Implicit methods

SDIRK methods measured off-archive, and the stiff problems that motivate them.

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.

Implicit methods solve an equation at every stage, paying for a Jacobian and Newton iterations to stay stable on stiff problems. A scored record has no field for them, so the run searches two-stage SDIRK methods in a lane of its own: cycles set aside for this class that write an unscored archive. The lane reports the modeled cycles a method needs to reach an accuracy target.

At a glance

family
Two-stage SDIRK over a dyadic grid of gamma and a21, with a Newton iteration count and a Jacobian policy
search
Deterministic enumeration, plus the side-track jobs
question
How many modeled cycles does a method need to reach an accuracy target? (cycles to tolerance)
problems
The validation problems
arithmetic
float64: there is no Q15 LU factorization
checked by
Nothing: not order-verified and not scored
cost basis
A design estimate that assumes a division cost
sources
implicit_archive/elites.json, the side-track ledger, validation/results.json, benchmark/results.json
points measured
47
one per parameter point in the plan
jobs
6
each closes one open design question
L-stable gammas
0
none at any sdirk.gamma_a21_scan point. Arithmetic: tableau and stability algebra exact over Fractions; the measured order is float64
side-track code hash
2b9f0b0f9005
the executor and prototypes, at the latest measurement

The validation suite has not reported a stiff subset yet, so the gap this class addresses is not stated here.

