Explicit Runge-Kutta tableaus scored end-to-end in Q15 at a fixed cycle budget.
The rk run ended on 2026-11-10, and no more compute goes to it. Epoch 2 is frozen and these pages are its final snapshot: the site is no longer rebuilt. The overview site sums up the final findings, and the harness is public for anyone to download and run.
Explicit Runge-Kutta methods build each stage from stages already computed, so a step is a fixed run of multiply-adds. For each order, stage count and cost, the grid keeps the verified tableau with the lowest held-out error, and that error, measured in Q15 at a fixed cycle budget, is the number each elite reports.
Built from cycle 17,364 of the run, 358,752 archive records. Newest elite recorded 2026-10-23 23:10 CT, after the last saturation check at 2026-09-27 17:23 CT.
Source: rk-work/archive, the elite in each cell when this page was built. Arithmetic: Q15 with floor rounding; cycles per step from the pinned m0plus_fast model.
| mark | order | stages | bucket | tableau | cycles per step | held-out error |
|---|---|---|---|---|---|---|
| archive elite | order 1 | 1 stage | bucket 0 | ff5d6ff6768c | 5 | 0.320268 |
| archive elite | order 2 | 2 stages | bucket 0 | 892bd5b10e39 | 14 | 0.0705737 |
| archive elite | order 2 | 2 stages | bucket 1 | 2d2c2b75d719 | 16 | 0.0530643 |
| archive elite | order 2 | 2 stages | bucket 2 | 9e7d5a6e28dd | 40 | 0.0560062 |
| archive elite | order 2 | 3 stages | bucket 0 | 6058c34391b1 | 13 | 0.0779237 |
| archive elite | order 2 | 3 stages | bucket 1 | 312c38c22850 | 29 | 0.0238649 |
| archive elite | order 2 | 3 stages | bucket 2 | ddfa9e5bdf2c | 60 | 0.0435731 |
| archive elite | order 2 | 3 stages | bucket 3 | 06bcc450bce4 | 64 | 0.0456747 |
| archive elite | order 2 | 4 stages | bucket 1 | 8b1c20657fe8 | 31 | 0.0216412 |
| archive elite | order 2 | 4 stages | bucket 2 | 3c3c46da6ed5 | 45 | 0.0158228 |
| archive elite | order 2 | 4 stages | bucket 3 | 9fef8aebd015 | 91 | 0.0362323 |
| archive elite | order 2 | 5 stages | bucket 2 | 8677ac52b717 | 60 | 0.0122299 |
| archive elite | order 2 | 5 stages | bucket 3 | ec6a1c6e2990 | 64 | 0.0135838 |
| archive elite | order 2 | 5 stages | bucket 4 | 6ede8aa89377 | 138 | 0.0260573 |
| archive elite | order 2 | 6 stages | bucket 2 | 977d34ecabe5 | 58 | 0.0129137 |
| archive elite | order 2 | 6 stages | bucket 3 | 4a62c0ce629f | 64 | 0.0135838 |
| archive elite | order 2 | 6 stages | bucket 4 | 1d4f752965ad | 163 | 0.0354374 |
| archive elite | order 3 | 3 stages | bucket 2 | 57af328f7be2 | 61 | 0.0410749 |
| archive elite | order 3 | 3 stages | bucket 3 | a4bfd55b3f91 | 68 | 0.0392167 |
| archive elite | order 3 | 4 stages | bucket 2 | 32f5efc1ebe8 | 61 | 0.0421996 |
| archive elite | order 3 | 4 stages | bucket 3 | cf7cbff1cb8a | 105 | 0.0281748 |
| archive elite | order 3 | 5 stages | bucket 2 | 62efd50c6fff | 63 | 0.0631408 |
| archive elite | order 3 | 5 stages | bucket 3 | 27688d49ea96 | 125 | 0.0255085 |
| archive elite | order 3 | 5 stages | bucket 4 | 5f506b0793cb | 156 | 0.0251585 |
| archive elite | order 3 | 6 stages | bucket 3 | 090b2e692801 | 100 | 0.0253993 |
| archive elite | order 3 | 6 stages | bucket 4 | ba6c87de14bc | 128 | 0.0228645 |
| archive elite | order 4 | 4 stages | bucket 2 | 016199da4761 | 49 | 0.0694907 |
| archive elite | order 4 | 5 stages | bucket 2 | 715e0970926e | 63 | 0.060571 |
| archive elite | order 4 | 5 stages | bucket 3 | 626d128171fb | 88 | 0.0371707 |
| archive elite | order 4 | 6 stages | bucket 3 | b619e3e62d5e | 104 | 0.0278667 |
| archive elite | order 4 | 6 stages | bucket 4 | 0d9f381b85c3 | 129 | 0.0366576 |
| classical baseline | euler | 5 | 0.320268 |
Order 1
Order 2
Order 3
Order 4
Source: rk-work/archive, the elite in each cell when this page was built. Arithmetic: Q15 with floor rounding; cycles per step from the pinned m0plus_fast model.
