Key findings

What the explicit search found, at snapshot 2026-09-10.

At a fixed budget of 65,536 cycles on a Cortex-M0+ cost model, under magnitude weighting on the analytic cost basis, which prices the stage and b combinations only, the search found Runge-Kutta methods with up to 2.9508× lower held-out error than the best classical method. On that weighting and basis, dropping any one of the four held-out problems leaves a lead of 1.9236× to 3.1739×, so there the lead does not rest on one of them. Of the five other cells in the grid in finding 1, only the reference-norm scale on the analytic basis keeps both claims; in the other four a classical anchor leads on at least one reduced held-out set, and under median-anchor equal weighting on the traced whole-step basis it leads on the full set too. The data also shows why: Q15's floor rounding adds a bias of about half an LSB to every multiply, and that alone reorders the classical methods before any search starts. The demo lets you watch it happen.

Unless a figure says otherwise: a 65,536-cycle budget per problem, the m0plus_fast cost model, Q15 with floor rounding, the four held-out problems. Charts come from tools/key_findings.json; finding 6 reads rk-work/validation/results.json. Findings 1 to 5 were computed at 141,364 archive records, finding 6 and the method matrix at 141,364, and each figure carries the state it was computed at.

Every chart on this page is a static drawing, and the table folded under it, which holds the same numbers, is the keyboard and tap path to them.

The anchor result

rk4 and rk38 are both four-stage, order-4 methods with the same stability polynomial, and textbooks treat them as interchangeable. The analytic cost model does not. With a 1-cycle multiplier it prices rk4 cheaper (33 against 36 cycles per step), and with a 32-cycle multiplier it prices rk38 cheaper (64 against 85), because rk4's coefficients need more shift-add work there. That swap belongs to the frozen epoch-1 rule, which charges each coefficient the cheaper of a shift-add chain and a hardware multiply, and the compiled, traced step is what checks it. At the scope the model prices, the stage and b combinations and nothing else, the trace has rk38 at 59 cycles and rk4 at 67 under m0plus_fast, and 59 and 67 under m0plus_slow, so there rk38 is the cheaper of the two under both multiplier models and the fast half of the swap does not reproduce. Those counts come from an emulator that is instruction accurate rather than cycle accurate, and the findings validation page has every traced row.

Analytic cycles per step for rk4 and rk38 under the fast and slow multiplier cost models. Bar height is the per-step cost and the printed number is the exact cycle count. Which of the two anchor methods is cheaper swaps between the multiplier models, and that swap is the cost model's main sanity check.as of 2026-09-10 (US Central), from the code and data at build time
m0plus_fast (1-cycle multiplier)m0plus_slow (32-cycle multiplier)
rk4 under m0plus_fast: 33 cycles per step33rk4 under m0plus_slow: 85 cycles per step85rk4rk38 under m0plus_fast: 36 cycles per step36rk38 under m0plus_slow: 64 cycles per step64rk38
Every plotted bar, in one table (4 rows)
methodcost modelcycles per step
rk4m0plus_fast33
rk4m0plus_slow85
rk38m0plus_fast36
rk38m0plus_slow64

1Discovered methods lead in all 16 cells

The archive keeps one elite per MAP-Elites cell (order, stage count, cycle-cost bucket). 18 cells are occupied: 16 by discovered methods and 2 by classical ones (euler and rk38). Each discovered cell is compared with every classical method of equal or lower per-step cost, since a cheaper method could always be swapped in. All 16 discovered cells have the lower held-out error. The median error ratio, discovered over classical, is 0.71, and the best is 0.34.

Per-step cost against held-out error at the shared 65,536-cycle budget, both on log scales, so down and left is better. Every blue dot under the dashed line has lower error than every classical method. Cycle counts are for one state (m0plus_fast, n_states=1); cost scales with the problem's state dimension, so the same tableau costs 44 cycles per step on a two-state problem.data: key_findings.json, series frontier_cycles_vs_heldoutas of 2026-09-10 (US Central), computed at 141,364 archive records
discovered (best per grid cell)classical methods
5102040800.020.050.10.20.4cycles per step at one state (log)held-out error (log)best classical method (0.0844)euler (classical): 5 cycles/step, held-out error 0.32eulermidpoint (classical): 11 cycles/step, held-out error 0.0844midpointheun2 (classical): 13 cycles/step, held-out error 0.0961heun2ralston2 (classical): 16 cycles/step, held-out error 0.0919ralston2heun3 (classical): 23 cycles/step, held-out error 0.0852heun3kutta3 (classical): 26 cycles/step, held-out error 0.0873kutta3rk4 (classical): 33 cycles/step, held-out error 0.0919rk4rk38 (classical): 36 cycles/step, held-out error 0.0853rk38discovered, order 2, 2 stages: 15 cycles/step, held-out error 0.0613, tier no_improvementdiscovered, order 2, 2 stages: 17 cycles/step, held-out error 0.0494, tier no_improvementdiscovered, order 2, 3 stages: 22 cycles/step, held-out error 0.0286, tier no_improvementdiscovered, order 3, 3 stages: 29 cycles/step, held-out error 0.0477, tier no_improvementdiscovered, order 2, 4 stages: 30 cycles/step, held-out error 0.0836, tier no_incumbentdiscovered, order 2, 4 stages: 39 cycles/step, held-out error 0.0587, tier heldout_verifieddiscovered, order 3, 4 stages: 46 cycles/step, held-out error 0.0651, tier heldout_verifieddiscovered, order 2, 5 stages: 57 cycles/step, held-out error 0.0533, tier heldout_verifieddiscovered, order 3, 5 stages: 59 cycles/step, held-out error 0.0587, tier no_improvementdiscovered, order 4, 6 stages: 63 cycles/step, held-out error 0.0603, tier no_incumbentdiscovered, order 3, 6 stages: 63 cycles/step, held-out error 0.0822, tier no_incumbentdiscovered, order 3, 5 stages: 65 cycles/step, held-out error 0.0604, tier no_improvementdiscovered, order 2, 5 stages: 65 cycles/step, held-out error 0.0639, tier no_improvementdiscovered, order 2, 6 stages: 72 cycles/step, held-out error 0.0453, tier heldout_verifieddiscovered, order 3, 6 stages: 75 cycles/step, held-out error 0.0377, tier heldout_verifieddiscovered, order 4, 6 stages: 77 cycles/step, held-out error 0.0808, tier no_incumbentbest discovered: 0.0286 at 22 cycles
The same points as a table (24 rows)
methodkindorderstagescycles/stepheld-out errortier
eulerclassical1150.32no_incumbent
midpointclassicaln/a2110.0844no_incumbent
heun2classicaln/a2130.0961no_incumbent
discovered 2d2c2b75d719discovered22150.0613no_improvement
ralston2classicaln/a2160.0919no_incumbent
discovered 4feb237eaaa3discovered22170.0494no_improvement
discovered 11e898cbab59discovered23220.0286no_improvement
heun3classicaln/a3230.0852no_incumbent
kutta3classicaln/a3260.0873no_incumbent
discovered 4f501ec6a5cediscovered33290.0477no_improvement
discovered 5fef0bd2b5fediscovered24300.0836no_incumbent
rk4classicaln/a4330.0919no_incumbent
rk38classical44360.0853no_incumbent
discovered 2bceb12dc1c2discovered24390.0587heldout_verified
discovered 19b66f506eecdiscovered34460.0651heldout_verified
discovered 6d0004de713fdiscovered25570.0533heldout_verified
discovered 274563728359discovered35590.0587no_improvement
discovered 196b1d17d0b7discovered46630.0603no_incumbent
discovered c10bc5417b5bdiscovered36630.0822no_incumbent
discovered 407c65d76723discovered35650.0604no_improvement
discovered 935100ec0e46discovered25650.0639no_improvement
discovered aea95967c0fddiscovered26720.0453heldout_verified
discovered 42863b93bf7ddiscovered36750.0377heldout_verified
discovered 013ba3020051discovered46770.0808no_incumbent

Held-out error here is the published figure, magnitude weighting on the analytic basis. The tier is read from the archive at build time: for a method that holds a cell it is the stored tier of that cell's elite, for a classical method that holds no cell it is the stored tier of its one seed record, and a method with no archive record prints n/a.

Which of the 136 searchable MAP-Elites cells hold an elite: 17 are occupied, 16 by discovered methods and 1 by a seeded classical method. 1 more cell (euler, at one stage) sits outside the searched range. The upper cost buckets are empty by construction, since no step at 2 to 6 stages costs that much under the fast multiplier.data: rk-work archive, MAP-Elites cells at generation timeas of 2026-09-10 (US Central), 141,364 archive records
held by a discovered methodheld by a classical method
order 1s1order 1, 1 stage, bucket 0: seeded baseline, outside the searched stage ranges2s3s4s5s6b0b7order 2s2order 2, 2 stages, bucket 0: discovered 2d2c2b75d719, held-out error 0.0613order 2, 2 stages, bucket 1: discovered 4feb237eaaa3, held-out error 0.0494s3order 2, 3 stages, bucket 1: discovered 11e898cbab59, held-out error 0.0286s4order 2, 4 stages, bucket 1: discovered 5fef0bd2b5fe, held-out error 0.0836order 2, 4 stages, bucket 2: discovered 2bceb12dc1c2, held-out error 0.0587s5order 2, 5 stages, bucket 2: discovered 6d0004de713f, held-out error 0.0533order 2, 5 stages, bucket 3: discovered 935100ec0e46, held-out error 0.0639s6order 2, 6 stages, bucket 3: discovered aea95967c0fd, held-out error 0.0453b0b7order 3s2s3order 3, 3 stages, bucket 1: discovered 4f501ec6a5ce, held-out error 0.0477s4order 3, 4 stages, bucket 2: discovered 19b66f506eec, held-out error 0.0651s5order 3, 5 stages, bucket 2: discovered 274563728359, held-out error 0.0587order 3, 5 stages, bucket 3: discovered 407c65d76723, held-out error 0.0604s6order 3, 6 stages, bucket 2: discovered c10bc5417b5b, held-out error 0.0822order 3, 6 stages, bucket 3: discovered 42863b93bf7d, held-out error 0.0377b0b7order 4s2s3s4order 4, 4 stages, bucket 2: classical baseline rk38, held-out error 0.0853s5s6order 4, 6 stages, bucket 2: discovered 196b1d17d0b7, held-out error 0.0603order 4, 6 stages, bucket 3: discovered 013ba3020051, held-out error 0.0808b0b7columns are cycle-cost buckets b0..b7 (log2 bands of cycles/step, m0plus_fast); rows are stage counts
The occupied cells as a table (18 rows)
orderstagescycle bucketheld byheld-out errorcounted
11b0seeded euler0.32outside the searched stage range
22b0discovered 2d2c2b75d7190.0613counted in the fraction
22b1discovered 4feb237eaaa30.0494counted in the fraction
23b1discovered 11e898cbab590.0286counted in the fraction
24b1discovered 5fef0bd2b5fe0.0836counted in the fraction
24b2discovered 2bceb12dc1c20.0587counted in the fraction
25b2discovered 6d0004de713f0.0533counted in the fraction
25b3discovered 935100ec0e460.0639counted in the fraction
26b3discovered aea95967c0fd0.0453counted in the fraction
33b1discovered 4f501ec6a5ce0.0477counted in the fraction
34b2discovered 19b66f506eec0.0651counted in the fraction
35b2discovered 2745637283590.0587counted in the fraction
35b3discovered 407c65d767230.0604counted in the fraction
36b2discovered c10bc5417b5b0.0822counted in the fraction
36b3discovered 42863b93bf7d0.0377counted in the fraction
44b2classical baseline rk380.0853counted in the fraction
46b2discovered 196b1d17d0b70.0603counted in the fraction
46b3discovered 013ba30200510.0808counted in the fraction

The best discovered method is a 3-stage, order-2 tableau at 22 cycles per step for one state (m0plus_fast, n_states=1), with held-out error 0.0286 against 0.0844 for midpoint, the best classical method. Cost scales with the problem's state dimension, so the same tableau costs 44 cycles per step on a two-state problem. Every coefficient is dyadic, so the method needs only shifts and adds.