Stability of each diagonal value

Each gamma the dyadic scan tried against |R(∞)|, the factor one step applies to a very stiff mode (log axis). Below 1 the mode is damped. None reaches 0, which L-stability needs; the smallest is 0.00124444, at gamma 75/256.
A-stablenot A-stable
00.20.40.60.810.0010.010.11101001000gamma, the diagonal value|R(∞)||R(∞)| = 1gamma 1/32: R(∞) = 449, not A-stable, measured order 2.1001gamma 1/16: R(∞) = 97, not A-stable, measured order 2.0912gamma 3/32: R(∞) = 329/9, not A-stable, measured order 2.0835gamma 1/8: R(∞) = 17, not A-stable, measured order 2.0779gamma 5/32: R(∞) = 217/25, not A-stable, measured order 2.0769gamma 11/64: R(∞) = 761/121, not A-stable, measured order 2.0811gamma 3/16: R(∞) = 41/9, not A-stable, measured order 2.096gamma 13/64: R(∞) = 553/169, not A-stable, measured order 2.1735gamma 7/32: R(∞) = 113/49, not A-stable, measured order 1.8356gamma 29/128: R(∞) = 1609/841, not A-stable, measured order 1.9477gamma 15/64: R(∞) = 353/225, not A-stable, measured order 1.9779gamma 31/128: R(∞) = 1217/961, not A-stable, measured order 1.9912gamma 1/4: R(∞) = 1, A-stable, measured order 1.9982gamma 33/128: R(∞) = 833/1089, A-stable, measured order 2.0023gamma 67/256: R(∞) = 2953/4489, A-stable, measured order 2.0036gamma 17/64: R(∞) = 161/289, A-stable, measured order 2.0046gamma 69/256: R(∞) = 2201/4761, A-stable, measured order 2.0054gamma 35/128: R(∞) = 457/1225, A-stable, measured order 2.0059gamma 71/256: R(∞) = 1457/5041, A-stable, measured order 2.0064gamma 143/512: R(∞) = 5089/20449, A-stable, measured order 2.0065gamma 9/32: R(∞) = 17/81, A-stable, measured order 2.0066gamma 145/512: R(∞) = 3617/21025, A-stable, measured order 2.0067gamma 73/256: R(∞) = 721/5329, A-stable, measured order 2.0068gamma 293/1024: R(∞) = 10073/85849, A-stable, measured order 2.0068gamma 147/512: R(∞) = 2153/21609, A-stable, measured order 2.0069gamma 295/1024: R(∞) = 7153/87025, A-stable, measured order 2.0069gamma 37/128: R(∞) = 89/1369, A-stable, measured order 2.0069gamma 593/2048: R(∞) = 19873/351649, A-stable, measured order 2.0069gamma 297/1024: R(∞) = 4241/88209, A-stable, measured order 2.0069gamma 595/2048: R(∞) = 14057/354025, A-stable, measured order 2.0069gamma 149/512: R(∞) = 697/22201, A-stable, measured order 2.0069gamma 1193/4096: R(∞) = 38801/1423249, A-stable, measured order 2.0069gamma 597/2048: R(∞) = 8249/356409, A-stable, measured order 2.0069gamma 1195/4096: R(∞) = 27193/1428025, A-stable, measured order 2.0069gamma 299/1024: R(∞) = 1337/89401, A-stable, measured order 2.0069gamma 1197/4096: R(∞) = 15593/1432809, A-stable, measured order 2.0069gamma 599/2048: R(∞) = 2449/358801, A-stable, measured order 2.0069gamma 1199/4096: R(∞) = 4001/1437601, A-stable, measured order 2.0069gamma 75/256: R(∞) = -7/5625, A-stable, measured order 2.0069gamma 1201/4096: R(∞) = -7583/1442401, A-stable, measured order 2.0069gamma 601/2048: R(∞) = -3343/361201, A-stable, measured order 2.0069gamma 1203/4096: R(∞) = -19159/1447209, A-stable, measured order 2.0069gamma 301/1024: R(∞) = -1559/90601, A-stable, measured order 2.0069gamma 1205/4096: R(∞) = -30727/1452025, A-stable, measured order 2.0069gamma 603/2048: R(∞) = -9127/363609, A-stable, measured order 2.0069gamma 1207/4096: R(∞) = -42287/1456849, A-stable, measured order 2.0069gamma 151/512: R(∞) = -751/22801, A-stable, measured order 2.0069gamma 605/2048: R(∞) = -14903/366025, A-stable, measured order 2.0069gamma 303/1024: R(∞) = -4447/91809, A-stable, measured order 2.0069gamma 607/2048: R(∞) = -20671/368449, A-stable, measured order 2.0069gamma 19/64: R(∞) = -23/361, A-stable, measured order 2.0069gamma 305/1024: R(∞) = -7327/93025, A-stable, measured order 2.0068gamma 153/512: R(∞) = -2191/23409, A-stable, measured order 2.0068gamma 307/1024: R(∞) = -10199/94249, A-stable, measured order 2.0068gamma 77/256: R(∞) = -727/5929, A-stable, measured order 2.0067gamma 155/512: R(∞) = -3623/24025, A-stable, measured order 2.0067gamma 39/128: R(∞) = -271/1521, A-stable, measured order 2.0066gamma 157/512: R(∞) = -5047/24649, A-stable, measured order 2.0065gamma 79/256: R(∞) = -1439/6241, A-stable, measured order 2.0064gamma 5/16: R(∞) = -7/25, A-stable, measured order 2.0062gamma 81/256: R(∞) = -2143/6561, A-stable, measured order 2.0059gamma 41/128: R(∞) = -623/1681, A-stable, measured order 2.0056gamma 83/256: R(∞) = -2839/6889, A-stable, measured order 2.0053gamma 21/64: R(∞) = -199/441, A-stable, measured order 2.005gamma 43/128: R(∞) = -967/1849, A-stable, measured order 2.0043gamma 11/32: R(∞) = -71/121, A-stable, measured order 2.0036gamma 45/128: R(∞) = -1303/2025, A-stable, measured order 2.0028gamma 23/64: R(∞) = -367/529, A-stable, measured order 2.0021gamma 3/8: R(∞) = -7/9, A-stable, measured order 2.0005gamma 25/64: R(∞) = -527/625, A-stable, measured order 1.999gamma 13/32: R(∞) = -151/169, A-stable, measured order 1.9976gamma 27/64: R(∞) = -679/729, A-stable, measured order 1.9963gamma 7/16: R(∞) = -47/49, A-stable, measured order 1.9952gamma 15/32: R(∞) = -223/225, A-stable, measured order 1.9936gamma 1/2: R(∞) = -1, A-stable, measured order 1.9929gamma 17/32: R(∞) = -287/289, A-stable, measured order 1.9937gamma 9/16: R(∞) = -79/81, A-stable, measured order 1.9963gamma 5/8: R(∞) = -23/25, A-stable, measured order 2.0105gamma 11/16: R(∞) = -103/121, A-stable, measured order 2.0514gamma 3/4: R(∞) = -7/9, A-stable, measured order 2.205gamma 13/16: R(∞) = -119/169, A-stable, measured order -2.7267

Source: side-track job sdirk.gamma_dyadic_scan, 9 ledger points, 81 distinct gammas plotted. Arithmetic: tableau and stability algebra exact over Fractions; the measured order is float64.

