_428d_4657_a899_f7e200a68a7c_a4Trial 7Hole 1 (outline 1)Proved24 steps · 8.6 min · 45k
p2m-3ad60fWilliamsonShmoys.planar_drawing_maximal_supergraph__p2m_97829890_428d_4657_a899_f7e200a68a7c_a4Trial 7Assembly (outline 1)Proved: our full check passed––
p2m-aec6a2thm_5_6Trial 7Outline 1Written, 2 holes26 steps · 6.9 min · 68k
p2m-aec6a2thm_5_6Trial 7Hole 1 (outline 1)Proved22 steps · 8.0 min · 37k
p2m-aec6a2thm_5_6Trial 7In place (outline 1)Filled in, checked by the assembly24 steps · 6.0 min · 34k
p2m-aec6a2thm_5_6Trial 7Assembly (outline 1)Proved: our full check passed––
p2m-d00cfcMatchIndiff.PublicSM.lemma_2_1Trial 7Outline 1Degenerate outline23 steps · 13 min · 76k
p2m-d00cfcMatchIndiff.PublicSM.lemma_2_1Trial 7Outline 2replaces outline 1Degenerate outline30 steps · 22 min · 103k
p2m-d00cfcMatchIndiff.PublicSM.lemma_2_1Trial 7Outline 3replaces outline 2Written, 2 holes17 steps · 9.2 min · 64k
p2m-d00cfcMatchIndiff.PublicSM.lemma_2_1Trial 7Hole 1 (outline 3)Proved9 steps · 2.1 min · 15k
p2m-d00cfcMatchIndiff.PublicSM.lemma_2_1Trial 7Hole 2 (outline 3)Proved7 steps · 1.2 min · 13k
p2m-d00cfcMatchIndiff.PublicSM.lemma_2_1Trial 7Assembly (outline 3)Proved: our full check passed––
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Outline 1Written, 8 holes36 steps · 19 min · 106k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Hole 2 (outline 1)Proved11 steps · 3.2 min · 14k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Hole 3 (outline 1)Proved20 steps · 5.0 min · 49k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Hole 4 (outline 1)Proved30 steps · 8.6 min · 47k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Hole 5 (outline 1)Proved31 steps · 13 min · 69k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Hole 6 (outline 1)A sorry left128 steps · 48 min · 234k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Hole 6 (outline 1) · try 2retry of try 1Disproved: the hole is false97 steps · 60 min · 298k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Hole 7 (outline 1)Did not compile128 steps · 67 min · 301k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Hole 7 (outline 1) · try 2retry of try 1A sorry left112 steps · 95 min · 527k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Hole 8 (outline 1)Disproved: the hole is false128 steps · 50 min · 258k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Outline 2replaces outline 1Written, 3 holes57 steps · 72 min · 241k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Hole 1 (outline 2)Proved38 steps · 28 min · 115k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Hole 2 (outline 2)Proved21 steps · 15 min · 50k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Hole 3 (outline 2)A sorry left106 steps · 116 min · 624k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Hole 3 (outline 2) · try 2retry of try 1A sorry left80 steps · 129 min · 666k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Outline 3replaces outline 2Written, 9 holes99 steps · 66 min · 241k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Hole 1 (outline 3)Proved24 steps · 4.4 min · 52k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Hole 2 (outline 3)A sorry left40 steps · 130 min · 517k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Hole 2 (outline 3) · try 2retry of try 1A sorry left119 steps · 110 min · 395k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Hole 3 (outline 3)Proved31 steps · 30 min · 125k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Hole 4 (outline 3)Proved66 steps · 27 min · 119k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Hole 5 (outline 3)Proved65 steps · 27 min · 90k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Hole 6 (outline 3)Proved104 steps · 44 min · 200k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Hole 7 (outline 3)Proved74 steps · 34 min · 135k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Hole 8 (outline 3)A sorry left19 steps · 60 min · 265k
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Hole 8 (outline 3) · try 2retry of try 1Cancelled (outline refuted)––
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7Hole 9 (outline 3)Disproved: the hole is false13 steps · 6.3 min · 39k
p2m-51d621QiFixedDesign.Concentration.lemma_4Trial 7 probesProbe: outlineA sorry left35 steps · 45 min
p2m-d00cfcMatchIndiff.PublicSM.lemma_2_1Trial 7 probesProbe: outlineDegenerate outline18 steps · 14 min
p2m-115f6cQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Trial 7 probesProbe: outlineCancelled–
p2m-17cf43MatchIndiff.PublicSM.lemma_D_9Trial 7 probesProbe: outlineCancelled–
p2m-aec6a2thm_5_6Trial 7 probesProbe: outlineCancelled–
p2m-3ad60fWilliamsonShmoys.planar_drawing_maximal_supergraph__p2m_97829890_428d_4657_a899_f7e200a68a7c_a4Trial 6Outline 1Written, 3 holes19 steps · 6.6 min · 55k
p2m-3ad60fWilliamsonShmoys.planar_drawing_maximal_supergraph__p2m_97829890_428d_4657_a899_f7e200a68a7c_a4Trial 6Hole 2 (outline 1)Proved6 steps · 1.7 min · 14k