Its tableau and its error on each held-out problem

A = [[0, 0, 0], [13/16, 0, 0], [5/32, 11/32, 0]], b = (0, 0, 1), c = (0, 13/16, 1/2). Measured order 2.02. Held-out error per problem, each with its share of the held-out sum of squares: pendulum 0.00474 (0.69%), dc_motor 0.0276 (23.33%), rc_thermal (flagged) 0.0461 (64.84%), quaternion 0.0191 (11.14%). rc_thermal is marked flagged because the degeneracy detector flags it over the 9 methods finding 1's counterfactual scores on the analytic basis (finding 2 gives every verdict and its reason).

Dropping one held-out problem at a time

Held-out error is the root mean square of the four per-problem errors, so a problem's influence grows with the square of its error and the four do not carry a quarter each. For the champion, rc_thermal carries 64.8% of the held-out sum of squares and pendulum 0.69%; for midpoint it is rc_thermal at 85.1%. Which problem dominates is a property of the method rather than of the metric: euler's largest share is pendulum at 86.1%. rc_thermal is the problem where every classical method lands on the same quantization floor (finding 4). Under magnitude weighting on the analytic basis (the stage and b combinations only), dropping rc_thermal leaves 1.9236×, the lowest ratio in the table below; the highest is 3.1739×, with dc_motor dropped, and the best classical method changes from midpoint to rk38 when quaternion goes.

held-out problem droppedchampionbest classical methodratio
none, all four kept0.0286midpoint 0.08442.9508×
pendulum0.0329midpoint 0.09752.9610×
dc_motor0.0289midpoint 0.09183.1739×
rc_thermal0.0196midpoint 0.03771.9236×
quaternion0.0311rk38 0.0953.0499×

Held-out error is an RMS, so each row is recomputed over the three problems it keeps, under magnitude weighting on the analytic basis, from tools/key_findings.json, efficiency.numbers.leave_one_out.

Two caveats. Elites are chosen by held-out error from 141,364 archived candidates, so the best values carry selection bias. The outer loop is shown one held-out error per occupied cell; only the inner optimizer is scored on search-set error alone. And a tier is a mechanical comparison with the cell's incumbent when the record is written, not a validation grade. The best cell's elite is stored as no_improvement, its cell had an incumbent and neither the heldout_verified nor the search_only rule applied. Of the 16 discovered elites, 5 are heldout_verified.

The same comparison under two other weightings and a second cost basis

The ratio above is one cell of a grid. It is measured under magnitude weighting, the aggregate the archive selects on (an unweighted RMS of the four per-problem errors), and on the analytic cost basis, the per-step cost the analytic cost model charges for the stage and b combinations. The grid adds two other weightings and one other cost basis, and all six combinations are below. Every archived per-problem error is already normalized before any weighting: the distance problems divide by their own peak amplitude, and pendulum and quaternion measure a relative drift. The median-anchor scale then divides each error by the median error of the eight classical anchors on that problem, and it is the equal-weighting answer, named that way wherever it appears. The reference-norm scale divides instead by the error a method would report if its state collapsed to zero, which is the size of the reference solution on that same normalized scale. It is published beside the median-anchor scale and never alone, because it does not equalize the weights: rc_thermal's reference norm is the smallest of the four held-out problems', so that scale lifts rc_thermal's share of the champion's sum of squares from 64.8% to 98.7% and midpoint's from 85.1% to 99.6%. The two scales disagree about the headline, which is why neither is printed without the other.

The second basis prices a step at what the compiled code executes, read from the traced document rather than from the model, and it changes the step count rather than the arithmetic. On damped_osc, a two-state problem, euler takes 1,598 steps inside the budget where the model gave it 6,553, and the champion 455 where the model gave it 1,489. Every method loses steps on every problem, and no two of the 9 lose the same fraction on damped_osc, so the field is reshuffled rather than shifted.

weightingcost basischampionbest classical anchorratio, anchor over championlowest leave-one-out ratioreduced sets where an anchor leadsahead on the full set
magnitude (published)analytic (published)0.0286midpoint 0.08442.9508x1.9236x (drop rc_thermal)0 of 4champion
magnitude (published)traced whole step0.0151heun3 0.01521.0112x0.6143x (drop rc_thermal)3 of 4champion
equal, median anchoranalytic (published)0.653midpoint 0.7191.1020x0.8144x (drop rc_thermal)3 of 4champion
equal, median anchortraced whole step3.16midpoint 0.1560.0494x0.0314x (drop dc_motor)3 of 4anchor
reference normanalytic (published)0.146midpoint 0.493.3646x1.9200x (drop rc_thermal)0 of 4champion
reference normtraced whole step0.0424heun3 0.0781.8386x0.6042x (drop rc_thermal)1 of 4champion

Each row is one cell of the grid, and a ratio above 1.0 means the champion has the lower error. The analytic basis prices the stage and b combinations only; the traced whole-step basis buys steps at the count for the whole rk_step with the derivative body excluded. Read from tools/key_findings.json, counterfactual.numbers.heldout.

Both published figures are measured under magnitude weighting on the analytic basis (the stage and b combinations only): the 2.9508x lead, and the 1.9236x floor with rc_thermal dropped, which is the claim that the champion stays ahead on every reduced set. Taken one at a time, one of the five other cells keeps both and four do not:

Two cells change the weighting and the cost basis together, moving to the traced whole-step basis (the whole rk_step, derivative body excluded). Under median-anchor equal weighting there the ratio reads 0.0494x. pendulum has the smallest median anchor error of the four held-out problems at 0.00412, so dividing by it amplifies any method whose pendulum error rises when steps are taken away. The champion's pendulum error goes from 0.00474 at 1,489 steps on the analytic basis to 0.026 at 455 steps on the traced one, which is most of why the median anchor ratio falls below one there. Read that cell as the metric magnifying one problem, not as the champion losing by that factor. Under the reference-norm scale on the same traced whole-step basis it reads 1.8386x, a classical anchor leads on 1 of the 4 reduced sets, and the floor is 0.6042x once rc_thermal is dropped.

Every leave-one-out row in the grid, as a table (30 rows)
weightingcost basisdroppedchampionbest classical anchorits errorratiorunner-up anchor
magnitude (published)analytic (published)none, all four kept0.0286midpoint0.08442.9508xheun3
magnitude (published)analytic (published)dc_motor0.0289midpoint0.09183.1739xheun3
magnitude (published)analytic (published)pendulum0.0329midpoint0.09752.9610xheun3
magnitude (published)analytic (published)quaternion0.0311rk380.0953.0499xkutta3
magnitude (published)analytic (published)rc_thermal0.0196midpoint0.03771.9236xheun3
magnitude (published)traced whole stepnone, all four kept0.0151heun30.01521.0112xmidpoint
magnitude (published)traced whole stepdc_motor0.0169heun30.01560.9229xmidpoint
magnitude (published)traced whole steppendulum0.00875heun30.01691.9323xmidpoint
magnitude (published)traced whole stepquaternion0.0173heun30.01710.9912xmidpoint
magnitude (published)traced whole steprc_thermal0.0157midpoint0.009670.6143xrk38
equal, median anchoranalytic (published)none, all four kept0.653midpoint0.7191.1020xrk38
equal, median anchoranalytic (published)dc_motor0.715midpoint0.6690.9362xrk38
equal, median anchoranalytic (published)pendulum0.356midpoint0.8312.3307xheun3
equal, median anchoranalytic (published)quaternion0.727kutta30.7220.9933xrk38
equal, median anchoranalytic (published)rc_thermal0.734midpoint0.5980.8144xrk38
equal, median anchortraced whole stepnone, all four kept3.16midpoint0.1560.0494xrk38
equal, median anchortraced whole stepdc_motor3.64midpoint0.1140.0314xrk38
equal, median anchortraced whole steppendulum0.0875heun30.171.9392xmidpoint
equal, median anchortraced whole stepquaternion3.64midpoint0.1720.0473xrk38
equal, median anchortraced whole steprc_thermal3.64midpoint0.1480.0407xrk38
reference normanalytic (published)none, all four kept0.146midpoint0.493.3646xheun3
reference normanalytic (published)dc_motor0.168midpoint0.5653.3738xheun3
reference normanalytic (published)pendulum0.168midpoint0.5663.3650xheun3
reference normanalytic (published)quaternion0.168midpoint0.5663.3700xheun3
reference normanalytic (published)rc_thermal0.0193midpoint0.03711.9200xheun3
reference normtraced whole stepnone, all four kept0.0424heun30.0781.8386xmidpoint
reference normtraced whole stepdc_motor0.0488heun30.08971.8380xmidpoint
reference normtraced whole steppendulum0.0466heun30.08991.9290xmidpoint
reference normtraced whole stepquaternion0.0489heun30.091.8384xmidpoint
reference normtraced whole steprc_thermal0.0157midpoint0.00950.6042xrk38
Every scored method's held-out error on both cost bases (9 rows)
methodkindweightingheld-out RMS, analytic (published)held-out RMS, traced whole step
11e898cbdiscoveredmagnitude (published)0.02860.0151
midpointclassicalmagnitude (published)0.08440.0161
heun3classicalmagnitude (published)0.08520.0152
rk38classicalmagnitude (published)0.08530.0179
kutta3classicalmagnitude (published)0.08730.0216
ralston2classicalmagnitude (published)0.09190.0223
rk4classicalmagnitude (published)0.09190.0204
heun2classicalmagnitude (published)0.09610.0261
eulerclassicalmagnitude (published)0.321.5

Each value is the root mean square of one method's four held-out errors under magnitude weighting. The analytic column prices the stage and b combinations only. The traced column buys steps at the count for the whole rk_step with the derivative body excluded, a budget denominator taken from an emulator that is instruction accurate and not cycle accurate. The two columns are errors at two step counts and are never divided by each other. Read from counterfactual.series.per_method.