The archive is a MAP-Elites structure: one grid per algebraic order 1 to 4, whose cells are keyed by (stage count, cycle bucket). A cell holds at most one record, its elite: the verified tableau with the lowest held-out error seen so far for that shape and cost; ties keep the earlier record. An empty cell means no verified tableau has landed there yet.
Stage rows span 2 to 6 by default, plus a row for any other stage count that holds an elite (order 1's single-stage cell appears as s=1). The cycle buckets are log2-spaced bins of the m0plus_fast cycles per step: bucket 0 means fewer than 16 cycles, bucket 1 is 16 to 31, bucket 2 is 32 to 63, doubling each column, and bucket 7 collects everything at 1024 cycles or more.
One row per occupied cell, every order together so cells can be compared across orders. The hash links to the full record. A large gap between search_error and heldout_error is overfitting to the visible search set; measured_order and the tier badge are defined in the glossary. 1 of these rows is a seeded classical baseline rather than a search result; the label column says which.
| order | stages | bucket | tableau_hash | tier | heldout_error | search_error | cycles fast | cycles slow | measured_order | label |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | 1 | 0 | ff5d6ff6768c | no_incumbent | 0.320268 | 0.01289 | 5 | 5 | 0.984858 | seeded classical baseline |
| 2 | 2 | 0 | 892bd5b10e39 | heldout_verified | 0.0705737 | 0.0106168 | 14 | 14 | 2.01094 | exhaustive: optimal within the enumerated space |
| 2 | 2 | 1 | 2d2c2b75d719 | no_improvement | 0.0530643 | 0.0192404 | 16 | 16 | 2.01094 | exhaustive: optimal within the enumerated space |
| 2 | 2 | 2 | 9e7d5a6e28dd | heldout_verified | 0.0560062 | 0.00481262 | 40 | 40 | 2.01094 | search result |
| 2 | 3 | 0 | 6058c34391b1 | no_incumbent | 0.0779237 | 0.0177026 | 13 | 13 | 2.01094 | search result |
| 2 | 3 | 1 | 312c38c22850 | no_improvement | 0.0238649 | 0.00469231 | 29 | 29 | 2.01094 | search result |
| 2 | 3 | 2 | ddfa9e5bdf2c | no_improvement | 0.0435731 | 0.0123266 | 60 | 184 | 2.01997 | search result |
| 2 | 3 | 3 | 06bcc450bce4 | no_improvement | 0.0456747 | 0.0154514 | 64 | 126 | 2.01202 | search result |
| 2 | 4 | 1 | 8b1c20657fe8 | no_incumbent | 0.0216412 | 0.0102529 | 31 | 31 | 2.01094 | search result |
| 2 | 4 | 2 | 3c3c46da6ed5 | no_improvement | 0.0158228 | 0.00535362 | 45 | 45 | 2.01094 | search result |
| 2 | 4 | 3 | 9fef8aebd015 | heldout_verified | 0.0362323 | 0.00337466 | 91 | 184 | 1.99285 | search result |
| 2 | 5 | 2 | 8677ac52b717 | heldout_verified | 0.0122299 | 0.00264942 | 60 | 60 | 2.01094 | search result |