p2m-3ad60fWilliamsonShmoys.planar_drawing_maximal_supergraph__p2m_97829890_428d_4657_a899_f7e200a68a7c_a4Trial 6Assembly (outline 1)Proved: our full check passed––
p2m-dbd69fPerturbedWF.eq_2_3Trial 6Outline 1Written, 2 holes59 steps · 27 min · 106k
p2m-dbd69fPerturbedWF.eq_2_3Trial 6Hole 2 (outline 1)Our check failed19 steps · 7.0 min · 39k
p2m-dbd69fPerturbedWF.eq_2_3Trial 6Hole 2 (outline 1) · try 2retry of try 1Cancelled (outline replaced)––
p2m-dbd69fPerturbedWF.eq_2_3Trial 6In place (outline 1)Cancelled (outline replaced)––
p2m-dbd69fPerturbedWF.eq_2_3Trial 6Outline 2replaces outline 1Written, 4 holes28 steps · 13 min · 50k
p2m-dbd69fPerturbedWF.eq_2_3Trial 6Hole 1 (outline 2)Proved21 steps · 4.8 min · 21k
p2m-dbd69fPerturbedWF.eq_2_3Trial 6Hole 3 (outline 2)Proved47 steps · 44 min · 176k
p2m-dbd69fPerturbedWF.eq_2_3Trial 6Hole 4 (outline 2)Proved15 steps · 7.6 min · 28k
p2m-dbd69fPerturbedWF.eq_2_3Trial 6In place (outline 2)Filled in, checked by the assembly118 steps · 62 min · 274k
p2m-dbd69fPerturbedWF.eq_2_3Trial 6Assembly (outline 2)Proved: our full check passed––
p2m-63eff5Ideal.Factors.quotientResidueEquiv_mkTrial 6 probesProbe: outlineProved: our full check passed19 steps · 14 min
p2m-26d409predictable_section_measurable_localTrial 6 probesProbe: outlineWritten, 2 holes41 steps · 39 min
p2m-b1d59eoccupation_bind_sections_integrable_localTrial 6 probesProbe: outlineProved: our full check passed11 steps · 4.3 min
p2m-dbd69fPerturbedWF.eq_2_3Trial 6 probesProbe: outlineDid not compile33 steps · 45 min
p2m-c1b761PochetWolsey.Variants.sales_extreme_pointsTrial 6 probesProbe: outlineDid not compile16 steps · 45 min
p2m-f181dfNestedSeatAlloc.IntPolicy.expRevenue_one_right_derivativeTrial 6 probesProbe: outlineOutline written43 steps · 33 min
p2m-a146ceCodedSorting.proposition_7_1_transposeTrial 5Outline 1Written, 7 holes29 steps · 25 min · 135k
p2m-a146ceCodedSorting.proposition_7_1_transposeTrial 5Hole 1 (outline 1)Proved25 steps · 20 min · 80k
p2m-a146ceCodedSorting.proposition_7_1_transposeTrial 5Hole 2 (outline 1)Proved35 steps · 39 min · 235k
p2m-a146ceCodedSorting.proposition_7_1_transposeTrial 5Hole 3 (outline 1)Proved32 steps · 20 min · 104k
p2m-a146ceCodedSorting.proposition_7_1_transposeTrial 5Hole 4 (outline 1)Proved81 steps · 94 min · 303k
p2m-a146ceCodedSorting.proposition_7_1_transposeTrial 5Hole 5 (outline 1)Proved20 steps · 6.3 min · 29k
p2m-a146ceCodedSorting.proposition_7_1_transposeTrial 5Hole 6 (outline 1)Did not compile92 steps · 130 min · 364k
p2m-a146ceCodedSorting.proposition_7_1_transposeTrial 5Hole 6 (outline 1) · try 2retry of try 1Did not compile65 steps · 131 min · 363k
p2m-a146ceCodedSorting.proposition_7_1_transposeTrial 5Hole 7 (outline 1)Did not compile45 steps · 130 min · 335k
p2m-a146ceCodedSorting.proposition_7_1_transposeTrial 5Hole 7 (outline 1) · try 2retry of try 1A sorry left105 steps · 130 min · 278k
p2m-a146ceCodedSorting.proposition_7_1_transposeTrial 5Outline 2replaces outline 1Written, 6 holes27 steps · 33 min · 87k
p2m-a146ceCodedSorting.proposition_7_1_transposeTrial 5Hole 1 (outline 2)Proved31 steps · 39 min · 102k
p2m-a146ceCodedSorting.proposition_7_1_transposeTrial 5Hole 2 (outline 2)Proved36 steps · 56 min · 141k
p2m-a146ceCodedSorting.proposition_7_1_transposeTrial 5Hole 3 (outline 2)Did not compile89 steps · 130 min · 428k
p2m-a146ceCodedSorting.proposition_7_1_transposeTrial 5Hole 3 (outline 2) · try 2retry of try 1Proved38 steps · 90 min · 245k
p2m-a146ceCodedSorting.proposition_7_1_transposeTrial 5Hole 4 (outline 2)Proved38 steps · 67 min · 166k
p2m-a146ceCodedSorting.proposition_7_1_transposeTrial 5Hole 5 (outline 2)Proved18 steps · 14 min · 38k
p2m-a146ceCodedSorting.proposition_7_1_transposeTrial 5Hole 6 (outline 2)Proved39 steps · 55 min · 218k
p2m-a146ceCodedSorting.proposition_7_1_transposeTrial 5Assembly (outline 2)Proved: our full check passed––
p2m-6f7dceInsiderDonations.price_recastTrial 5Outline 1Written, 1 hole66 steps · 110 min · 371k
p2m-6f7dceInsiderDonations.price_recastTrial 5Hole 1 (outline 1)Proved25 steps · 25 min · 97k
p2m-6f7dceInsiderDonations.price_recastTrial 5Assembly (outline 1)Assembly did not check––
p2m-6f7dceInsiderDonations.price_recastTrial 5Assembly (outline 1) · try 2retry of try 1Proved: our full check passed––
p2m-83aae8BBPSSW.Distill.typical_set_sqrtTrial 5Outline 1Written, 6 holes32 steps · 56 min · 115k
p2m-83aae8BBPSSW.Distill.typical_set_sqrtTrial 5Hole 1 (outline 1)Stopped by a fault18 steps · 19 min · 77k
p2m-83aae8BBPSSW.Distill.typical_set_sqrtTrial 5Hole 1 (outline 1) · try 2retry of try 1Stopped, no progress91 steps · 110 min · 300k
p2m-83aae8BBPSSW.Distill.typical_set_sqrtTrial 5Hole 2 (outline 1)Stopped, no progress25 steps · 110 min · 260k
p2m-83aae8BBPSSW.Distill.typical_set_sqrtTrial 5Hole 2 (outline 1) · try 2retry of try 1A sorry left127 steps · 88 min · 268k