Every per-problem error behind the grid, with its step count and its shares (126 rows)
cost basissetmethodproblemstatescycles/stepstepserrorshare, magnitudeshare, median anchorshare, reference norm
analytic (published)heldout11e898cbdc_motor2441,4890.027623.33%10.12%0.86%
analytic (published)heldout11e898cbpendulum2441,4890.004740.69%77.65%0.026%
analytic (published)heldout11e898cbquaternion4887440.019111.14%7.11%0.43%
analytic (published)heldout11e898cbrc_thermal (flagged)3669920.046164.84%5.12%98.68%
analytic (published)heldouteulerdc_motor2106,5530.1717.13%0.032%2.09%
analytic (published)heldouteulerpendulum2106,5530.59486.08%99.96%26.27%
analytic (published)heldouteulerquaternion4203,2760.05960.87%0.0057%0.26%
analytic (published)heldouteulerrc_thermal (flagged)3154,3690.1565.92%0.0048%71.37%
analytic (published)heldoutheun2dc_motor2262,5200.097425.69%40.98%0.94%
analytic (published)heldoutheun2pendulum2262,5200.004210.048%19.99%0.0018%
analytic (published)heldoutheun2quaternion4521,2600.05598.47%19.90%0.32%
analytic (published)heldoutheun2rc_thermal (flagged)3391,6800.15665.79%19.13%98.74%
analytic (published)heldoutheun3dc_motor2461,4240.054310.17%8.58%0.29%
analytic (published)heldoutheun3pendulum2461,4240.009710.33%71.49%0.0098%
analytic (published)heldoutheun3quaternion4927120.04065.67%7.04%0.17%
analytic (published)heldoutheun3rc_thermal (flagged)3699490.15683.84%12.89%99.53%
analytic (published)heldoutkutta3dc_motor2521,2600.04827.63%20.22%0.22%
analytic (published)heldoutkutta3pendulum2521,2600.0003830.00048%0.33%1.5e-05%
analytic (published)heldoutkutta3quaternion41046300.055710.18%39.72%0.31%
analytic (published)heldoutkutta3rc_thermal (flagged)3788400.15882.20%39.72%99.46%
analytic (published)heldoutmidpointdc_motor2222,9780.056811.30%35.14%0.32%
analytic (published)heldoutmidpointpendulum2222,9786.45e-051.5e-05%0.012%4.3e-07%
analytic (published)heldoutmidpointquaternion4441,4890.03223.64%16.66%0.11%
analytic (published)heldoutmidpointrc_thermal (flagged)3331,9850.15685.06%48.19%99.57%
analytic (published)heldoutralston2dc_motor2322,0480.078818.40%29.30%0.62%
analytic (published)heldoutralston2pendulum2322,0480.004730.066%27.54%0.0023%
analytic (published)heldoutralston2quaternion4641,0240.05669.50%22.27%0.33%
analytic (published)heldoutralston2rc_thermal (flagged)3481,3650.15672.03%20.90%99.05%
analytic (published)heldoutrk38dc_motor2729100.0529.30%26.01%0.27%
analytic (published)heldoutrk38pendulum2729100.00110.0042%3.04%0.00013%
analytic (published)heldoutrk38quaternion41444550.04476.86%28.24%0.21%
analytic (published)heldoutrk38rc_thermal (flagged)31086060.15683.83%42.72%99.52%
analytic (published)heldoutrk4dc_motor2669920.076417.26%30.81%0.56%
analytic (published)heldoutrk4pendulum2669920.004030.048%22.39%0.0016%
analytic (published)heldoutrk4quaternion41324960.05418.65%22.73%0.29%
analytic (published)heldoutrk4rc_thermal (flagged)3996610.15874.04%24.08%99.14%
analytic (published)search11e898cbdahlquist (flagged)1222,9780.0001990.01%94.46%99.74%
analytic (published)search11e898cbdamped_osc2441,4890.003943.93%2.51%0.26%
analytic (published)search11e898cbvanderpol_mild2441,4890.019596.06%3.03%0.0033%
analytic (published)searcheulerdahlquist (flagged)1513,1074.54e-050.00041%12.36%60.64%
analytic (published)searcheulerdamped_osc2106,5530.014240.48%81.69%39.33%
analytic (published)searcheulervanderpol_mild2106,5530.017259.52%5.95%0.03%
analytic (published)searchheun2dahlquist (flagged)1135,0414.54e-050.00023%29.15%90.89%
analytic (published)searchheun2damped_osc2262,5200.005543.38%29.33%8.98%
analytic (published)searchheun2vanderpol_mild2262,5200.029696.62%41.53%0.13%
analytic (published)searchheun3dahlquist (flagged)1232,8490.0001670.0035%83.81%99.00%
analytic (published)searchheun3damped_osc2461,4240.006525.29%8.57%0.99%
analytic (published)searchheun3vanderpol_mild2461,4240.027694.71%7.61%0.0093%
analytic (published)searchkutta3dahlquist (flagged)1262,5200.0001670.00031%46.96%99.22%
analytic (published)searchkutta3damped_osc2521,2600.005350.32%3.24%0.67%
analytic (published)searchkutta3vanderpol_mild2521,2600.094399.68%49.80%0.11%
analytic (published)searchmidpointdahlquist (flagged)1115,9577.67e-050.0011%56.10%95.18%
analytic (published)searchmidpointdamped_osc2222,9780.006688.61%28.77%4.79%
analytic (published)searchmidpointvanderpol_mild2222,9780.021891.39%15.13%0.026%
analytic (published)searchralston2dahlquist (flagged)1164,0964.54e-050.00012%9.10%57.09%
analytic (published)searchralston2damped_osc2322,0480.015313.82%69.45%42.77%
analytic (published)searchralston2vanderpol_mild2322,0480.038186.18%21.45%0.14%
analytic (published)searchrk38dahlquist (flagged)1361,8204.54e-050.00014%18.44%82.16%
analytic (published)searchrk38damped_osc2729100.008174.63%40.35%17.65%
analytic (published)searchrk38vanderpol_mild2729100.037195.37%41.21%0.19%
analytic (published)searchrk4dahlquist (flagged)1331,9850.0001670.00027%35.38%94.24%
analytic (published)searchrk4damped_osc2669920.01592.43%21.60%5.65%
analytic (published)searchrk4vanderpol_mild2669920.10197.57%43.02%0.12%
traced whole stepheldout11e898cbdc_motor21444550.007335.93%0.03%0.72%
traced whole stepheldout11e898cbpendulum21444550.02674.66%99.94%9.40%
traced whole stepheldout11e898cbquaternion44901330.003531.38%0.01%0.17%
traced whole stepheldout11e898cbrc_thermal32972200.012818.04%0.017%89.70%
traced whole stepheldouteulerdc_motor2411,5980.02960.0097%3.7e-05%0.0091%
traced whole stepheldouteulerpendulum2411,598399.85%100.00%97.10%
traced whole stepheldouteulerquaternion41295080.07580.064%0.00036%0.062%
traced whole stepheldouteulerrc_thermal3521,2600.08140.074%5.2e-05%2.83%
traced whole stepheldoutheun2dc_motor2927120.026425.61%6.41%0.89%
traced whole stepheldoutheun2pendulum2927120.00611.37%89.02%0.049%
traced whole stepheldoutheun2quaternion43052140.01023.84%1.42%0.14%
traced whole stepheldoutheun2rc_thermal31604090.043469.18%3.15%98.92%
traced whole stepheldoutheun3dc_motor21643990.014221.65%1.09%0.79%
traced whole stepheldoutheun3pendulum21643990.008337.48%97.92%0.28%
traced whole stepheldoutheun3quaternion45431200.006975.25%0.39%0.20%
traced whole stepheldoutheun3rc_thermal33631800.024765.63%0.60%98.72%
traced whole stepheldoutkutta3dc_motor21873500.016915.23%0.19%0.77%
traced whole stepheldoutkutta3pendulum21873500.024131.21%99.58%1.65%
traced whole stepheldoutkutta3quaternion45821120.01146.99%0.13%0.37%
traced whole stepheldoutkutta3rc_thermal34191560.029546.56%0.10%97.21%
traced whole stepheldoutmidpointdc_motor2748850.01624.74%59.63%0.82%
traced whole stepheldoutmidpointpendulum2748853.67e-050.00013%0.081%4.5e-06%
traced whole stepheldoutmidpointquaternion42962210.004912.33%8.26%0.08%
traced whole stepheldoutmidpointrc_thermal31444550.027572.93%32.03%99.10%
traced whole stepheldoutralston2dc_motor21165640.020821.63%0.44%0.94%
traced whole stepheldoutralston2pendulum21165640.019318.66%99.22%0.84%
traced whole stepheldoutralston2quaternion43621810.009484.51%0.14%0.20%
traced whole stepheldoutralston2rc_thermal32642480.033255.20%0.21%98.02%
traced whole stepheldoutrk38dc_motor22352780.014716.86%39.47%0.54%
traced whole stepheldoutrk38pendulum22352780.0002360.0044%2.67%0.00015%
traced whole stepheldoutrk38quaternion4785830.009727.42%25.56%0.25%
traced whole stepheldoutrk38rc_thermal35321230.031175.72%32.30%99.22%
traced whole stepheldoutrk4dc_motor22562560.018821.30%5.52%0.72%
traced whole stepheldoutrk4pendulum22562560.004671.31%88.74%0.046%
traced whole stepheldoutrk4quaternion4716910.01026.23%2.38%0.22%
traced whole stepheldoutrk4rc_thermal34781370.034471.15%3.36%99.01%
traced whole stepsearch11e898cbdahlquist (flagged)1867620.0001991.27%99.91%99.99%
traced whole stepsearch11e898cbdamped_osc21444550.00062812.71%0.067%0.0066%
traced whole stepsearch11e898cbvanderpol_mild21444550.0016386.02%0.023%2.3e-05%
traced whole stepsearcheulerdahlquist (flagged)1322,0484.54e-059.7e-05%12.25%72.00%
traced whole stepsearcheulerdamped_osc2411,5980.0115.67%48.12%27.76%
traced whole stepsearcheulervanderpol_mild2411,5980.044794.33%39.63%0.24%
traced whole stepsearchheun2dahlquist (flagged)1551,1914.54e-050.00068%57.51%97.50%
traced whole stepsearchheun2damped_osc2927120.00282.57%14.75%2.45%
traced whole stepsearchheun2vanderpol_mild2927120.017297.43%27.75%0.048%
traced whole stepsearchheun3dahlquist (flagged)1937044.54e-050.00084%62.81%97.99%
traced whole stepsearchheun3damped_osc21643990.00252.56%12.89%1.97%
traced whole stepsearchheun3vanderpol_mild21643990.015497.44%24.30%0.039%
traced whole stepsearchkutta3dahlquist (flagged)11026424.54e-050.00055%61.47%99.70%
traced whole stepsearchkutta3damped_osc21873500.0008680.20%1.52%0.24%
traced whole stepsearchkutta3vanderpol_mild21873500.019399.80%37.01%0.062%
traced whole stepsearchmidpointdahlquist (flagged)1501,3107.67e-050.00091%70.18%99.38%
traced whole stepsearchmidpointdamped_osc2748850.002280.80%4.17%0.58%
traced whole stepsearchmidpointvanderpol_mild2748850.025399.20%25.64%0.037%
traced whole stepsearchralston2dahlquist (flagged)1669924.54e-050.0004%40.91%93.99%
traced whole stepsearchralston2damped_osc21165640.004433.83%26.33%5.94%
traced whole stepsearchralston2vanderpol_mild21165640.022296.17%32.75%0.077%
traced whole stepsearchrk38dahlquist (flagged)11305040.000290.031%98.11%99.91%
traced whole stepsearchrk38damped_osc22352780.003354.19%0.89%0.089%
traced whole stepsearchrk38vanderpol_mild22352780.01695.78%1.01%0.0011%
traced whole stepsearchrk4dahlquist (flagged)11394710.0001670.0064%93.71%99.84%
traced whole stepsearchrk4damped_osc22562560.002541.46%1.45%0.15%
traced whole stepsearchrk4vanderpol_mild22562560.020898.53%4.84%0.0053%

(flagged) marks a problem the degeneracy detector flags over the 9 methods in this table on that row's own cost basis: on the analytic basis it flags dahlquist and rc_thermal; on the traced whole-step basis it flags dahlquist (finding 2 gives every verdict and its reason).