Every plotted gamma, in one table (81 rows)
gammaR at infinityA-stablemeasured order
1/32449no2.1001
1/1697no2.0912
3/32329/9no2.0835
1/817no2.0779
5/32217/25no2.0769
11/64761/121no2.0811
3/1641/9no2.096
13/64553/169no2.1735
7/32113/49no1.8356
29/1281609/841no1.9477
15/64353/225no1.9779
31/1281217/961no1.9912
1/41yes1.9982
33/128833/1089yes2.0023
67/2562953/4489yes2.0036
17/64161/289yes2.0046
69/2562201/4761yes2.0054
35/128457/1225yes2.0059
71/2561457/5041yes2.0064
143/5125089/20449yes2.0065
9/3217/81yes2.0066
145/5123617/21025yes2.0067
73/256721/5329yes2.0068
293/102410073/85849yes2.0068
147/5122153/21609yes2.0069
295/10247153/87025yes2.0069
37/12889/1369yes2.0069
593/204819873/351649yes2.0069
297/10244241/88209yes2.0069
595/204814057/354025yes2.0069
149/512697/22201yes2.0069
1193/409638801/1423249yes2.0069
597/20488249/356409yes2.0069
1195/409627193/1428025yes2.0069
299/10241337/89401yes2.0069
1197/409615593/1432809yes2.0069
599/20482449/358801yes2.0069
1199/40964001/1437601yes2.0069
75/256-7/5625yes2.0069
1201/4096-7583/1442401yes2.0069
601/2048-3343/361201yes2.0069
1203/4096-19159/1447209yes2.0069
301/1024-1559/90601yes2.0069
1205/4096-30727/1452025yes2.0069
603/2048-9127/363609yes2.0069
1207/4096-42287/1456849yes2.0069
151/512-751/22801yes2.0069
605/2048-14903/366025yes2.0069
303/1024-4447/91809yes2.0069
607/2048-20671/368449yes2.0069
19/64-23/361yes2.0069
305/1024-7327/93025yes2.0068
153/512-2191/23409yes2.0068
307/1024-10199/94249yes2.0068
77/256-727/5929yes2.0067
155/512-3623/24025yes2.0067
39/128-271/1521yes2.0066
157/512-5047/24649yes2.0065
79/256-1439/6241yes2.0064
5/16-7/25yes2.0062
81/256-2143/6561yes2.0059
41/128-623/1681yes2.0056
83/256-2839/6889yes2.0053
21/64-199/441yes2.005
43/128-967/1849yes2.0043
11/32-71/121yes2.0036
45/128-1303/2025yes2.0028
23/64-367/529yes2.0021
3/8-7/9yes2.0005
25/64-527/625yes1.999
13/32-151/169yes1.9976
27/64-679/729yes1.9963
7/16-47/49yes1.9952
15/32-223/225yes1.9936
1/2-1yes1.9929
17/32-287/289yes1.9937
9/16-79/81yes1.9963
5/8-23/25yes2.0105
11/16-103/121yes2.0514
3/4-7/9yes2.205
13/16-119/169yes-2.7267