p2m-83aae8BBPSSW.Distill.typical_set_sqrtTrial 5Hole 3 (outline 1)A sorry left82 steps · 130 min · 403k
p2m-83aae8BBPSSW.Distill.typical_set_sqrtTrial 5In place (outline 1)Stopped, no progress100 steps · 68 min · 335k
p2m-83aae8BBPSSW.Distill.typical_set_sqrtTrial 5Outline 2replaces outline 1Written, 12 holes44 steps · 61 min · 162k
p2m-83aae8BBPSSW.Distill.typical_set_sqrtTrial 5Hole 1 (outline 2)Cancelled––
p2m-83aae8BBPSSW.Distill.typical_set_sqrtTrial 5In place (outline 2)Cancelled––
p2m-63eff5Ideal.Factors.quotientResidueEquiv_mkTrial 4RunProved: our full check passed20 steps · 3.7 min · 21k
p2m-6f7dceInsiderDonations.price_recastTrial 4RunA sorry left15 steps · 50 min · 159k
p2m-12ea27QAlgorithms.Papers.Independent.Intakethird100.P20.complete_quantum_presheaf_trivialTrial 4RunStopped, no progress13 steps · 65 min · 311k
p2m-f10c9dSongZipkinFluct.Monotone.corollary2Trial 4RunA sorry left43 steps · 83 min · 232k
p2m-a146ceCodedSorting.proposition_7_1_transposeTrial 4RunA sorry left45 steps · 81 min · 666k
p2m-c0eca8HolevoCQ.decoder_subPOVMTrial 4RunStopped, no progress52 steps · 65 min · 348k
p2m-83aae8BBPSSW.Distill.typical_set_sqrtTrial 4RunA sorry left48 steps · 85 min · 268k
p2m-d1f097DigitalPrivacy.DataStrategy.proof_thm1_case_psi_lt_PrhoTrial 4RunA sorry left39 steps · 85 min · 249k
p2m-a146ceCodedSorting.proposition_7_1_transposeTrial 3RunA sorry left59 steps · 79 min · 270k
p2m-c3c3c4APSCounting893.Main.generic_linear_noether_normalizationTrial 3RunA sorry left42 steps · 78 min · 516k
p2m-8503e6BirthdayEntropy.Soundness.pair_avoidanceTrial 3RunA sorry left19 steps · 70 min · 209k
p2m-c0eca8HolevoCQ.decoder_subPOVMTrial 3RunDid not compile59 steps · 79 min · 321k
p2m-83aae8BBPSSW.Distill.typical_set_sqrtTrial 3RunA sorry left49 steps · 95 min · 874k
p2m-d1f097DigitalPrivacy.DataStrategy.proof_thm1_case_psi_lt_PrhoTrial 3RunDid not compile35 steps · 81 min · 233k
p2m-7f145aSeqPostedPricing.MatroidSPM.accept_prob_eq_quantileTrial 3RunProved: our full check passed47 steps · 54 min · 132k
p2m-3ad60fWilliamsonShmoys.planar_drawing_maximal_supergraph__p2m_97829890_428d_4657_a899_f7e200a68a7c_a4Trial 3RunProved: our full check passed60 steps · 64 min · 231k
p2m-eb3303AvramDividend.Classical.avram_ubv_finite9_uniformIntegrable_one_of_common_integrable_dominatorTrial 3RunDid not compile65 steps · 42 min · 175k
p2m-2b384dFiniteEnergyPlane.integral_norm_sq_le_of_skewTrial 3RunA sorry left48 steps · 58 min · 242k
p2m-479463BookProof.ChapterF3.dressed_vacuum_besselTrial 3RunProved: our full check passed21 steps · 9.0 min · 63k
p2m-7ab928DPDeconstructed.lemma_6_1Trial 2RunProved: our full check passed27 steps · 21 min · 57k
p2m-81b93aMarkovProcess.measurable_finTailTrial 2RunProved: our full check passed17 steps · 12 min · 19k
p2m-12f452OnlineCombOpt.Exp2Full.theorem_1_full_domainTrial 2RunA sorry left27 steps · 65 min · 176k
p2m-12087dWilliamsonShmoys.face_accounting_implies_edge_boundTrial 2RunProved: our full check passed12 steps · 5.2 min · 13k
p2m-8b8abcGroundwork.Colouring.card_prescribedTrial 2RunA sorry left68 steps · 50 min · 157k
p2m-dfe8e4BianchiIX.measurableEmbedding_gaussWordTrial 2RunProved: our full check passed57 steps · 62 min · 205k
p2m-6c35d7WilliamsonShmoys.connected_planar_supergraph_from_step__p2m_f2188066_bdee_41e3_86df_42250492e63d_a4Trial 2RunDid not compile45 steps · 68 min · 143k
p2m-2499fdQAlgorithms.Papers.Independent.Intakethird100.P20.regularized_base_is_quasipointTrial 2RunDid not compile46 steps · 64 min · 162k
p2m-12ea27QAlgorithms.Papers.Independent.Intakethird100.P20.complete_quantum_presheaf_trivialTrial 2RunA sorry left36 steps · 60 min · 323k
p2m-f10c9dSongZipkinFluct.Monotone.corollary2Trial 2RunA sorry left45 steps · 63 min · 196k
p2m-6c35d7WilliamsonShmoys.connected_planar_supergraph_from_step__p2m_f2188066_bdee_41e3_86df_42250492e63d_a4Trial 1RunDid not compile29 steps · 25 min · 66k
p2m-2499fdQAlgorithms.Papers.Independent.Intakethird100.P20.regularized_base_is_quasipointTrial 1RunDid not compile7 steps · 26 min · 78k
p2m-12ea27QAlgorithms.Papers.Independent.Intakethird100.P20.complete_quantum_presheaf_trivialTrial 1RunNo solution file25 steps · 30 min · 121k
p2m-f10c9dSongZipkinFluct.Monotone.corollary2Trial 1RunA sorry left20 steps · 30 min · 83k
p2m-35e2daFasterSFM.lemma_13Trial 1RunNo solution file14 steps · 29 min · 87k
p2m-ca216eOAI.MUB6.Certificate.GramData.Block3.entry_11_11Trial 1RunNo solution file3 steps · 0.2 min · 54k
p2m-87ebddTenChallenges.challenge_04_r3Trial 1RunNo solution file29 steps · 16 min · 52k
p2m-5e7987FalconerPath.completionTrial 1RunNo solution file26 steps · 29 min · 88k
p2m-42603fQuantumMaxCutEnergy.exists_0562_approximationTrial 1RunNo solution file19 steps · 19 min · 68k
p2m-6ed316QMCLeafB.SnapLedger.ledger_round_okTrial 1RunNo solution file30 steps · 9.9 min · 68k