What the search would have kept under the other two weightings

The champion was selected under magnitude weighting, so a ratio for that same tableau under another weighting describes the metric and not the search. Every archived record stores its four held-out errors, so re-aggregating all of them answers the other question directly, with no step count and no solve changed. Under magnitude weighting the champion is rank 1 of 141,364; under the median-anchor scale it is rank 9,974, and the tableau that would have won has 6 stages and costs 73 cycles per step; under the reference-norm scale it is rank 15, and the tableau that would have won has 2 stages and costs 17 cycles per step. No classical anchor is ahead of the champion under any of the three, so neither the median-anchor scale nor the reference-norm scale changes whether a discovered tableau would have won, only which one. This is an analytic-basis result by construction: re-pricing steps for 141,364 records would mean replaying the archive rather than re-aggregating it, so it does not say what the search would have found on a part.

weightingtableau that would have wonits cycles/stepits errorthe published champion's rankclassical anchors ahead of it
magnitude (published)11e898cbab59 (the published champion)220.02861 of 141,3640
equal, median anchorb1a8c184b541730.4189,974 of 141,3640
reference norm4feb237eaaa3170.080115 of 141,3640

Every archived record re-aggregated under each weighting and reduced to one row per unique tableau, on the analytic cost basis, from tools/key_findings.json, counterfactual.numbers.archive_reranking.

The ten best tableaus under each weighting (30 rows)
weightingranktableaustagescycles/steptiermeasured orderheld-out errorthe published champion
magnitude (published)111e898cbab59322no_improvement2.0180.0286yes
magnitude (published)242863b93bf7d675heldout_verified3.0410.0377
magnitude (published)36fd075d05984675no_improvement3.0410.0384
magnitude (published)4a93e22525856676heldout_verified3.0410.042
magnitude (published)583a12de98425318search_only2.0110.0424
magnitude (published)68e9f832cef67683no_improvement3.0350.0436
magnitude (published)7b1a8c184b541673no_improvement3.0730.0445
magnitude (published)8b5cf281cc448675no_improvement3.0110.045
magnitude (published)9aea95967c0fd672heldout_verified1.9890.0453
magnitude (published)1070b5da1e1bda673no_improvement3.0410.0462
equal, median anchor1b1a8c184b541673no_improvement3.0730.418
equal, median anchor26b7e429a36fd676no_improvement3.0260.432
equal, median anchor30b1d0c191699674no_improvement3.0030.44
equal, median anchor49bbd04f081b1673no_improvement3.0270.442
equal, median anchor5de4bc8f53f01673no_improvement3.0270.446
equal, median anchor6d8f394874283674no_improvement3.1040.447
equal, median anchor7296cd47861a6670no_improvement3.0300.447
equal, median anchor82d21aaf11d93671no_improvement3.0250.448
equal, median anchor93bcc9564053a676no_improvement3.0440.448
equal, median anchor104a9a0c03cf41676no_improvement3.0050.45
reference norm14feb237eaaa3217no_improvement2.0110.0801
reference norm2757d041459cf217no_improvement2.0110.081
reference norm3f7e5852d9797217search_only2.0110.0814
reference norm4704f56750516217no_improvement2.0110.0816
reference norm598d9a14cdd80217no_improvement2.0110.0819
reference norm618e3d22acbe7217search_only2.0110.0819
reference norm7205485ec87aa217no_improvement2.0110.0832
reference norm899a2da54ecd1217no_improvement2.0110.0847
reference norm9e36296c9d6f9217no_improvement2.0110.0848
reference norm10b46b6eaea2ac217search_only2.0110.085

The grid leaves out 15 discovered cell elites that hold a MAP-Elites cell but were never compiled and traced, so they are named as excluded rather than priced by a fit. Every method in the grid has a trace row and every state count the problem set uses is priced, so nothing in the grid is extrapolated. The excluded elites are still counted in the re-ranking above, which changes the aggregate and re-prices nothing.

The discovered cell elites with no trace row, named rather than priced (15 rows)
tableauanalytic cycles/stepheld-out error
2d2c2b75d719150.0613
4feb237eaaa3170.0494
4f501ec6a5ce290.0477
5fef0bd2b5fe300.0836
2bceb12dc1c2390.0587
19b66f506eec460.0651
6d0004de713f570.0533
274563728359590.0587
196b1d17d0b7630.0603
c10bc5417b5b630.0822
407c65d76723650.0604
935100ec0e46650.0639
aea95967c0fd720.0453
42863b93bf7d750.0377
013ba3020051770.0808

Whole rk_step, derivative body excluded. Neither basis is the full per-step cost on a part: the analytic model prices less than the traced count, the traced count prices less than a real application step, and the pair is a bracket rather than a correction. The traced count is used only as a budget denominator, because a real part pays those cycles out of the same 65,536: quoting it against the analytic count as a ratio would report a scope difference as a model error, which DECISIONS.md D42 and D43 forbid. The prices come from an emulator that is instruction accurate and not cycle accurate, under TRACE_HASH 61f3a1bc and VERIFIER_HASH de5bec22, from rk-work/trace/results.json; the aggregates are from tools/key_findings.json, key counterfactual.

2Floor rounding reorders the classical methods

Every Q15 multiply rounds toward negative infinity, as ARM's ASRS does, so the error from each product leans the same way. At these budgets that bias piles up faster than truncation error shrinks. The counterfactual reruns the same 4 methods, budget and problems with round-to-nearest.

Each line is one method, ranked by search-set RMS error under each rounding mode, with the value beside each end. Blue is euler, orange rk4. Round-to-nearest was rerun outside the archive as the counterfactual.data: key_findings.json, numbers aggregate.search_rmsas of 2026-09-10 (US Central), computed at 141,364 archive records
round-to-nearest(counterfactual)floor (ASRS)(what the hardware does)rank 1rank 2rank 3rank 4euler: search-set RMS 0.0469 under round-to-nearest, 0.0129 under flooreuler: search-set RMS 0.0469 under round-to-nearest, 0.0129 under flooreuler · 0.0469euler · 0.0129heun2: search-set RMS 0.061 under round-to-nearest, 0.0174 under floorheun2: search-set RMS 0.061 under round-to-nearest, 0.0174 under floorheun2 · 0.061heun2 · 0.0174rk38: search-set RMS 0.0196 under round-to-nearest, 0.0219 under floorrk38: search-set RMS 0.0196 under round-to-nearest, 0.0219 under floorrk38 · 0.0196rk38 · 0.0219rk4: search-set RMS 0.0289 under round-to-nearest, 0.059 under floorrk4: search-set RMS 0.0289 under round-to-nearest, 0.059 under floorrk4 · 0.0289rk4 · 0.059
The same ranks and errors as a table (4 rows)
methodrank, round-to-nearestRMS, round-to-nearestrank, floorRMS, floor
euler30.046910.0129
heun240.06120.0174
rk3810.019630.0219
rk420.028940.059
Per problem: floor against round-to-nearest
Final-state error per problem and method, log scale, left is better. The bar between the two dots is what the rounding mode alone changes.data: key_findings.json, series per_problem_floor_vs_roundas of 2026-09-10 (US Central), computed at 141,364 archive records
floor (ASRS, as measured)round-to-nearest (counterfactual)
1e-41e-31e-21e-11e0final-state error at the shared budget (log)dahlquist (flagged)eulerdahlquist (flagged) / euler: floor 4.54e-05 (rank 1), round-to-nearest 0.0799 (rank 3)dahlquist (flagged) / euler: floor 4.54e-05 (rank 1), round-to-nearest 0.0799 (rank 3)heun2dahlquist (flagged) / heun2: floor 4.54e-05 (rank 2), round-to-nearest 0.0922 (rank 4)dahlquist (flagged) / heun2: floor 4.54e-05 (rank 2), round-to-nearest 0.0922 (rank 4)rk4dahlquist (flagged) / rk4: floor 0.000167 (rank 4), round-to-nearest 0.0362 (rank 2)dahlquist (flagged) / rk4: floor 0.000167 (rank 4), round-to-nearest 0.0362 (rank 2)rk38dahlquist (flagged) / rk38: floor 4.54e-05 (rank 3), round-to-nearest 0.0333 (rank 1)dahlquist (flagged) / rk38: floor 4.54e-05 (rank 3), round-to-nearest 0.0333 (rank 1)damped_osceulerdamped_osc / euler: floor 0.0142 (rank 3), round-to-nearest 0.00801 (rank 3)damped_osc / euler: floor 0.0142 (rank 3), round-to-nearest 0.00801 (rank 3)heun2damped_osc / heun2: floor 0.00554 (rank 1), round-to-nearest 0.013 (rank 4)damped_osc / heun2: floor 0.00554 (rank 1), round-to-nearest 0.013 (rank 4)rk4damped_osc / rk4: floor 0.0159 (rank 4), round-to-nearest 0.00736 (rank 2)damped_osc / rk4: floor 0.0159 (rank 4), round-to-nearest 0.00736 (rank 2)rk38damped_osc / rk38: floor 0.00817 (rank 2), round-to-nearest 0.000533 (rank 1)damped_osc / rk38: floor 0.00817 (rank 2), round-to-nearest 0.000533 (rank 1)vanderpol_mildeulervanderpol_mild / euler: floor 0.0172 (rank 1), round-to-nearest 0.0123 (rank 2)vanderpol_mild / euler: floor 0.0172 (rank 1), round-to-nearest 0.0123 (rank 2)heun2vanderpol_mild / heun2: floor 0.0296 (rank 2), round-to-nearest 0.05 (rank 4)vanderpol_mild / heun2: floor 0.0296 (rank 2), round-to-nearest 0.05 (rank 4)rk4vanderpol_mild / rk4: floor 0.101 (rank 4), round-to-nearest 0.0337 (rank 3)vanderpol_mild / rk4: floor 0.101 (rank 4), round-to-nearest 0.0337 (rank 3)rk38vanderpol_mild / rk38: floor 0.0371 (rank 3), round-to-nearest 0.00708 (rank 1)vanderpol_mild / rk38: floor 0.0371 (rank 3), round-to-nearest 0.00708 (rank 1)pendulumeulerpendulum / euler: floor 0.594 (rank 4), round-to-nearest 0.554 (rank 4)pendulum / euler: floor 0.594 (rank 4), round-to-nearest 0.554 (rank 4)heun2pendulum / heun2: floor 0.00421 (rank 3), round-to-nearest 0.00116 (rank 1)pendulum / heun2: floor 0.00421 (rank 3), round-to-nearest 0.00116 (rank 1)rk4pendulum / rk4: floor 0.00403 (rank 2), round-to-nearest 0.00277 (rank 3)pendulum / rk4: floor 0.00403 (rank 2), round-to-nearest 0.00277 (rank 3)rk38pendulum / rk38: floor 0.0011 (rank 1), round-to-nearest 0.00244 (rank 2)pendulum / rk38: floor 0.0011 (rank 1), round-to-nearest 0.00244 (rank 2)dc_motoreulerdc_motor / euler: floor 0.171 (rank 4), round-to-nearest 0.0162 (rank 3)dc_motor / euler: floor 0.171 (rank 4), round-to-nearest 0.0162 (rank 3)heun2dc_motor / heun2: floor 0.0974 (rank 3), round-to-nearest 0.0718 (rank 4)dc_motor / heun2: floor 0.0974 (rank 3), round-to-nearest 0.0718 (rank 4)rk4dc_motor / rk4: floor 0.0764 (rank 2), round-to-nearest 0.0121 (rank 2)dc_motor / rk4: floor 0.0764 (rank 2), round-to-nearest 0.0121 (rank 2)rk38dc_motor / rk38: floor 0.052 (rank 1), round-to-nearest 0.0114 (rank 1)dc_motor / rk38: floor 0.052 (rank 1), round-to-nearest 0.0114 (rank 1)rc_thermal (flagged)eulerrc_thermal (flagged) / euler: floor 0.156 (rank 1), round-to-nearest 0.149 (rank 3)rc_thermal (flagged) / euler: floor 0.156 (rank 1), round-to-nearest 0.149 (rank 3)heun2rc_thermal (flagged) / heun2: floor 0.156 (rank 2), round-to-nearest 0.197 (rank 4)rc_thermal (flagged) / heun2: floor 0.156 (rank 2), round-to-nearest 0.197 (rank 4)rk4rc_thermal (flagged) / rk4: floor 0.158 (rank 4), round-to-nearest 0.0663 (rank 2)rc_thermal (flagged) / rk4: floor 0.158 (rank 4), round-to-nearest 0.0663 (rank 2)rk38rc_thermal (flagged) / rk38: floor 0.156 (rank 3), round-to-nearest 0.0511 (rank 1)rc_thermal (flagged) / rk38: floor 0.156 (rank 3), round-to-nearest 0.0511 (rank 1)quaternioneulerquaternion / euler: floor 0.0596 (rank 4), round-to-nearest 0.0124 (rank 3)quaternion / euler: floor 0.0596 (rank 4), round-to-nearest 0.0124 (rank 3)heun2quaternion / heun2: floor 0.0559 (rank 3), round-to-nearest 0.0274 (rank 4)quaternion / heun2: floor 0.0559 (rank 3), round-to-nearest 0.0274 (rank 4)rk4quaternion / rk4: floor 0.0541 (rank 2), round-to-nearest 0.00171 (rank 1)quaternion / rk4: floor 0.0541 (rank 2), round-to-nearest 0.00171 (rank 1)rk38quaternion / rk38: floor 0.0447 (rank 1), round-to-nearest 0.00558 (rank 2)quaternion / rk38: floor 0.0447 (rank 1), round-to-nearest 0.00558 (rank 2)
The same pairs as a table (28 rows)
problemmethodfloor errorfloor weight sharefloor rankround-to-nearest errorround-to-nearest weight shareround-to-nearest rankverdict over these 4 methods
dahlquisteuler4.54e-050.00041%10.079996.74%3degenerate
dahlquistheun24.54e-050.00023%20.092276.12%4degenerate
dahlquistrk40.0001670.00027%40.036252.44%2degenerate
dahlquistrk384.54e-050.00014%30.033395.65%1degenerate
damped_osceuler0.014240.48%30.008010.97%3clean
damped_oscheun20.005543.38%10.0131.52%4clean
damped_oscrk40.01592.43%40.007362.17%2clean
damped_oscrk380.008174.63%20.0005330.025%1clean
vanderpol_mildeuler0.017259.52%10.01232.29%2clean
vanderpol_mildheun20.029696.62%20.0522.36%4clean
vanderpol_mildrk40.10197.57%40.033745.40%3clean
vanderpol_mildrk380.037195.37%30.007084.33%1clean
pendulumeuler0.59486.08%40.55493.16%4clean
pendulumheun20.004210.048%30.001160.003%1clean
pendulumrk40.004030.048%20.002770.17%3clean
pendulumrk380.00110.0042%10.002440.21%2clean
dc_motoreuler0.1717.13%40.01620.08%3clean
dc_motorheun20.097425.69%30.071811.54%4clean
dc_motorrk40.076417.26%20.01213.21%2clean
dc_motorrk380.0529.30%10.01144.68%1clean
rc_thermaleuler0.1565.92%10.1496.71%3degenerate
rc_thermalheun20.15665.79%20.19786.78%4degenerate
rc_thermalrk40.15874.04%40.066396.56%2degenerate
rc_thermalrk380.15683.83%30.051193.98%1degenerate
quaternioneuler0.05960.87%40.01240.047%3clean
quaternionheun20.05598.47%30.02741.68%4clean
quaternionrk40.05418.65%20.001710.064%1clean
quaternionrk380.04476.86%10.005581.12%2clean