Where one step's cycles go

Modeled cycles per SDIRK2 step by where they go, for each Jacobian policy. Neither count includes the rhs or Jacobian evaluations themselves; the table under the chart gives them per step beside each bar's total.
Newton iterationsLU factorizationother stage arithmeticfinite-difference Jacobian
0200400600800enzyme_qssaanalyticenzyme_qssa, analytic Jacobian: Newton iterations, 204 cycles per step (newton_iterations)enzyme_qssa, analytic Jacobian: LU factorization, 84 cycles per step (lu_factor)enzyme_qssa, analytic Jacobian: other stage arithmetic, 60 cycles per step (combine_b, form_m, stage_base); 348 in all, plus 6 rhs evaluations per step that the count leaves out348finite differenceenzyme_qssa, finite difference Jacobian: Newton iterations, 204 cycles per step (newton_iterations)enzyme_qssa, finite difference Jacobian: LU factorization, 84 cycles per step (lu_factor)enzyme_qssa, finite difference Jacobian: other stage arithmetic, 60 cycles per step (combine_b, form_m, stage_base)enzyme_qssa, finite difference Jacobian: finite-difference Jacobian, 24 cycles per step (fd_jacobian_arith); 372 in all, plus 8 rhs evaluations per step that the count leaves out372robertson_scaledanalyticrobertson_scaled, analytic Jacobian: Newton iterations, 342 cycles per step (newton_iterations)robertson_scaled, analytic Jacobian: LU factorization, 148 cycles per step (lu_factor)robertson_scaled, analytic Jacobian: other stage arithmetic, 108 cycles per step (combine_b, form_m, stage_base); 598 in all, plus 6 rhs evaluations per step that the count leaves out598finite differencerobertson_scaled, finite difference Jacobian: Newton iterations, 342 cycles per step (newton_iterations)robertson_scaled, finite difference Jacobian: LU factorization, 148 cycles per step (lu_factor)robertson_scaled, finite difference Jacobian: other stage arithmetic, 108 cycles per step (combine_b, form_m, stage_base)robertson_scaled, finite difference Jacobian: finite-difference Jacobian, 54 cycles per step (fd_jacobian_arith); 652 in all, plus 9 rhs evaluations per step that the count leaves out652servo_load_stepanalyticservo_load_step, analytic Jacobian: Newton iterations, 204 cycles per step (newton_iterations)servo_load_step, analytic Jacobian: LU factorization, 84 cycles per step (lu_factor)servo_load_step, analytic Jacobian: other stage arithmetic, 60 cycles per step (combine_b, form_m, stage_base); 348 in all, plus 6 rhs evaluations per step that the count leaves out348finite differenceservo_load_step, finite difference Jacobian: Newton iterations, 204 cycles per step (newton_iterations)servo_load_step, finite difference Jacobian: LU factorization, 84 cycles per step (lu_factor)servo_load_step, finite difference Jacobian: other stage arithmetic, 60 cycles per step (combine_b, form_m, stage_base)servo_load_step, finite difference Jacobian: finite-difference Jacobian, 24 cycles per step (fd_jacobian_arith); 372 in all, plus 8 rhs evaluations per step that the count leaves out372

Source: side-track job sdirk.jacobian_cost, 3 points; cycle model m0plus_fast. Arithmetic: float64 only; no Q15 effects are included.

Every plotted segment, in one table (21 rows)
problemJacobianwhere the cycles gocycles per stepstep totalrhs evaluations per step
enzyme_qssaanalyticNewton iterations2043486
enzyme_qssaanalyticLU factorization843486
enzyme_qssaanalyticother stage arithmetic603486
enzyme_qssafinite differenceNewton iterations2043728
enzyme_qssafinite differenceLU factorization843728
enzyme_qssafinite differenceother stage arithmetic603728
enzyme_qssafinite differencefinite-difference Jacobian243728
robertson_scaledanalyticNewton iterations3425986
robertson_scaledanalyticLU factorization1485986
robertson_scaledanalyticother stage arithmetic1085986
robertson_scaledfinite differenceNewton iterations3426529
robertson_scaledfinite differenceLU factorization1486529
robertson_scaledfinite differenceother stage arithmetic1086529
robertson_scaledfinite differencefinite-difference Jacobian546529
servo_load_stepanalyticNewton iterations2043486
servo_load_stepanalyticLU factorization843486
servo_load_stepanalyticother stage arithmetic603486
servo_load_stepfinite differenceNewton iterations2043728
servo_load_stepfinite differenceLU factorization843728
servo_load_stepfinite differenceother stage arithmetic603728
servo_load_stepfinite differencefinite-difference Jacobian243728

Measured against real counterparts, at matched accuracy

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.