| 2 | 5 | 3 | ec6a1c6e2990 | no_improvement | 0.0135838 | 0.00584245 | 64 | 64 | 2.01094 | search result |
| 2 | 5 | 4 | 6ede8aa89377 | heldout_verified | 0.0260573 | 0.00444157 | 138 | 355 | 1.99631 | search result |
| 2 | 6 | 2 | 977d34ecabe5 | no_improvement | 0.0129137 | 0.00808785 | 58 | 58 | 2.01094 | search result |
| 2 | 6 | 3 | 4a62c0ce629f | no_improvement | 0.0135838 | 0.00584245 | 64 | 64 | 2.01094 | search result |
| 2 | 6 | 4 | 1d4f752965ad | no_improvement | 0.0354374 | 0.00894625 | 163 | 225 | 1.97865 | search result |
| 3 | 3 | 2 | 57af328f7be2 | no_improvement | 0.0410749 | 0.0157948 | 61 | 92 | 3.04041 | exhaustive: optimal within the enumerated space |
| 3 | 3 | 3 | a4bfd55b3f91 | heldout_verified | 0.0392167 | 0.0121371 | 68 | 68 | 3.04041 | exhaustive: optimal within the enumerated space |
| 3 | 4 | 2 | 32f5efc1ebe8 | no_improvement | 0.0421996 | 0.00598947 | 61 | 61 | 3.03441 | search result |
| 3 | 4 | 3 | cf7cbff1cb8a | no_improvement | 0.0281748 | 0.0141172 | 105 | 198 | 3.04079 | search result |
| 3 | 5 | 2 | 62efd50c6fff | heldout_verified | 0.0631408 | 0.00399856 | 63 | 94 | 4.07063 | search result |
| 3 | 5 | 3 | 27688d49ea96 | heldout_verified | 0.0255085 | 0.00937325 | 125 | 311 | 3.00774 | search result |
| 3 | 5 | 4 | 5f506b0793cb | heldout_verified | 0.0251585 | 0.00300234 | 156 | 311 | 3.02822 | search result |
| 3 | 6 | 3 | 090b2e692801 | no_improvement | 0.0253993 | 0.0102562 | 100 | 100 | 3.0246 | search result |
| 3 | 6 | 4 | ba6c87de14bc | no_improvement | 0.0228645 | 0.0122833 | 128 | 128 | 3.09694 | search result |
| 4 | 4 | 2 | 016199da4761 | no_improvement | 0.0694907 | 0.0201858 | 49 | 49 | 4.07063 | search result |
| 4 | 5 | 2 | 715e0970926e | no_improvement | 0.060571 | 0.0280446 | 63 | 63 | 4.07063 | search result |
| 4 | 5 | 3 | 626d128171fb | no_improvement | 0.0371707 | 0.0130353 | 88 | 88 | 4.07063 | search result |
| 4 | 6 | 3 | b619e3e62d5e | heldout_verified | 0.0278667 | 0.0021239 | 104 | 104 | 4.07063 | search result |
| 4 | 6 | 4 | 0d9f381b85c3 | no_incumbent | 0.0366576 | 0.00360065 | 129 | 222 | 4.08062 | search result |
Scored under one verifier hash: 2db0816c621c44aa5b4b8efbadfde3a37497780460be138cd8af5439ab176611 (31 rows). Each record's own hash is on its detail page.
rk-work/validation/results.json has not been written in this work directory, so the practical problems are not reported here.
rk-work/benchmark/results.json has not been written in this work directory, so the matched-accuracy comparison is not stated here. It appears once rk_harness.benchmark has been run against the current code; nothing is inferred in its place. The epoch 1 comparison, run under that epoch's own pin, is on the frozen epoch 1 site.