The weight share is the problem's share of the sum of squares behind its own aggregate, the search set or the held-out set, which is the weight an RMS gives it. The verdict is the findings site's own degeneracy detector run over the four methods in this table, and it names that field because a verdict belongs to the field it was computed over; the drawing marks the same problems (flagged). The table under the Dahlquist paragraph gives every problem's verdict on this field and on two others.

On the three search problems, under magnitude weighting on the analytic basis (the stage and b combinations only), floor puts euler at rank 1 of these 4 methods (RMS 0.0129) and rk4 at rank 4 (0.059, 4.6× worse), a rank 1 that survives none of the five other cells of the grid (below). Round-to-nearest, on the same weighting and basis, restores the textbook order, with rk4 ahead of euler (0.0289 against 0.0469). Per problem, the best method changes on 5 of the 7 problems, and one of those 5, rc_thermal, is flagged degenerate over these four methods; dahlquist, also flagged over these four methods, is among the problems where it does not change.

Dahlquist shows the mechanism most clearly. Under floor, euler, heun2 and rk38 all report error 4.54e-05, which is the entire reference value: the Q15 state decays to exactly zero, the true answer is below one LSB, and floor's downward bias lands on it. Under round-to-nearest the state stalls above zero and the error is up to 2,032× larger.

Two of the seven problems are flagged degenerate by the same detector the findings validation page uses, computed over the four methods in the table above: dahlquist and rc_thermal. Over the 9 methods finding 1's counterfactual scores on the analytic basis, 8 classical and one discovered, it flags dahlquist and rc_thermal, and over the same 9 on the traced whole-step basis it flags dahlquist. On rc_thermal, 11e898cb, the one discovered method in that field, is the only finisher more than 5 percent from the reference norm, and the other 8 span 1.59 percent, under the 5 percent spread threshold. Of the problems flagged over both the 4-method and the 9-method analytic fields, the rule that fires differs between the two for rc_thermal, the spread rule over the 4 and the pair rule over the 9; the same rule fires on both for dahlquist, the pair rule. The table below gives every problem's verdict and its reason on all three fields.

Every problem's degeneracy verdict and reason on the three fields (7 rows)
problemover the 4 methods in this findingover the 9 methods, analytic basisover the 9 methods, traced whole-step basis
dahlquistflagged: 3 of the 4 finishers sit within 5 percent of the reference solution's norm of 4.53999e-05 and agree with each other to within 5 percent, so those runs returned approximately the reference and the comparison is reporting the problem rather than the methodflagged: 4 of the 9 finishers sit within 5 percent of the reference solution's norm of 4.53999e-05 and agree with each other to within 5 percent, so those runs returned approximately the reference and the comparison is reporting the problem rather than the methodflagged: 5 of the 9 finishers sit within 5 percent of the reference solution's norm of 4.53999e-05 and agree with each other to within 5 percent, so those runs returned approximately the reference and the comparison is reporting the problem rather than the method
damped_oscnot flagged: the 4 finishers' Q15 errors span 187.30 percent from best to worst, over the 5 percent threshold, and none of them sits within 5 percent of the reference solution's norm of 0.0176384, so the comparison is about the methodsnot flagged: the 9 finishers' Q15 errors span 304.24 percent from best to worst, over the 5 percent threshold, and none of them sits within 5 percent of the reference solution's norm of 0.0176384, so the comparison is about the methodsnot flagged: the 9 finishers' Q15 errors span 1643.90 percent from best to worst, over the 5 percent threshold, and none of them sits within 5 percent of the reference solution's norm of 0.0176384, so the comparison is about the methods
vanderpol_mildnot flagged: the 4 finishers' Q15 errors span 486.22 percent from best to worst, over the 5 percent threshold, and none of them sits within 5 percent of the reference solution's norm of 0.77394, so the comparison is about the methodsnot flagged: the 9 finishers' Q15 errors span 486.22 percent from best to worst, over the 5 percent threshold, and none of them sits within 5 percent of the reference solution's norm of 0.77394, so the comparison is about the methodsnot flagged: the 9 finishers' Q15 errors span 2632.54 percent from best to worst, over the 5 percent threshold, and none of them sits within 5 percent of the reference solution's norm of 0.77394, so the comparison is about the methods
pendulumnot flagged: the 4 finishers' Q15 errors span 53902.14 percent from best to worst, over the 5 percent threshold, and none of them sits within 5 percent of the reference solution's norm of 1, so the comparison is about the methodsnot flagged: the 9 finishers' Q15 errors span 921909.64 percent from best to worst, over the 5 percent threshold, and none of them sits within 5 percent of the reference solution's norm of 1, so the comparison is about the methodsnot flagged: the 9 finishers' Q15 errors span 8181338.72 percent from best to worst, over the 5 percent threshold, and none of them sits within 5 percent of the reference solution's norm of 1, so the comparison is about the methods
dc_motornot flagged: the 4 finishers' Q15 errors span 228.95 percent from best to worst, over the 5 percent threshold, and none of them sits within 5 percent of the reference solution's norm of 1.01989, so the comparison is about the methodsnot flagged: the 9 finishers' Q15 errors span 518.66 percent from best to worst, over the 5 percent threshold, and none of them sits within 5 percent of the reference solution's norm of 1.01989, so the comparison is about the methodsnot flagged: the 9 finishers' Q15 errors span 303.76 percent from best to worst, over the 5 percent threshold, and none of them sits within 5 percent of the reference solution's norm of 1.01989, so the comparison is about the methods
rc_thermalflagged: the 4 finishers' Q15 errors span 1.45 percent from best to worst, under the 5 percent threshold, so the comparison is reporting the problem rather than the methodflagged: 8 of the 9 finishers sit within 5 percent of the reference solution's norm of 0.159158 and agree with each other to within 5 percent, so those runs returned approximately the reference and the comparison is reporting the problem rather than the methodnot flagged: the 9 finishers' Q15 errors span 536.87 percent from best to worst, over the 5 percent threshold, and none of them sits within 5 percent of the reference solution's norm of 0.159158, so the comparison is about the methods
quaternionnot flagged: the 4 finishers' Q15 errors span 33.41 percent from best to worst, over the 5 percent threshold, and none of them sits within 5 percent of the reference solution's norm of 1, so the comparison is about the methodsnot flagged: the 9 finishers' Q15 errors span 211.86 percent from best to worst, over the 5 percent threshold, and none of them sits within 5 percent of the reference solution's norm of 1, so the comparison is about the methodsnot flagged: the 9 finishers' Q15 errors span 2045.50 percent from best to worst, over the 5 percent threshold, and none of them sits within 5 percent of the reference solution's norm of 1, so the comparison is about the methods

Whether rank 1 survives a different weighting or a compiled cost basis

euler holds rank 1 of these four methods in exactly one cell of the six: magnitude weighting on the analytic basis, which is the cell the paragraph above is written from. At the traced whole-step price (the whole rk_step, derivative body excluded) under magnitude weighting it is rank 4 of 4 behind rk38, so its rank 1 does not survive the compiled cost basis. On the analytic basis (the stage and b combinations only) it is rank 3 of 4 behind heun2 under both the median-anchor scale and the reference-norm scale. On the traced whole-step basis (the whole rk_step, derivative body excluded) it is rank 2 of 4 behind heun2 under both the median-anchor scale and the reference-norm scale. So euler's rank 1 survives in none of the five other cells.