Lane elites, ranked by cycles to tolerance

The archive fixes a cycle budget and measures the error a method reaches. The lane fixes an accuracy target on the axis-T ladder and measures the modeled cycles a candidate needs to reach it, so neither number converts into the other. Every run is float64. The pinned order-condition checker never saw these coefficients, so none is order-verified.

candidates measured
15,154
under the lane search code hash below
candidates ranked
77
the subset this document put in order; the leading 32 are listed below, the document's cap
source
rk-work/implicit_archive/elites.json
elite target
0.0009765625 final-state error
lane search code hash
d99e4d2dbe284c58
document written
cycle 1199, 2026-09-24 20:36 CT
The 32 entries plotted here, each at its a21 against the median modeled cycles it needed at the elite target, by Jacobian policy. Every plotted entry has gamma 15/16.
analytic Jacobianfinite difference Jacobian
-0.500.511.522.5222522502275230023252350a21median cycles at the elite targetrank 29, 6f1a9b4e86e38932: gamma 15/16, a21 1/32, analytic Jacobian, median 2292 cycles at the elite targetrank 3, fa97537319554fa4: gamma 15/16, a21 3/16, finite difference Jacobian, median 2292 cycles at the elite targetrank 25, 069ad01e0253b1c4: gamma 15/16, a21 5/16, analytic Jacobian, median 2292 cycles at the elite targetrank 4, 67edd925c6be74cb: gamma 15/16, a21 13/32, analytic Jacobian, median 2292 cycles at the elite targetrank 19, 389b50e08ac720cf: gamma 15/16, a21 17/32, analytic Jacobian, median 2292 cycles at the elite targetrank 7, fc39b04a4e72454f: gamma 15/16, a21 19/32, finite difference Jacobian, median 2292 cycles at the elite targetrank 13, 133fec9793a988ee: gamma 15/16, a21 5/8, analytic Jacobian, median 2292 cycles at the elite targetrank 21, f15d825048bccb57: gamma 15/16, a21 21/32, finite difference Jacobian, median 2292 cycles at the elite targetrank 8, 86ac7bbe1a248f56: gamma 15/16, a21 11/16, finite difference Jacobian, median 2292 cycles at the elite targetrank 15, 60b0436d36617396: gamma 15/16, a21 3/4, finite difference Jacobian, median 2292 cycles at the elite targetrank 23, b227649593685f17: gamma 15/16, a21 13/16, analytic Jacobian, median 2292 cycles at the elite targetrank 16, 5a6b77f3ae25626e: gamma 15/16, a21 31/32, analytic Jacobian, median 2292 cycles at the elite targetrank 32, a14ab7216ef0e65e: gamma 15/16, a21 17/16, analytic Jacobian, median 2292 cycles at the elite targetrank 1, 06e1f29458dd750f: gamma 15/16, a21 9/8, analytic Jacobian, median 2292 cycles at the elite targetrank 22, fb8d2f3c250b7bbd: gamma 15/16, a21 21/16, finite difference Jacobian, median 2292 cycles at the elite targetrank 24, 01e4a8c9819a50d3: gamma 15/16, a21 43/32, finite difference Jacobian, median 2292 cycles at the elite targetrank 26, ff34288badbbe1f0: gamma 15/16, a21 11/8, finite difference Jacobian, median 2292 cycles at the elite targetrank 11, 19ee189981109dfc: gamma 15/16, a21 45/32, analytic Jacobian, median 2292 cycles at the elite targetrank 20, 4b479ab96e6b80e2: gamma 15/16, a21 47/32, finite difference Jacobian, median 2292 cycles at the elite targetrank 28, 76cc74810d7fb334: gamma 15/16, a21 49/32, analytic Jacobian, median 2292 cycles at the elite targetrank 9, c9d8bf682781441f: gamma 15/16, a21 25/16, finite difference Jacobian, median 2292 cycles at the elite targetrank 2, bbc824514dbbc487: gamma 15/16, a21 13/8, finite difference Jacobian, median 2292 cycles at the elite targetrank 30, a88503e2477f2c4d: gamma 15/16, a21 53/32, finite difference Jacobian, median 2292 cycles at the elite targetrank 12, 9b82b33b5cbfc0be: gamma 15/16, a21 55/32, finite difference Jacobian, median 2292 cycles at the elite targetrank 17, 06c51ca12015352e: gamma 15/16, a21 57/32, finite difference Jacobian, median 2292 cycles at the elite targetrank 27, 8279882fa4390725: gamma 15/16, a21 29/16, analytic Jacobian, median 2292 cycles at the elite targetrank 14, b35265e162bc977e: gamma 15/16, a21 59/32, analytic Jacobian, median 2292 cycles at the elite targetrank 18, ddddc7baf7724544: gamma 15/16, a21 59/32, finite difference Jacobian, median 2292 cycles at the elite targetrank 5, ba3f68cfe1b56931: gamma 15/16, a21 15/8, finite difference Jacobian, median 2292 cycles at the elite targetrank 6, 08ef684cdd3c0ca2: gamma 15/16, a21 61/32, analytic Jacobian, median 2292 cycles at the elite targetrank 10, f1ca7c6f9b025ca1: gamma 15/16, a21 63/32, analytic Jacobian, median 2292 cycles at the elite targetrank 31, 93cd26808d2259a8: gamma 15/16, a21 63/32, finite difference Jacobian, median 2292 cycles at the elite target