Trials

Each trial changed the agent’s scaffold (its instructions, tools and rounds) from what the one before taught, so the trials are steps, not repeats. Tap a trial for every leaf it ran.

TrialDatesScaffoldProvedCalibrationOpenModel hoursTokens computed
Trial 1v10 of 10–0 of 103.6763k
Trial 2v24 of 104 of 60 of 47.81.5M
Trial 3v33 of 112 of 41 of 711.83.3M
Trial 4v41 of 81 of 20 of 68.72.3M
Trial 5v52 of 31 of 11 of 233.16.0M
Trial 6 probesv6–––3.0558k
Trial 6v72 of 21 of 11 of 12.9763k
Trial 7 probesv7–––1.0251k
Trial 7v83 of 41 of 12 of 323.56.4M
All15 of 4810 of 155 of 3395.321.7M
Trial 1 · 7 November 2026 · 0 of 10 proved

No leaf proved. Seven of ten runs never wrote a solution file: the agent spent most of its steps searching library source. Two runs were cut off by the gateway's personal-data filter, which took long integers in the mathematics for card numbers; that filter is fixed and the two runs are not scored.

Budget per leaf: one round of 30 steps / 30 model minutes.

LeafTheoremKindResultRoundsModel minTokens computed
p2m-6c35d7A01WilliamsonShmoys.connected_planar_supergraph_from_step__p2m_f2188066_bdee_41e3_86df_42250492e63d_a4Finite maximality turns strict planar extensions into connectivityOpenNot proved: did not compile124.866k
p2m-2499fdA02QAlgorithms.Papers.Independent.Intakethird100.P20.regularized_base_is_quasipointLemma 3.4OpenNot proved: did not compile126.478k
p2m-12ea27A03QAlgorithms.Papers.Independent.Intakethird100.P20.complete_quantum_presheaf_trivialProposition 3.1OpenNot proved: no solution file129.5121k
p2m-f10c9dA04SongZipkinFluct.Monotone.corollary2Corollary 2: the finite-horizon basestock levels y^*_n(i) converge as n\to\inftyOpenNot proved: a sorry left130.083k
p2m-35e2daB01FasterSFM.lemma_13Lemma 13 — inserted and cutting-plane row bounds: k\le k_n, Q_n<Kp_n^2, \chi_n<1OpenNot proved: no solution file129.087k
p2m-ca216eB02OAI.MUB6.Certificate.GramData.Block3.entry_11_11OAI.MUB6.Certificate.GramData.Block3.entry_11_11: intermediate lemma from OpenAI, The maximum number of mutually unbiased bases in dimension sixOpenNot scoredcut off by a false positive of the gateway's personal-data filter (since fixed)10.254k
p2m-87ebddB03TenChallenges.challenge_04_r3Sidorenko's inequality for sides of size at most 3OpenNot proved: no solution file116.152k
p2m-5e7987B04FalconerPath.completionA nonzero absolutely continuous measure on the original distance setOpenNot proved: no solution file129.088k
p2m-42603fB05QuantumMaxCutEnergy.exists_0562_approximationClassical 0.562 approximation of Quantum Max-Cut energyOpenNot scoredcut off by a false positive of the gateway's personal-data filter (since fixed)119.268k
p2m-6ed316B06QMCLeafB.SnapLedger.ledger_round_okSnap displacement versus the recorded margin (margin–ledger joint)OpenNot proved: no solution file19.968k
Trial 2 · 8 November 2026 · 4 of 10 proved

A new scaffold: write the solution file early, build after every edit, ask Lean instead of searching, and up to two fix rounds with Lean's errors. 4 of 6 calibration leaves proved, 3 of them in the first round; none of the 4 open leaves.

Budget per leaf: 30 steps / 30 min, then up to 2 fix rounds of 20 / 20.

LeafTheoremKindResultRoundsModel minTokens computed
p2m-7ab928C01DPDeconstructed.lemma_6_1Lemma 6.1 — fixed points and iterates pass around the cycle \tau_0,\tau_1,\tau_2CalibrationProved120.857k
p2m-81b93aC02MarkovProcess.measurable_finTailMeasurability of dropping the first finite coordinateCalibrationProved111.619k
p2m-12f452C03OnlineCombOpt.Exp2Full.theorem_1_full_domainTheorem 1 — full-domain EXP2 regret lower bound (corrected: excludes dimension one)CalibrationNot proved: a sorry left365.1176k
p2m-12087dC04WilliamsonShmoys.face_accounting_implies_edge_boundFace accounting implies the planar edge boundCalibrationProved15.213k
p2m-8b8abcC05Groundwork.Colouring.card_prescribedColouring: card prescribedCalibrationNot proved: a sorry left349.5157k
p2m-dfe8e4C06BianchiIX.measurableEmbedding_gaussWordMeasurable embeddings defined by finite Gauss wordsCalibrationProved in round 3361.7205k
p2m-6c35d7A01WilliamsonShmoys.connected_planar_supergraph_from_step__p2m_f2188066_bdee_41e3_86df_42250492e63d_a4Finite maximality turns strict planar extensions into connectivityOpentried before in trial 1Not proved: did not compile368.1143k
p2m-2499fdA02QAlgorithms.Papers.Independent.Intakethird100.P20.regularized_base_is_quasipointLemma 3.4Opentried before in trial 1Not proved: did not compile363.5162k
p2m-12ea27A03QAlgorithms.Papers.Independent.Intakethird100.P20.complete_quantum_presheaf_trivialProposition 3.1Opentried before in trial 1Not proved: a sorry left360.1323k
p2m-f10c9dA04SongZipkinFluct.Monotone.corollary2Corollary 2: the finite-horizon basestock levels y^*_n(i) converge as n\to\inftyOpentried before in trial 1Not proved: a sorry left363.4196k
Trial 3 · 9 to 12 November 2026 · 3 of 11 proved

A pre-written skeleton and the remaining goals in each fix message. 3 of 11 proved, among them O04, the first open leaf proved in any trial, which prove2me then accepted. O04 turned out short, and nobody had tried it: an existence proof, not a rate. About two thirds of the trial's 58 hours were lost to outages; no leaf lost model time.

Budget per leaf: 45 / 45 for open leaves (30 / 30 for calibration), then 2 fix rounds of 20 / 20.

LeafTheoremKindResultRoundsModel minTokens computed
p2m-a146ceF01CodedSorting.proposition_7_1_transposeProposition 7.1 (rearrangement part; companion: proposition_7_1_correctness) — \pi and \pi^{-1} are three adjacent field interchangesOpenNot proved: a sorry left379.3270k
p2m-c3c3c4F02APSCounting893.Main.generic_linear_noether_normalizationLemma 3.15 — Generic linear Noether normalizationOpenNot proved: a sorry left378.0516k
p2m-8503e6F03BirthdayEntropy.Soundness.pair_avoidanceLemma 4.1 — Pair avoidance under fixed-size samplingOpenNot proved: a sorry left369.8209k
p2m-c0eca8O01HolevoCQ.decoder_subPOVM§4: the decision operators (11) satisfy \sum_uX_u\le IOpenNot proved: did not compile378.5321k
p2m-83aae8O02BBPSSW.Distill.typical_set_sqrtTypical sets of size 2^{H+O(\sqrt n)} carry almost all the weightOpenNot proved: a sorry left394.5874k
p2m-d1f097O03DigitalPrivacy.DataStrategy.proof_thm1_case_psi_lt_PrhoProof of Theorem 1, case \psi<P\rho (corrected: \psi>0): C_{{inv}}(\xi)<C_h(\xi) for every \xi>0OpenNot proved: did not compile380.6233k
p2m-7f145aO04SeqPostedPricing.MatroidSPM.accept_prob_eq_quantile§4: offered the price F_i^{-1}(1-q_i), agent i accepts with probability exactly q_iOpenProved in round 2253.9132k
p2m-3ad60fC07WilliamsonShmoys.planar_drawing_maximal_supergraph__p2m_97829890_428d_4657_a899_f7e200a68a7c_a4A finite maximal drawable supergraph existsCalibrationProved in round 3363.6231k
p2m-eb3303C08AvramDividend.Classical.avram_ubv_finite9_uniformIntegrable_one_of_common_integrable_dominatorAny uniformly dominated L1 family is uniformly integrable at exponent oneCalibrationNot proved: did not compile342.4175k
p2m-2b384dC09FiniteEnergyPlane.integral_norm_sq_le_of_skewPoincaré inequality on an interval for a curve vanishing at the left endpoint, with a skew-twisted derivativeCalibrationNot proved: a sorry left357.5242k
p2m-479463C10BookProof.ChapterF3.dressed_vacuum_bessel`BookProof.ChapterF3.dressed_vacuum_bessel` {ι : Type*} (x : ι → E') (hx : Orthonormal ℂ x) (v : E') (hv : ‖v‖ = 1) (s : Finset ι) : ∑ j ∈ s, ‖inner (𝕜CalibrationProved19.063k
Trial 4 · 12 to 13 November 2026 · 1 of 8 proved