Search-set error and rank in every cell of the grid (54 rows)
weightingcost basismethodsearch-set RMSrank of the nine scoredrank of the four in this finding
magnitude (published)analytic (published)11e898cb0.01151 of 9n/a
magnitude (published)analytic (published)euler0.01292 of 91 of 4
magnitude (published)analytic (published)midpoint0.01313 of 9n/a
magnitude (published)analytic (published)heun30.01644 of 9n/a
magnitude (published)analytic (published)heun20.01745 of 92 of 4
magnitude (published)analytic (published)rk380.02196 of 93 of 4
magnitude (published)analytic (published)ralston20.02377 of 9n/a
magnitude (published)analytic (published)kutta30.05458 of 9n/a
magnitude (published)analytic (published)rk40.0599 of 94 of 4
magnitude (published)traced whole step11e898cb0.001021 of 9n/a
magnitude (published)traced whole stepheun30.009032 of 9n/a
magnitude (published)traced whole steprk380.009463 of 91 of 4
magnitude (published)traced whole stepheun20.01014 of 92 of 4
magnitude (published)traced whole stepkutta30.01115 of 9n/a
magnitude (published)traced whole steprk40.01216 of 93 of 4
magnitude (published)traced whole stepralston20.01317 of 9n/a
magnitude (published)traced whole stepmidpoint0.01478 of 9n/a
magnitude (published)traced whole stepeuler0.02659 of 94 of 4
equal, median anchoranalytic (published)heun20.7951 of 91 of 4
equal, median anchoranalytic (published)midpoint0.9682 of 9n/a
equal, median anchoranalytic (published)rk3813 of 92 of 4
equal, median anchoranalytic (published)euler1.224 of 93 of 4
equal, median anchoranalytic (published)ralston21.425 of 9n/a
equal, median anchoranalytic (published)heun31.736 of 9n/a
equal, median anchoranalytic (published)11e898cb1.937 of 9n/a
equal, median anchoranalytic (published)kutta32.318 of 9n/a
equal, median anchoranalytic (published)rk42.669 of 94 of 4
equal, median anchortraced whole stepheun30.5421 of 9n/a
equal, median anchortraced whole stepkutta30.5482 of 9n/a
equal, median anchortraced whole stepheun20.5663 of 91 of 4
equal, median anchortraced whole stepralston20.6714 of 9n/a
equal, median anchortraced whole stepmidpoint0.8665 of 9n/a
equal, median anchortraced whole stepeuler1.236 of 92 of 4
equal, median anchortraced whole steprk41.647 of 93 of 4
equal, median anchortraced whole step11e898cb1.888 of 9n/a
equal, median anchortraced whole steprk382.779 of 94 of 4
reference normanalytic (published)heun20.6061 of 91 of 4
reference normanalytic (published)rk380.6372 of 92 of 4
reference normanalytic (published)euler0.7413 of 93 of 4
reference normanalytic (published)ralston20.7644 of 9n/a
reference normanalytic (published)midpoint0.9995 of 9n/a
reference normanalytic (published)kutta32.146 of 9n/a
reference normanalytic (published)heun32.147 of 9n/a
reference normanalytic (published)rk42.198 of 94 of 4
reference normanalytic (published)11e898cb2.539 of 9n/a
reference normtraced whole stepkutta30.5781 of 9n/a
reference normtraced whole stepheun30.5832 of 9n/a
reference normtraced whole stepheun20.5853 of 91 of 4
reference normtraced whole stepralston20.5964 of 9n/a
reference normtraced whole stepeuler0.685 of 92 of 4
reference normtraced whole stepmidpoint0.9786 of 9n/a
reference normtraced whole steprk42.137 of 93 of 4
reference normtraced whole step11e898cb2.538 of 9n/a
reference normtraced whole steprk383.689 of 94 of 4

Search-set RMS over the three search problems, one row per method in every cell. The analytic basis prices the stage and b combinations only; the traced whole-step basis buys steps at the count for the whole rk_step with the derivative body excluded, from an emulator that is instruction accurate and not cycle accurate, and it is used only as a budget denominator.

3Where quantization overtakes truncation

The premise was tested before the search started. If truncation error dominated roundoff at practical step sizes, classical analysis would already give the right answer and the search would be pointless. The falsification run swept the step size on damped_osc, computing each method's error twice, in Q15 and in float64 over exactly the same steps, so the gap between the curves is the arithmetic alone.

Final-state error on damped_osc against step size, both axes log. Solid lines are Q15, dashed lines float64 over the same steps. Left of each vertical marker the Q15 line turns up while float64 keeps falling. Q15 runs that overflowed at the largest steps are left out. So are float64 errors above 1, which no Q15 state can hold: rk4 at h = 5; heun2 at h = 1.25, 2.5 and 5.data: key_findings.json, series sweeps (falsification run)as of 2026-09-10 (US Central), computed at 141,364 archive records
rk4 (solid Q15, dashed float64)heun2 (solid Q15, dashed float64)
0.010.050.10.5121e-101e-81e-61e-41e-21e01e2step size h (log; smaller steps to the left)final-state error (log)rk4 crossover h = 0.156heun2 crossover h = 0.0391rk4 Q15 at h = 0.00977: error 0.0258rk4 Q15 at h = 0.0195: error 0.0399rk4 Q15 at h = 0.0391: error 0.0156rk4 Q15 at h = 0.0781: error 0.00479rk4 Q15 at h = 0.156: error 0.00254rk4 Q15 at h = 0.312: error 0.00113rk4 Q15 at h = 0.625: error 0.00134rk4 float64 at h = 0.00977: error 5.84e-11rk4 float64 at h = 0.0195: error 9.35e-10rk4 float64 at h = 0.0391: error 1.5e-08rk4 float64 at h = 0.0781: error 2.41e-07rk4 float64 at h = 0.156: error 3.88e-06rk4 float64 at h = 0.312: error 6.29e-05rk4 float64 at h = 0.625: error 0.00102rk4 float64 at h = 1.25: error 0.0125rk4 float64 at h = 2.5: error 0.0176rk4heun2 Q15 at h = 0.00977: error 0.0276heun2 Q15 at h = 0.0195: error 0.0126heun2 Q15 at h = 0.0391: error 0.00642heun2 Q15 at h = 0.0781: error 0.00297heun2 Q15 at h = 0.156: error 0.00401heun2 Q15 at h = 0.312: error 0.0128heun2 Q15 at h = 0.625: error 0.0467heun2 float64 at h = 0.00977: error 1.23e-05heun2 float64 at h = 0.0195: error 4.92e-05heun2 float64 at h = 0.0391: error 0.000197heun2 float64 at h = 0.0781: error 0.000786heun2 float64 at h = 0.156: error 0.00313heun2 float64 at h = 0.312: error 0.0123heun2 float64 at h = 0.625: error 0.0467heun2
Every sweep point as a table, the ones the chart leaves out included (20 rows)
methodstep size hQ15 errorfloat64 error
rk40.009770.02585.84e-11
rk40.01950.03999.35e-10
rk40.03910.01561.5e-08
rk40.07810.004792.41e-07
rk40.1560.002543.88e-06
rk40.3120.001136.29e-05
rk40.6250.001340.00102
rk41.25n/a0.0125
rk42.5n/a0.0176
rk45n/a2.37e+10
heun20.009770.02761.23e-05
heun20.01950.01264.92e-05
heun20.03910.006420.000197
heun20.07810.002970.000786
heun20.1560.004010.00313
heun20.3120.01280.0123
heun20.6250.04670.0467
heun21.25n/a17.5
heun22.5n/a3.68e+07
heun25n/a4.41e+08

rk4's Q15 error bottoms out at 0.00113, separates from the float curve at h = 0.156, and climbs back to 0.0258 by h = 0.00977: below the crossover, more steps mean more floored multiplies and less accuracy. heun2 separates at h = 0.0391. Both crossovers fall inside the practical step-size range, as the premise needed.

The stored verdict is "mixed": coefficient arithmetic is 50 to 56% of per-step cycles under the fast multiplier, above the 30% proceed line, but heun2 under the slow one sits at 14.8%, just under the 15% kill line. At the shared budget under floor, rk4 is the best of the 4 methods on none of the 7 problems, and dahlquist and rc_thermal, two of those 7, are flagged degenerate over the same 4 methods (finding 2).

The shape of this is not new, and reporting it that way is the point. Analysis of low-precision Runge-Kutta integration finds that a round-to-nearest solution stagnates once the timestep is small enough, with the rounding term growing as the step shrinks. Measuring the same turn-up in fixed point under floor rounding says the harness is seeing a real effect rather than an artifact of its own arithmetic (related work).

4The rc_thermal quantization floor

rc_thermal is the archive's stiff problem: a three-state RC network decaying toward equilibrium, with its derivative stored at DERIV_SCALE = 1/8 and one LSB worth 0.000122 in physical units. All 8 classical methods report errors between 0.1558 and 0.1582, a spread of 1.6%. They agree because every one of them collapses.

Final-state error on rc_thermal. Every floor bar (blue) reaches the dashed line, the size of the true solution, because the state collapsed to near zero. Orange: round-to-nearest, measured for 4 methods. Green: the best discovered method on rc_thermal (not the finding 1 champion), under the same floor arithmetic.data: key_findings.json, series per_methodas of 2026-09-10 (US Central), computed at 141,364 archive records
floor (ASRS, as measured)round-to-nearest (counterfactual)best discovered, still under floor
00.050.10.150.2final-state erroreuler floor: error 0.156; 4,369 steps; final Q15 state (14, 15, 17)euler round-to-nearest: error 0.149eulermidpoint floor: error 0.156; 1,985 steps; final Q15 state (15, 15, 18)midpointheun2 floor: error 0.156; 1,680 steps; final Q15 state (14, 15, 17)heun2 round-to-nearest: error 0.197heun2ralston2 floor: error 0.156; 1,365 steps; final Q15 state (14, 15, 17)ralston2heun3 floor: error 0.156; 949 steps; final Q15 state (14, 15, 16)heun3kutta3 floor: error 0.158; 840 steps; final Q15 state (4, 4, 5)kutta3rk4 floor: error 0.158; 661 steps; final Q15 state (4, 4, 6)rk4 round-to-nearest: error 0.06630.0663rk4rk38 floor: error 0.156; 606 steps; final Q15 state (13, 14, 16)rk38 round-to-nearest: error 0.05110.0511rk38best discovered (order 2, 2 stages, 17 cycles/step), still under floor: error 0.02050.0205bestdiscoveredtrue solution norm 0.159floor: 0.156 to 0.158 for all 8
The same bars as a table (9 rows)
methodfloor errorshare of its held-out sum of squaresround-to-nearest errorstepsfinal Q15 state
euler0.1565.92%0.1494,369(14, 15, 17)
midpoint0.15685.06%n/a1,985(15, 15, 18)
heun20.15665.79%0.1971,680(14, 15, 17)
ralston20.15672.03%n/a1,365(14, 15, 17)
heun30.15683.84%n/a949(14, 15, 16)
kutta30.15882.20%n/a840(4, 4, 5)
rk40.15874.04%0.0663661(4, 4, 6)
rk380.15683.83%0.0511606(13, 14, 16)
best discovered, under floor0.0205n/an/an/an/a

The share is rc_thermal's part of that method's held-out sum of squares under magnitude weighting on the analytic basis, from counterfactual.series.per_method; the best discovered method on rc_thermal is not one of the methods scored there, so its row has none. rc_thermal (flagged): the degeneracy detector flags it over the 4 methods finding 2 publishes and the 9 methods finding 1's counterfactual scores on the analytic basis (finding 2).

Each method drives the Q15 state to within 5 to 18 LSBs of the origin (euler ends at raw state (14, 15, 17), rk4 at (4, 4, 6)) while the true solution still has norm 0.1592 at t = 4, so the reported error is the reference itself. Late in the decay the scaled derivative terms are a few LSBs, and floor turns them into extra decay.

It is a floor effect, not a limit of Q15: under round-to-nearest rk38 reaches 0.0511, and the best discovered method on this problem, a 2-stage, order-2 tableau at 17 cycles per step, reaches 0.0205 while still using floor.

The degeneracy detector flags rc_thermal over the 4 methods finding 2 publishes and the 9 methods finding 1's counterfactual scores on the analytic basis, and not over those 9 on the traced whole-step basis. The share column in the table above is rc_thermal's part of each method's held-out sum of squares under magnitude weighting on the analytic basis.