Source: rk-work/implicit_archive/elites.json. float64 runs, cycles priced on design_estimate_div32; not order-verified.

Where the cycle levels sit

Analytic entries need 2292 cycles at a21 1/32 to 63/32; finite difference entries need 2292 cycles at a21 3/16 to 63/32.

Every plotted entry, in one table (32 rows)
rankkeygammaa21Jacobianmedian cycles at the elite target
106e1f29458dd750f15/169/8analytic2292
2bbc824514dbbc48715/1613/8finite difference2292
3fa97537319554fa415/163/16finite difference2292
467edd925c6be74cb15/1613/32analytic2292
5ba3f68cfe1b5693115/1615/8finite difference2292
608ef684cdd3c0ca215/1661/32analytic2292
7fc39b04a4e72454f15/1619/32finite difference2292
886ac7bbe1a248f5615/1611/16finite difference2292
9c9d8bf682781441f15/1625/16finite difference2292
10f1ca7c6f9b025ca115/1663/32analytic2292
1119ee189981109dfc15/1645/32analytic2292
129b82b33b5cbfc0be15/1655/32finite difference2292
13133fec9793a988ee15/165/8analytic2292
14b35265e162bc977e15/1659/32analytic2292
1560b0436d3661739615/163/4finite difference2292
165a6b77f3ae25626e15/1631/32analytic2292
1706c51ca12015352e15/1657/32finite difference2292
18ddddc7baf772454415/1659/32finite difference2292
19389b50e08ac720cf15/1617/32analytic2292
204b479ab96e6b80e215/1647/32finite difference2292
21f15d825048bccb5715/1621/32finite difference2292
22fb8d2f3c250b7bbd15/1621/16finite difference2292
23b227649593685f1715/1613/16analytic2292
2401e4a8c9819a50d315/1643/32finite difference2292
25069ad01e0253b1c415/165/16analytic2292
26ff34288badbbe1f015/1611/8finite difference2292
278279882fa439072515/1629/16analytic2292
2876cc74810d7fb33415/1649/32analytic2292
296f1a9b4e86e3893215/161/32analytic2292
30a88503e2477f2c4d15/1653/32finite difference2292
3193cd26808d2259a815/1663/32finite difference2292
32a14ab7216ef0e65e15/1617/16analytic2292
The 32 listed entries, the ranking rule and how the lane prices a cycle

Ranking rule, quoted from the document: lowest median cycles at the elite target across the problems where the status is reached; a candidate that reached the elite target nowhere ranks below every candidate that reached it somewhere; ties broken by the count of targets reached, highest leading, then by record_hash. No clock is read and no value is sampled, so the order is total and two runs over the same records agree.

Rank is the position the document gives under the rule above. Each row shows the candidate's shape, the median modeled cycles it needed at the elite target, how many problems it reached that target on, and the cost basis its cycles were priced on.