One sorry at a time, and a stop when a fix round changes nothing. Leaves now run as tasks on the gateway's queue. 1 of 8 proved (calibration leaf C11). Six leaves that earlier trials had failed were tried again: none proved, and no hole closed. Every hard first round ran into its 45-minute cap, half of the model time in long reasoning turns.

Budget per leaf: 45 / 45, then 2 fix rounds of 20 / 20.

LeafTheoremKindResultRoundsModel minTokens computed
p2m-63eff5C11Ideal.Factors.quotientResidueEquiv_mkFactors: quotient Residue Equiv mkCalibrationProved13.721k
p2m-6f7dceC12InsiderDonations.price_recast§3.2, p.9 — P^*=\eta v+[1-\eta]\bar v-\lambda^*u-\sum_i(\psi_i-\bar\psi)/(2\tau_o)CalibrationNot scoreda fix round ended with no tool call (a tooling fault, since fixed), so the leaf lost its last round250.4159k
p2m-12ea27A03QAlgorithms.Papers.Independent.Intakethird100.P20.complete_quantum_presheaf_trivialProposition 3.1Opentried before in trials 1 and 2Not proved: stopped, no progress265.4311k
p2m-f10c9dA04SongZipkinFluct.Monotone.corollary2Corollary 2: the finite-horizon basestock levels y^*_n(i) converge as n\to\inftyOpentried before in trials 1 and 2Not proved: a sorry left383.0232k
p2m-a146ceF01CodedSorting.proposition_7_1_transposeProposition 7.1 (rearrangement part; companion: proposition_7_1_correctness) — \pi and \pi^{-1} are three adjacent field interchangesOpentried before in trial 3Not proved: a sorry left381.1666k
p2m-c0eca8O01HolevoCQ.decoder_subPOVM§4: the decision operators (11) satisfy \sum_uX_u\le IOpentried before in trial 3Not proved: stopped, no progress265.2348k
p2m-83aae8O02BBPSSW.Distill.typical_set_sqrtTypical sets of size 2^{H+O(\sqrt n)} carry almost all the weightOpentried before in trial 3Not proved: a sorry left385.3268k
p2m-d1f097O03DigitalPrivacy.DataStrategy.proof_thm1_case_psi_lt_PrhoProof of Theorem 1, case \psi<P\rho (corrected: \psi>0): C_{{inv}}(\xi)<C_h(\xi) for every \xi>0Opentried before in trial 3Not proved: a sorry left385.2249k
Trial 5 · 13 to 14 November 2026 · 2 of 3 proved

The model first writes an outline of the proof in which every step it has not proved is a sorry hole. Each hole then becomes a task of its own, after a quick pass of standard tactics, a ledger keeps every hole Lean has checked, and an assembly puts the pieces together. 2 of 3 leaves proved. F01, an open leaf that trials 3 and 4 had failed, was proved and accepted by prove2me: its first outline left the main claim in two holes, which failed twice; a second outline split the index into four fields, as the leaf’s own statement does, and all 6 of its holes were proved and assembled on the first try. Calibration leaf C12 was proved on the second try of its assembly. We stopped O02, which is not scored.

Budget per leaf: 90 steps / 90 min for the outline and for each hole, then 2 fix rounds of 20 / 20; up to 2 tries per task.

LeafTheoremKindResultTasksModel minTokens computed
p2m-a146ceF01CodedSorting.proposition_7_1_transposeProposition 7.1 (rearrangement part; companion: proposition_7_1_correctness) — \pi and \pi^{-1} are three adjacent field interchangesOpentried before in trials 3 and 4Proved with outline 2192 outlines1208.83.6M
p2m-6f7dceC12InsiderDonations.price_recast§3.2, p.9 — P^*=\eta v+[1-\eta]\bar v-\lambda^*u-\sum_i(\psi_i-\bar\psi)/(2\tau_o)Calibrationtried before in trial 4Proved41 outline135.3468k
p2m-83aae8O02BBPSSW.Distill.typical_set_sqrtTypical sets of size 2^{H+O(\sqrt n)} carry almost all the weightOpentried before in trials 3 and 4Not scoredstopped by us: its first outline proved none of its 6 holes, and a harness bug (since fixed) left 11 of its second outline’s 12 holes in place102 outlines641.91.9M

Each leaf’s outlines, newest first, with what became of every hole; under each, every task (outline, hole, try and assembly) with its counts, as in “Working on now”.

p2m-a146ce (F01)

Proved: outline 2 assembled

Outline 2proved (assembled)written in 27 steps · 33.1 min

  • Hole 1proved
  • Hole 2proved
  • Hole 3proved in try 2 (try 1: did not compile)
  • Hole 4proved
  • Hole 5proved
  • Hole 6proved
  • Assemblyproved: our full check passed

Outline 1abandoned: holes 6 and 7 not proved in 2 trieswritten in 29 steps · 25.4 min