5Phase 0, closed by enumeration

Phase 0 is the one part of the space small enough to close completely: two-stage methods with a21 on a dyadic lattice, 256 candidates, of which 16 have exactly representable b weights. All 16 were enumerated, verified and archived, so the result holds over the whole space rather than a sample. It is an exhaustive evaluation and not a theorem, and it holds only for these four held-out problems, this 65,536-cycle budget, and Q15 with floor rounding. The top two members finish 0.04% apart, which is a tie at this resolution rather than an ordering.

All 16 members of the phase-0 space by a21, held-out error on a linear scale (lower is better). Blue is the optimum, orange the 2 textbook methods, gray the rest.data: key_findings.json, series all_membersas of 2026-09-10 (US Central), computed at 141,364 archive records
0.060.080.10.120.14a21, ordered by value (one dot per exactly representable tableau, 16 in all)held-out errora21 = -2/1, b = (5/4, -1/4): held-out error 0.109, 15 cycles/step, rank 14 of 16-2/1a21 = -1/1, b = (3/2, -1/2): held-out error 0.11, 14 cycles/step, rank 15 of 16-1/1a21 = -1/2, b = (2/1, -1/1): held-out error 0.0781, 13 cycles/step, rank 6 of 16-1/2a21 = -1/4, b = (3/1, -2/1): held-out error 0.0614, 15 cycles/step, rank 2 of 16near-tie 0.0614-1/4a21 = -1/8, b = (5/1, -4/1): held-out error 0.0613, 15 cycles/step, rank 1 of 16optimum 0.0613-1/8a21 = -1/16, b = (9/1, -8/1): held-out error 0.0889, 15 cycles/step, rank 9 of 16-1/16a21 = -1/32, b = (17/1, -16/1): held-out error 0.071, 15 cycles/step, rank 3 of 16-1/32a21 = -1/64, b = (33/1, -32/1): held-out error 0.0849, 15 cycles/step, rank 8 of 16-1/64a21 = 1/64, b = (-31/1, 32/1): held-out error 0.136, 15 cycles/step, rank 16 of 161/64a21 = 1/32, b = (-15/1, 16/1): held-out error 0.0898, 15 cycles/step, rank 10 of 161/32a21 = 1/16, b = (-7/1, 8/1): held-out error 0.102, 15 cycles/step, rank 13 of 161/16a21 = 1/8, b = (-3/1, 4/1): held-out error 0.0763, 15 cycles/step, rank 4 of 161/8a21 = 1/4, b = (-1/1, 2/1): held-out error 0.0781, 13 cycles/step, rank 5 of 161/4a21 = 1/2, b = (0/1, 1/1): held-out error 0.0844, 11 cycles/step, rank 7 of 16 (midpoint)midpoint1/2a21 = 1/1, b = (1/2, 1/2): held-out error 0.0961, 13 cycles/step, rank 12 of 16 (heun2)heun21/1a21 = 2/1, b = (3/4, 1/4): held-out error 0.0937, 15 cycles/step, rank 11 of 162/1
The same members as a table, best first (16 rows)
a21bheld-out errorcycles/stepranktextbook name
-1/8(5/1, -4/1)0.0613151
-1/4(3/1, -2/1)0.0614152
-1/32(17/1, -16/1)0.071153
1/8(-3/1, 4/1)0.0763154
1/4(-1/1, 2/1)0.0781135
-1/2(2/1, -1/1)0.0781136
1/2(0/1, 1/1)0.0844117midpoint
-1/64(33/1, -32/1)0.0849158
-1/16(9/1, -8/1)0.0889159
1/32(-15/1, 16/1)0.08981510
2/1(3/4, 1/4)0.09371511
1/1(1/2, 1/2)0.09611312heun2
1/16(-7/1, 8/1)0.1021513
-2/1(5/4, -1/4)0.1091514
-1/1(3/2, -1/2)0.111415
1/64(-31/1, 32/1)0.1361516

The two best members are a21 = −1/8 with b = (5, −4), at held-out error 0.0613 and 15 cycles per step, and a21 = −1/4 at 0.0614, 0.04% behind it: a tie, not a ranking. Neither is a textbook method, and both have lower held-out error than all 8 classical methods (the best of those is midpoint at 0.0844). Within the family, midpoint ranks 7 and heun2 12 of 16.

Re-ranking with one held-out problem dropped

A ranking of 16 members whose top two are 0.04% apart is worth testing against the problem set that produced it. Dropping each held-out problem in turn and re-ranking every member gives five orderings. The member ranked first changes across them (−1/32, −1/4, −1/8 and 1/16), and the top two places are not the same pair every time. The two that lead the full set never fall below rank 3 in any of the 5 orderings. The sharpest move is a21 = 1/16, which ranks 13 on the full set and first once dc_motor is dropped.

a21all four keptdrop pendulumdrop dc_motordrop rc_thermaldrop quaternion
−1/813212
−1/422321
−1/3231763
1/845674
1/457935
−1/266846
1/2 (midpoint)781057
−1/648411108
−1/16994129
1/32101051111
2111214810
1 (heun2)121313912
1/16131411513
−21415151314
−11516121415
1/641611161616

Rank by held-out RMS over the problems each column keeps, from tools/key_findings.json, phase0_exhaustive.numbers.rank_stability. The first column is the published ranking, recomputed here from the per-problem errors.

6Practical problems nobody tuned for

Everything above uses the archive's own search and held-out problems. The validation suite asks whether the winners transfer to 5 equations from embedded applications (power electronics, battery management, vehicle dynamics / ADAS, communications / clocking and biomedical devices) that no optimizer saw, run at the same 65,536-cycle budget in Q15 with floor rounding and scored against independent reference solutions.

The comparator is fixed. It is the champion from finding 1, picked at cycle 33 out of 141,364 archived records, before this suite ran, with its coefficients unchanged. Against the best of the 5 classical anchors on each problem it has the lower Q15 error on 3 of 5 non-stiff problems, and on 0 of the 1 stiff problems where both sides finish, with 1 tie inside the 2 percent band. Taking the anchors one at a time instead of the best of them, one discovered method against 5 classical ones with no maximum on either side, it is lower in 27 of 30 cells, higher in 2 and tied in 1. The champion is not ahead everywhere: bicycle_lateral, where midpoint reaches 0.000395 against the champion's 0.000539 (ratio 1.37) and pll_lock, where rk38 reaches 0.0128 against the champion's 0.0159 (ratio 1.24). On bicycle_lateral it is also not the discovered method with the lowest error: 196b1d17d0b7 reaches 0.000177 against 0.000539, so that row counts for the discovered side in the best-of-N view and against the champion here.

problemchampion Q15 errorbest classical anchorits Q15 errorchampion against it
buck_converter0.00198rk380.0208lower
battery_2rc0.0369heun20.0776lower
bicycle_lateral0.000539midpoint0.000395higher
pll_lock0.0159rk380.0128higher
glucose_minimal0.0207midpoint0.0373lower
servo_load_step (stiff)0.0116midpoint0.0116tie
enzyme_qssa (stiff) (flagged)0.00966midpoint0.0097left out, degenerate
robertson_scaled (stiff)n/amidpoint0.0894no finish

Every problem in the suite, with the champion against the best of the anchors that finished it. Read from rk-work/validation/results.json. (flagged) marks a problem the degeneracy detector flags over the finishers among all 8 methods the suite ran on it, and a flagged problem is left out of both tallies.

One problem is left out of both tallies: enzyme_qssa because the 6 finishers' Q15 errors span 2.68 percent from best to worst, under the 5 percent threshold, so the comparison is reporting the problem rather than the method. A ratio within 2% of 1.0 is counted as a tie rather than a win, on both sides of the comparison. The findings validation page applies the same two rules to the same document, so the two sites report one tally. This suite is summarized by medians of per-problem ratios and by win counts rather than by an RMS, so no problem carries a weight in its tables and there is no share to report in them, unlike the per-problem tables under finding 1 and finding 2.

Final-state Q15 error per practical problem at the 65,536-cycle budget, log scale, left is better. Orange is the best classical method (named), blue the best of the discovered ones, which is ahead on 4 of 5; on pll_lock the classical method keeps the win. Each dot is a maximum: best of 3 discovered against best of 5 classical.data: rk-work/validation/results.json, verdicts.per_problemcomputed at 141,364 archive records; the validation document records no date of its own
best discoveredbest classical method
1e-41e-31e-21e-1final-state Q15 error at the 65,536-cycle budget (log)buck_converterbuck_converter: best classical rk38 at 0.0208, best discovered at 0.00198 (ratio 0.0953)buck_converter: best classical rk38 at 0.0208, best discovered at 0.00198 (ratio 0.0953)0.00198rk38 0.0208battery_2rcbattery_2rc: best classical heun2 at 0.0776, best discovered at 0.0369 (ratio 0.475)battery_2rc: best classical heun2 at 0.0776, best discovered at 0.0369 (ratio 0.475)0.0369heun2 0.0776bicycle_lateralbicycle_lateral: best classical midpoint at 0.000395, best discovered at 0.000177 (ratio 0.45)bicycle_lateral: best classical midpoint at 0.000395, best discovered at 0.000177 (ratio 0.45)0.000177midpoint 0.000395pll_lockpll_lock: best classical rk38 at 0.0128, best discovered at 0.0159 (ratio 1.24)pll_lock: best classical rk38 at 0.0128, best discovered at 0.0159 (ratio 1.24)rk38 0.01280.0159glucose_minimalglucose_minimal: best classical midpoint at 0.0373, best discovered at 0.0207 (ratio 0.556)glucose_minimal: best classical midpoint at 0.0373, best discovered at 0.0207 (ratio 0.556)0.0207midpoint 0.0373
The same problems as a table (5 rows)
problembest classicalits Q15 errorbest discoveredits Q15 errorratio, discovered over classical
buck_converterrk380.020811e898cbab590.001980.095
battery_2rcheun20.077611e898cbab590.03690.475
bicycle_lateralmidpoint0.000395196b1d17d0b70.0001770.450
pll_lockrk380.012811e898cbab590.01591.240
glucose_minimalmidpoint0.037311e898cbab590.02070.556
The other way to count it: best of 3 discovered against best of 5 classical

Taking the lowest error on each side reads 4 of 5 on the non-stiff problems, with a median error ratio of 0.475 and the widest margin on buck_converter (0.00198 against 0.0208 for rk38, 10.5×). Both sides are maxima there, and the discovered side is a maximum over methods already selected on error, so that count rises with the number of discovered methods run. That is why the finding leads with one fixed comparator instead.

Float64 runs of the same tableaus and step counts are at least 337× more accurate in every case, so these errors measure quantization, and the largest raw Q15 value in any run is 20,130 of 32,767, so none leaned on overflow luck. The 3 stiff problems tell a different story (overflow, not accuracy); they and the full tables are on the findings validation page.

Discovered, classical and library methods

The chart shows one column of the full matrix folded below it; no single column ranks the methods.

Held-out RMS error at the 65,536-cycle budget for every scored method, best at the top, log scale. The libraries never ran this protocol, so they have no dot.data: tools/key_findings.json, frontier series and classical anchorsas of 2026-09-10 (US Central), computed at 141,364 archive records
discoveredclassical
1e-21e-11e011e898cb (discovered champion)11e898cb (discovered champion): held-out error 0.0286 at the 65,536-cycle budget0.028642863b93 (best order-3 elite)42863b93 (best order-3 elite): held-out error 0.0377 at the 65,536-cycle budget0.0377196b1d17 (best order-4 elite)196b1d17 (best order-4 elite): held-out error 0.0603 at the 65,536-cycle budget0.0603midpointmidpoint: held-out error 0.0844 at the 65,536-cycle budget0.0844heun3heun3: held-out error 0.0852 at the 65,536-cycle budget0.0852rk38rk38: held-out error 0.0853 at the 65,536-cycle budget0.0853kutta3kutta3: held-out error 0.0873 at the 65,536-cycle budget0.0873ralston2ralston2: held-out error 0.0919 at the 65,536-cycle budget0.0919rk4rk4: held-out error 0.0919 at the 65,536-cycle budget0.0919heun2heun2: held-out error 0.0961 at the 65,536-cycle budget0.0961eulereuler: held-out error 0.32 at the 65,536-cycle budget0.32held-out RMS error at the 65,536-cycle budget (log; left is better)
The same methods as a table, best first (11 rows)
methodkindheld-out RMS error
11e898cb (discovered champion)discovered0.0286
42863b93 (best order-3 elite)discovered0.0377
196b1d17 (best order-4 elite)discovered0.0603
midpointclassical0.0844
heun3classical0.0852
rk38classical0.0853
kutta3classical0.0873
ralston2classical0.0919
rk4classical0.0919
heun2classical0.0961
eulerclassical0.32

At matched tolerance the float64 libraries are far more accurate than any Q15 method (median error ratio, best Q15 over best library, 519), and at identical step counts float64 rk4 has the lower error in all 56 comparable cells. That is the price of 16-bit floor arithmetic. The Q15 methods offer bounded integer arithmetic at a fixed cycle cost known in advance, which no library here provides, and within that regime the discovered methods have the lowest held-out error (finding 1). The stiff column is the weak spot of the whole explicit class: an expensive tableau takes large steps at a fixed budget, leaves the stability region and overflows. That failure is the argument for the implicit class.

The full matrix: every method, every column, and where each column comes from
methodkindorderstagesheld-out error at budgetavalidation wins (practical / stiff)bcycles/step fastccycles/step slowcCSD weightcmeasured µs/stepdstiff validation subsetenotes
eulerclassical110.320 / 05506finishes all 3Cheapest step and smallest code. Its tiny budget-driven steps stay inside the stability region on the stiff runs; the least accurate method elsewhere.
midpointclassical220.08440 / 21111112finishes all 3The best classical method at this budget and arithmetic, and the strongest on the stiff problems.
heun2classical220.09610 / 01313214finishes all 3Same cost class as midpoint and slightly worse throughout.
rk4classical440.09190 / 033853432finishes 2 of 3; overflows on robertson_scaledThe textbook default. Its classically optimal coefficients buy nothing at this budget in Q15, and it is expensive under a slow multiplier.
rk38classical440.08531 / 036642233finishes 2 of 3; overflows on robertson_scaledSame order and stages as rk4; under the frozen epoch-1 rule it is cheaper with a slow multiplier.
11e898cb (discovered champion)discovered230.02863 / 12229820finishes 2 of 3; overflows on robertson_scaledThe three-stage, order-2 discovered champion, with the lowest held-out error in the archive. Its edge comes from how it handles floor bias, not from truncation order.
42863b93 (best order-3 elite)discovered360.03770 / 0751204866overflows on all 3The best order-3 archive elite. Strong on the held-out set, but six stages mean large steps at a fixed budget, which the stiff problems punish.
196b1d17 (best order-4 elite)discovered460.06031 / 0631054072overflows on all 3The best order-4 archive elite, with the same six-stage stiff weakness.
RK45library (SciPy, float64, adaptive)n/an/an/aan/abn/acn/acn/ac64n/aeSciPy's general-purpose default: explicit, adaptive, float64. It picks its step count at run time, so there is no fixed cycle cost to certify.
Radaulibrary (SciPy, float64, adaptive)n/an/an/aan/abn/acn/acn/ac224n/aeImplicit Radau IIA, built for the stiff problems that limit the explicit methods.
BDFlibrary (SciPy, float64, adaptive)n/an/an/aan/abn/acn/acn/ac133n/aeImplicit multistep for stiff problems, with adaptive step counts and no cycle model.
LSODAlibrary (SciPy, float64, adaptive)n/an/an/aan/abn/acn/acn/ac15n/aeSwitches between stiff and non-stiff modes on its own; the same caveats as the other library rows.

Where each column comes from

  1. a. Held-out RMS error at the 65,536-cycle budget, m0plus_fast, Q15 with floor rounding, from tools/key_findings.json (series frontier_cycles_vs_heldout, matched by tableau hash). The libraries never ran this protocol, because they are adaptive and float64.
  2. b. Problems where this method has the lowest Q15 error of the 8 tested, from rk-work/validation/results.json, split into practical (5 non-stiff) and stiff (3). The libraries are not part of the validation suite.
  3. c. Analytic cycles per step for one state and total CSD weight, recomputed from each tableau with the pinned costmodel and coeffrep modules. CSD weight is the shift-add length of the coefficient multiplies, a proxy for code size. Adaptive libraries have no fixed cycles per step.
  4. d. Python wall clock per step: the median across the scored problems of per-problem medians (15 repeats after 3 warmups, gc paused), from rk-work/benchmark/results.json. Q15 rows run the pinned solver in the interpreter and library rows run compiled code, so times compare like with like only within a regime. Methods outside the benchmark run have no time.
  5. e. Behavior on the 3 stiff validation problems (stiffness ratios 292 to 1,030), from rk-work/validation/results.json. An overflow is a Q15OverflowError from the pinned solver. The libraries were not run on these problems.

Measured speed

The benchmark checks the cycle model against a clock: it runs the champion and rk4 through the same pinned Q15 solver on all 7 scored problems, so only the tableau differs and the per-step time ratio isolates what the coefficients cost.

Per-step speedup of the champion over rk4 on each scored problem. Blue is measured wall clock (median of 15 repeats after 3 warmups), open orange the cycle model's prediction of 1.50×; the measured geometric mean is 1.45×.data: rk-work/benchmark/results.json, speedup.rowstableaus taken from validation/results.json at 141,364 archive records; the benchmark document records no archive state or date of its own
1×1.2×1.4×1.6×paritydahlquistdahlquist: measured 1.602x (13.2 vs 21.1 µs/step over 2,978 vs 1,985 steps); cycle model predicts 1.50x (22 vs 33 cycles); error ratio champion/rk4 1.187dahlquist: measured 1.602x (13.2 vs 21.1 µs/step over 2,978 vs 1,985 steps); cycle model predicts 1.50x (22 vs 33 cycles); error ratio champion/rk4 1.1871.60×cycle modelmeasureddamped_oscdamped_osc: measured 1.698x (20.5 vs 34.8 µs/step over 1,489 vs 992 steps); cycle model predicts 1.50x (44 vs 66 cycles); error ratio champion/rk4 0.247damped_osc: measured 1.698x (20.5 vs 34.8 µs/step over 1,489 vs 992 steps); cycle model predicts 1.50x (44 vs 66 cycles); error ratio champion/rk4 0.2471.70×vanderpol_mildvanderpol_mild: measured 1.615x (19.8 vs 32.0 µs/step over 1,489 vs 992 steps); cycle model predicts 1.50x (44 vs 66 cycles); error ratio champion/rk4 0.193vanderpol_mild: measured 1.615x (19.8 vs 32.0 µs/step over 1,489 vs 992 steps); cycle model predicts 1.50x (44 vs 66 cycles); error ratio champion/rk4 0.1931.61×pendulumpendulum: measured 1.332x (21.8 vs 29.1 µs/step over 1,489 vs 992 steps); cycle model predicts 1.50x (44 vs 66 cycles); error ratio champion/rk4 1.177pendulum: measured 1.332x (21.8 vs 29.1 µs/step over 1,489 vs 992 steps); cycle model predicts 1.50x (44 vs 66 cycles); error ratio champion/rk4 1.1771.33×dc_motordc_motor: measured 1.660x (18.6 vs 30.9 µs/step over 1,489 vs 992 steps); cycle model predicts 1.50x (44 vs 66 cycles); error ratio champion/rk4 0.362dc_motor: measured 1.660x (18.6 vs 30.9 µs/step over 1,489 vs 992 steps); cycle model predicts 1.50x (44 vs 66 cycles); error ratio champion/rk4 0.3621.66×rc_thermalrc_thermal: measured 1.175x (32.8 vs 38.6 µs/step over 992 vs 661 steps); cycle model predicts 1.50x (66 vs 99 cycles); error ratio champion/rk4 0.291rc_thermal: measured 1.175x (32.8 vs 38.6 µs/step over 992 vs 661 steps); cycle model predicts 1.50x (66 vs 99 cycles); error ratio champion/rk4 0.2911.18×quaternionquaternion: measured 1.191x (38.1 vs 45.4 µs/step over 744 vs 496 steps); cycle model predicts 1.50x (88 vs 132 cycles); error ratio champion/rk4 0.353quaternion: measured 1.191x (38.1 vs 45.4 µs/step over 744 vs 496 steps); cycle model predicts 1.50x (88 vs 132 cycles); error ratio champion/rk4 0.3531.19×per-step speedup, rk4 time over champion time (above 1 means the champion is faster)
The same problems as a table (7 rows)
problemmeasured ratiocycle model ratiochampion µs/steprk4 µs/steperror ratio, champion over rk4
dahlquist (flagged)1.6021.5013.1721.091.187
damped_osc1.6981.5020.4734.750.247
vanderpol_mild1.6151.5019.8131.990.193
pendulum1.3321.5021.8429.081.177
dc_motor1.6601.5018.6130.880.362
rc_thermal (flagged)1.1751.5032.8338.580.291
quaternion1.1911.5038.1045.370.353

(flagged) marks a problem the degeneracy detector flags over the 9 methods finding 1's counterfactual scores on the analytic basis, a field that holds both methods in this table (finding 2 gives every verdict and its reason).

Every problem lands within 22% of the predicted ratio. At the same budget the champion also has the lower error on 5 of the 7 problems (rk4 keeps dahlquist and pendulum; median error ratio 0.35), so the time saved per step does not cost accuracy. Two of those 7, dahlquist and rc_thermal, are flagged degenerate over the 9 methods finding 1's counterfactual scores on the analytic basis, a field that holds both rk4 and the champion: dahlquist is among the problems rk4 keeps and rc_thermal among the champion's. Python timings check the model rather than stand in for the chip; the protocol and its caveats are on the findings validation page.

How the numbers were made

Code assigns each record's tier when it is written, by comparison with its cell's incumbent: heldout_verified when it improved on the incumbent in both search-set and held-out error across at least two problem families, search_only when its search error improved and its held-out error did not, no_incumbent when the cell was empty, and no_improvement when there was an incumbent and neither of those rules applied, which includes a held-out improvement without a search improvement. Before the split, no_incumbent and no_improvement were stored as one word, unreplicated, which is still a legal stored tier. Of the 141,364 archive records this page was built from, 40 are heldout_verified, 59,125 search_only, 21 no_incumbent, 82,178 no_improvement and 0 unreplicated. The model proposes hypotheses only as predicates in a closed grammar, and code writes the verdict: no data or an effect below Cohen's d of 0.2 is inconclusive, and refuted hypotheses are fed back into later prompts. The premise got the same treatment, a falsification run with thresholds fixed in advance and its verdict published either way (finding 3).

Limits

The cycle model is analytic and has never been checked against a physical chip inside the loop. Measured order comes from one scalar problem in float64. Error is judged at the final time, not along the trajectory, and floor is one hardware rounding convention among several. Optimality claims from the enumerated phases hold only within their lattices, and four held-out problems are a strong filter but not a statistical guarantee. The full list, with the reasoning, is on the findings methodology page.

The other two classes

Everything on this page is the explicit class. The implicit and adaptive classes are measured in float64, outside the scored archive, and their results are on the findings site's implicit and adaptive pages.