  • Hole 1proved
  • Hole 2proved
  • Hole 3proved
  • Hole 4proved
  • Hole 5proved
  • Hole 6not proved in 2 tries (did not compile, then did not compile)
  • Hole 7not proved in 2 tries (did not compile, then a sorry left)
p2m-a146ce (F01): every task (19)
TaskOutlineTryResultStepsModel minTokens computedWith cache re-reads
Outline11Outline written2925.4135k1.2M
Hole 1pre-pass timed out11Proved2519.580k823k
Hole 2pre-pass did not close it11Proved3538.8235k1.3M
Hole 3pre-pass did not close it11Proved3219.6104k880k
Hole 4pre-pass did not close it11Proved8193.9303k5.1M
Hole 5pre-pass did not close it11Proved206.329k332k
Hole 6pre-pass did not close it11Not proved: did not compile92130.2364k5.4M
Hole 612Not proved: did not compile65130.6363k3.4M
Hole 7pre-pass did not close it11Not proved: did not compile45130.3335k3.2M
Hole 712Not proved: a sorry left105130.4278k6.3M
Outline21Outline written2733.187k1.3M
Hole 1pre-pass did not close it21Proved3139.0102k1.4M
Hole 2pre-pass did not close it21Proved3655.8141k2.4M
Hole 3pre-pass did not close it21Not proved: did not compile89130.3428k5.9M
Hole 322Proved3890.1245k2.0M
Hole 4pre-pass did not close it21Proved3866.7166k2.8M
Hole 5pre-pass did not close it21Proved1813.538k515k
Hole 6pre-pass did not close it21Proved3955.3218k2.2M
Assembly21Proved00.000
All 19 tasks8451208.83.6M46.8M

Outline 1: 7 holes, all extracted as lemmas; the pre-pass of standard tactics closed 0 of 7 (709 s).

Outline 2: 6 holes, all extracted as lemmas; the pre-pass of standard tactics closed 0 of 6 (470 s).

p2m-6f7dce (C12)

Proved: outline 1 assembled

Outline 1proved (assembled)written in 66 steps · 110.2 min

  • Hole 1proved
  • Assemblyproved: our full check passed (try 2)
p2m-6f7dce (C12): every task (4)
TaskOutlineTryResultStepsModel minTokens computedWith cache re-reads
Outline11Outline written66110.2371k3.9M
Hole 1pre-pass did not close it11Proved2525.197k917k
Assembly11Assembly did not check00.000
Assembly12Proved00.000
All 4 tasks91135.3468k4.8M

Outline 1: 1 hole, extracted as a lemma; the pre-pass of standard tactics closed 0 of 1 (71 s).

p2m-83aae8 (O02)

Stopped: outline 2 was abandoned (its tasks were cancelled)

Outline 2abandoned: its tasks were cancelledwritten in 44 steps · 61.0 min

  • Hole 1not tried (outline abandoned)
  • In placeholes 2 to 12, not tried (outline abandoned)

Outline 1abandoned: holes 1, 2 and 3 and the holes in place not provedwritten in 32 steps · 55.8 min

  • Hole 1not proved in 2 tries (a fault on our side, then stopped, no progress)
  • Hole 2not proved in 2 tries (stopped, no progress, then a sorry left)
  • Hole 3not proved: a sorry left
  • In placeholes 4 to 6, not proved: stopped, no progress
p2m-83aae8 (O02): every task (10)
TaskOutlineTryResultStepsModel minTokens computedWith cache re-reads
Outline11Outline written3255.8115k1.8M
Hole 1pre-pass did not close it11Not proved: a fault on our side1818.977k428k
Hole 112Not proved: stopped, no progress91110.0300k5.9M
Hole 2pre-pass did not close it11Not proved: stopped, no progress25110.2260k1.1M
Hole 212Not proved: a sorry left12787.6268k7.5M
Hole 3pre-pass did not close it11Not proved: a sorry left82130.4403k5.4M
In placethe 3 holes left in place, in one task11Not proved: stopped, no progress10068.0335k6.2M
Outline21Outline written4461.0162k2.5M
Hole 1pre-pass timed out21Cancelled––––
In placethe 11 holes left in place, in one task21Cancelled––––
All 10 tasks519641.91.9M30.7M

Outline 1: 6 holes, 3 extracted as lemmas, 3 left in place; the pre-pass of standard tactics closed 0 of 3 (268 s).

Outline 2: 12 holes, 1 extracted as a lemma, 11 left in place; the pre-pass of standard tactics closed 0 of 1 (118 s).

Trial 6 · 15 November 2026 · 2 of 2 proved

Trial 6 ran on 15 November 2026 on scaffold v7: 2 leaves (1 calibration, 1 open), 2 of 2 runs proved; prove2me accepted 1 proof from it and 2 leaves closed by reuse.

The outline task now starts with a plan, every step of it becomes a sorry hole (at most 12 holes and 300 lines), and an outline that does not compile is repaired by turning the parts Lean rejects into holes. All 3 leaves were closed. Open leaf p2m-dbd69f was proved by the model and accepted by prove2me: its first outline had a false hole, which the model disproved; its second outline, the paper's duality argument in four steps, had every hole proved and was assembled. Open leaf p2m-f181df was closed by reuse of a theorem already proved on prove2me, and calibration leaf p2m-3ad60f was proved by assembly.

Budget per leaf: one round of 110 steps / 110 min for the outline and for each hole, then one fix round of 20 / 20; up to 2 tries per task.

LeafTheoremKindResultTasksModel minTokens computed
p2m-3ad60fC07WilliamsonShmoys.planar_drawing_maximal_supergraph__p2m_97829890_428d_4657_a899_f7e200a68a7c_a4A finite maximal drawable supergraph existsCalibrationtried before in trial 3ProvedProved: outline 1 assembled31 outline8.369k
p2m-dbd69fN03PerturbedWF.eq_2_3Eq. (2.3): \langle v,x-y\rangle\le\|x-y\| for v\in B^* and x,y\in\mathcal COpenProved with outline 2Proved: outline 2 assembled102 outlines165.6695k

Each leaf’s outlines, newest first, with what became of every hole; under each, every task (outline, hole, try and assembly) with its counts, as in “Working on now”.

p2m-3ad60f (C07)

Proved: outline 1 assembled

Outline 1proved (assembled)written in 19 steps · 6.6 min

  • Hole 1closed by the index (reuse)
  • Hole 2proved
  • Hole 3closed by the pre-pass
  • Assemblyproved: our full check passed
p2m-3ad60f (C07): every task (3)
TaskOutlineTryResultStepsModel minTokens computedWith cache re-reads
Outline11Outline written196.655k440k
Hole 2pre-pass did not close it11Proved61.714k86k
Assembly11Proved00.000
All 3 tasks258.369k526k

Outline 1: 3 holes, all extracted as lemmas; the pre-pass of standard tactics closed 2 of 3 (479 s).

p2m-dbd69f (N03)

Proved: outline 2 assembled

Outline 2proved (assembled)written in 28 steps · 12.8 min

  • Hole 1proved
  • Hole 2in place, proved (checked in the assembly)
  • Hole 3proved
  • Hole 4proved
  • Assemblyproved: our full check passed

Outline 1abandoned: its tasks were cancelledwritten in 59 steps · 27.0 min

  • Hole 1in place, not tried (outline abandoned)
  • Hole 2try 1: our check failed; try 2 not tried (outline abandoned)
p2m-dbd69f (N03): every task (10)
TaskOutlineTryResultStepsModel minTokens computedWith cache re-reads
Outline11Outline written5927.0106k2.6M
Hole 2pre-pass did not close it11Our check failed197.039k386k
Hole 212Cancelled––––
In placethe 1 hole left in place, in one task11Cancelled––––
Outline21Outline written2812.850k1.1M
Hole 1pre-pass did not close it21Proved214.821k362k
Hole 3pre-pass did not close it21Proved4744.2176k2.4M
Hole 4pre-pass did not close it21Proved157.628k341k
In placethe 1 hole left in place, in one task21Filled in, checked by the assembly11862.2274k9.2M
Assembly21Proved00.000
All 10 tasks307165.6695k16.3M

Outline 1: 2 holes, 1 extracted as a lemma, 1 left in place; the pre-pass of standard tactics closed 0 of 1 (89 s).

Outline 2: 4 holes, 3 extracted as lemmas, 1 left in place; the pre-pass of standard tactics closed 0 of 3 (316 s).

Trial 7 · 15 to 17 November 2026 · 3 of 4 proved

Trial 7 ran from 15 to 17 November 2026 on scaffold v8: 4 leaves (1 calibration, 3 open), 3 of 4 runs proved; prove2me accepted 2 proofs from it.

Scaffold v8: v7 with what trial 6 taught. A counterexample to a false hole goes into the next outline's prompt, and an outline that proves the leaf ends it. 3 of 4 leaves were proved, and prove2me accepted both open leaves among them. Open leaf p2m-aec6a2 was proved from its first outline: its two holes were closed and the proof assembled. Open leaf p2m-d00cfc took three outlines: standard tactics closed every hole of the first two, which the outline check then refused as having no real step (since changed: such an outline is a proof); the third had 2 holes, both proved, and was assembled. Calibration leaf p2m-3ad60f was proved again, through an outline with one hole. Open leaf p2m-115f6c (Pinsker's inequality, the hard one) was not proved: of its three outlines, with 8, 3 and 9 holes, most holes were proved, but in each some hole failed both tries or was shown to be false.

Budget per leaf: one round of 110 steps / 110 min for the outline and for each hole, then one fix round of 20 / 20; up to 2 tries per task.

LeafTheoremKindResultTasksModel minTokens computed
p2m-3ad60fC07WilliamsonShmoys.planar_drawing_maximal_supergraph__p2m_97829890_428d_4657_a899_f7e200a68a7c_a4A finite maximal drawable supergraph existsCalibrationtried before in trials 3 and 6ProvedProved: outline 1 assembled31 outline71.4202k
p2m-aec6a2N10thm_5_6Chapter 5, thm_5_6OpenProvedProved: outline 1 assembled41 outline20.9138k
p2m-d00cfcN07MatchIndiff.PublicSM.lemma_2_1Lemma 2.1 — accepted profiles of a fixed board form a product, and a row and board fix the decoded componentOpenProved with outline 3Proved: outline 3 assembled63 outlines47.3270k
p2m-115f6cN08QAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1Pinsker inequality in the source full-L1/nats conventionOpenNot provedNot proved: outline 3 was refuted (hole 9 is false)273 outlines1267.65.8M

Each leaf’s outlines, newest first, with what became of every hole; under each, every task (outline, hole, try and assembly) with its counts, as in “Working on now”.

p2m-3ad60f (C07)

Proved: outline 1 assembled

Outline 1proved (assembled)written in 73 steps · 62.8 min

  • Hole 1proved
  • Assemblyproved: our full check passed
p2m-3ad60f (C07): every task (3)
TaskOutlineTryResultStepsModel minTokens computedWith cache re-reads
Outline11Outline written7362.8157k4.2M
Hole 1pre-pass did not close it11Proved248.645k741k
Assembly11Proved00.000
All 3 tasks9771.4202k4.9M

Outline 1: 1 hole, extracted as a lemma; the pre-pass of standard tactics closed 0 of 1 (91 s).

p2m-aec6a2 (N10)

Proved: outline 1 assembled

Outline 1proved (assembled)written in 26 steps · 6.9 min

  • Hole 1proved
  • Hole 2in place, proved (checked in the assembly)
  • Assemblyproved: our full check passed
p2m-aec6a2 (N10): every task (4)
TaskOutlineTryResultStepsModel minTokens computedWith cache re-reads
Outline11Outline written266.968k676k
Hole 1pre-pass did not close it11Proved228.037k429k
In placethe 1 hole left in place, in one task11Filled in, checked by the assembly246.034k526k
Assembly11Proved00.000
All 4 tasks7220.9138k1.6M

Outline 1: 2 holes, 1 extracted as a lemma, 1 left in place; the pre-pass of standard tactics closed 0 of 1 (83 s).

p2m-d00cfc (N07)

Proved: outline 3 assembled

Outline 3proved (assembled)written in 17 steps · 9.2 min

  • Hole 1proved
  • Hole 2proved
  • Assemblyproved: our full check passed

Outline 2refuted: degenerate (a hole is the leaf’s own claim)written in 30 steps · 22.3 min

Outline 1refuted: degenerate (a hole is the leaf’s own claim)written in 23 steps · 12.5 min

p2m-d00cfc (N07): every task (6)
TaskOutlineTryResultStepsModel minTokens computedWith cache re-reads
Outline11Ended: degenerate outline2312.576k687k
Outline21Ended: degenerate outline3022.3103k1.3M
Outline31Outline written179.264k551k
Hole 1pre-pass did not close it31Proved92.115k129k
Hole 2pre-pass did not close it31Proved71.213k93k
Assembly31Proved00.000
All 6 tasks8647.3270k2.7M

Outline 3: 2 holes, all extracted as lemmas; the pre-pass of standard tactics closed 0 of 2 (192 s).

p2m-115f6c (N08)

Not proved: outline 3 was refuted (hole 9 is false)

Outline 3refuted: hole 9 is falsewritten in 99 steps · 66.1 min

  • Hole 1proved
  • Hole 2not proved in 2 tries (a sorry left, then a sorry left)
  • Hole 3proved
  • Hole 4proved
  • Hole 5proved
  • Hole 6proved
  • Hole 7proved
  • Hole 8try 1: a sorry left; try 2 not tried (outline refuted)
  • Hole 9false, disproved by the model

Outline 2abandoned: hole 3 not proved in 2 trieswritten in 57 steps · 72.3 min

  • Hole 1proved
  • Hole 2proved
  • Hole 3not proved in 2 tries (a sorry left, then a sorry left)

Outline 1refuted: holes 6 and 8 are falsewritten in 36 steps · 19.4 min

  • Hole 1closed by the pre-pass
  • Hole 2proved
  • Hole 3proved
  • Hole 4proved
  • Hole 5proved
  • Hole 6false, disproved by the model in try 2 (try 1: a sorry left)
  • Hole 7not proved in 2 tries (did not compile, then a sorry left)
  • Hole 8false, disproved by the model
p2m-115f6c (N08): every task (27)
TaskOutlineTryResultStepsModel minTokens computedWith cache re-reads
Outline11Outline written3619.4106k2.1M
Hole 2pre-pass did not close it11Proved113.214k144k
Hole 3pre-pass did not close it11Proved205.049k519k
Hole 4pre-pass did not close it11Proved308.647k768k
Hole 5pre-pass did not close it11Proved3113.469k1.1M
Hole 6pre-pass did not close it11Not proved: a sorry left12848.1234k8.8M
Hole 612Disproved: the hole is false9760.0298k7.5M
Hole 7pre-pass did not close it11Not proved: did not compile12867.0301k9.2M
Hole 712Not proved: a sorry left11295.4527k10.3M
Hole 8pre-pass did not close it11Disproved: the hole is false12849.7258k9.8M
Outline21Outline written5772.3241k5.0M
Hole 1pre-pass did not close it21Proved3828.2115k1.7M
Hole 2pre-pass did not close it21Proved2114.850k480k
Hole 3pre-pass did not close it21Not proved: a sorry left106116.2624k8.5M
Hole 322Not proved: a sorry left80129.2666k7.9M
Outline31Outline written9966.1241k6.4M
Hole 1pre-pass did not close it31Proved244.452k643k
Hole 2pre-pass did not close it31Not proved: a sorry left40130.2517k4.6M
Hole 232Not proved: a sorry left119109.7395k9.1M
Hole 3pre-pass did not close it31Proved3129.9125k1.5M
Hole 4pre-pass did not close it31Proved6626.9119k3.8M
Hole 5pre-pass did not close it31Proved6526.690k2.5M
Hole 6pre-pass did not close it31Proved10444.0200k6.0M
Hole 7pre-pass did not close it31Proved7433.5135k3.3M
Hole 8pre-pass did not close it31Not proved: a sorry left1959.5265k1.1M
Hole 832Cancelled––––
Hole 9pre-pass did not close it31Disproved: the hole is false136.339k409k
All 27 tasks16771267.65.8M113.2M

Outline 1: 8 holes, all extracted as lemmas; the pre-pass of standard tactics closed 1 of 8 (840 s).

Outline 2: 3 holes, all extracted as lemmas; the pre-pass of standard tactics closed 0 of 3 (262 s).

Outline 3: 9 holes, all extracted as lemmas; the pre-pass of standard tactics closed 0 of 9 (300 s).

How to read this

  • The Lean kernel is the judge. A run counts as proved only when our full check passes (the proof builds, its theorem is exactly the leaf’s, and it uses no sorry and only the axioms propext, Classical.choice and Quot.sound), with Lean v4.33.1 and Mathlib 0df444a as on prove2me. A submission counts as accepted only when prove2me’s verdict says so. We do not grade the model’s work ourselves.
  • Calibration and open leaves. A calibration leaf had already been proved by someone else when we chose it. It tells us whether the setup works; the model is not shown that proof, and we never submit one. An open leaf had no accepted proof when we chose it. An open leaf “closed by reuse” has the same statement as a theorem already proved on prove2me, or is a part of one: we submit the one-line proof that uses it, name its author, and do not count it as proved by the model.
  • Every run counts. A run is one leaf in one trial: a first round, and from trial 2 up to two fix rounds in which the model gets Lean’s errors. From trial 5 a run is all of a leaf’s tasks: its outlines, one task per hole, the tries and the assembly. Failed runs stay in the tables. A leaf tried again in a later trial counts as one more run, and once among the open leaves tried. Runs cut short by a fault on our side, not the model’s, or stopped by us because of one, say “Not scored” and why (4 so far); they still count among the runs.
  • Model minutes are the time the model spent generating, summed over a run’s rounds (from trial 5, its tasks). Tokens computed are the work the GPU did: what the model read that was not already in the cache, plus what it wrote. An agent reads the same files again and again, and the cache serves those re-reads: counted with them, the tokens are about 13 times as many (92% of everything read and written were re-reads). Hover a token count for both numbers. Two first tries in trial 2 that the gateway’s personal-data filter cut off are not in the counts.
  • Working on now and the trials from trial 5 show each leaf’s outlines, newest first. An outline is in progress, refuted (one of its holes is false, or its shape was refused), proved (its holes put together passed our full check) or abandoned for a new outline. Each hole is proved by its own task, false (the model disproved it), closed before any task by the pre-pass of standard tactics, by our index of prove2me’s library (reuse) or by the ledger (proved before), or it is running, queued, not tried or cancelled; the holes left in the outline’s file (“in place”) share one task. A cancelled task says why: its outline refuted, replaced or abandoned after a hole failed, or its leaf proved. A leaf’s tokens count every task, each outline and each try included: tokens computed as above, and the count with the cache re-reads beside them. Open “Every task” for each task’s result, steps, model minutes and tokens. A running task’s steps and tokens show as the gateway counts them where it can, else once the task ends.
  • Small samples. 48 runs in 7 trials, each with a different scaffold, budget and set of leaves. They show what happened, not a success rate.
  • These are our own measurements, from the records of each run, not estimates. The model runs on the AIs@home network, which today is mostly the maintainer’s own GPUs, some of them rented. The data is published as JSON: /science/prove2me.json.

What comes next

  • Trial 5 asked whether one task per hole closes holes that whole-leaf rounds could not. On F01 it did, once; on O02 it closed none before we stopped it.
  • Trial 8, above: whether its scaffold, described under “Working on now”, lets the model close easier open leaves, chosen by quick outline probes.
  • Then a random sample of open leaves at a fixed budget, with the pass line written down before the run: the test of whether this works at a useful rate, and what a proof costs.
  • The question behind it: can an open model on one consumer GPU close open leaves often enough that volunteers’ GPUs could do real work here?