rankcandidategammaa21newton iterationsjacobianA-stablemedian cycles at targetproblems at targettargets reachedcost basis
106e1f29458dd750f15/169/82analytictrue22928 of 832 of 32design_estimate_div32
2bbc824514dbbc48715/1613/82finite_differencetrue22928 of 832 of 32design_estimate_div32
3fa97537319554fa415/163/162finite_differencetrue22928 of 832 of 32design_estimate_div32
467edd925c6be74cb15/1613/322analytictrue22928 of 832 of 32design_estimate_div32
5ba3f68cfe1b5693115/1615/82finite_differencetrue22928 of 832 of 32design_estimate_div32
608ef684cdd3c0ca215/1661/322analytictrue22928 of 832 of 32design_estimate_div32
7fc39b04a4e72454f15/1619/322finite_differencetrue22928 of 832 of 32design_estimate_div32
886ac7bbe1a248f5615/1611/162finite_differencetrue22928 of 832 of 32design_estimate_div32
9c9d8bf682781441f15/1625/162finite_differencetrue22928 of 832 of 32design_estimate_div32
10f1ca7c6f9b025ca115/1663/322analytictrue22928 of 832 of 32design_estimate_div32
1119ee189981109dfc15/1645/322analytictrue22928 of 832 of 32design_estimate_div32
129b82b33b5cbfc0be15/1655/322finite_differencetrue22928 of 832 of 32design_estimate_div32
13133fec9793a988ee15/165/82analytictrue22928 of 832 of 32design_estimate_div32
14b35265e162bc977e15/1659/322analytictrue22928 of 832 of 32design_estimate_div32
1560b0436d3661739615/163/42finite_differencetrue22928 of 832 of 32design_estimate_div32
165a6b77f3ae25626e15/1631/322analytictrue22928 of 832 of 32design_estimate_div32
1706c51ca12015352e15/1657/322finite_differencetrue22928 of 832 of 32design_estimate_div32
18ddddc7baf772454415/1659/322finite_differencetrue22928 of 832 of 32design_estimate_div32
19389b50e08ac720cf15/1617/322analytictrue22928 of 832 of 32design_estimate_div32
204b479ab96e6b80e215/1647/322finite_differencetrue22928 of 832 of 32design_estimate_div32
21f15d825048bccb5715/1621/322finite_differencetrue22928 of 832 of 32design_estimate_div32
22fb8d2f3c250b7bbd15/1621/162finite_differencetrue22928 of 832 of 32design_estimate_div32
23b227649593685f1715/1613/162analytictrue22928 of 832 of 32design_estimate_div32
2401e4a8c9819a50d315/1643/322finite_differencetrue22928 of 832 of 32design_estimate_div32
25069ad01e0253b1c415/165/162analytictrue22928 of 832 of 32design_estimate_div32
26ff34288badbbe1f015/1611/82finite_differencetrue22928 of 832 of 32design_estimate_div32
278279882fa439072515/1629/162analytictrue22928 of 832 of 32design_estimate_div32
2876cc74810d7fb33415/1649/322analytictrue22928 of 832 of 32design_estimate_div32
296f1a9b4e86e3893215/161/322analytictrue22928 of 832 of 32design_estimate_div32
30a88503e2477f2c4d15/1653/322finite_differencetrue22928 of 832 of 32design_estimate_div32
3193cd26808d2259a815/1663/322finite_differencetrue22928 of 832 of 32design_estimate_div32
32a14ab7216ef0e65e15/1617/162analytictrue22928 of 832 of 32design_estimate_div32

How this lane prices a cycle

design_estimate_div32

direction
estimate, unsigned
excludes
derivative and analytic-Jacobian evaluations, and every branch. The LU rests on sdirk.DIV_CYCLES = 32, a stated guess for a software reciprocal, because the cost model prices no divide
grade
a design estimate from the EPOCH3-DESIGN solver terms, not an assembly-verified count
includes
forming M = I - h*gamma*J, one LU factorization per step, the stage combination, the fixed Newton iterations with their back substitutions, the combination over b, and the finite difference assembly arithmetic when the Jacobian is not analytic
source
sdirk.estimate_sdirk2_cycles

What these numbers are not

Every measured point (47 points)

One table per job and one row per point, from its newest ledger line. A point counts as measured only under the code hash that measured it, so an older line from earlier code is left out; that rule and the retry policy are in the measurement ledger. Plan, job catalogue and invariants: docs/SIDETRACK-AUTOMATION.md in the harness repository. Designs these feed: docs/EPOCH2-DESIGN.md and docs/EPOCH3-DESIGN.md.

sdirk.gamma_a21_scan

25 points

Closes: EPOCH3-DESIGN: the L-stability window in J4 is 17 candidates wide and nothing pins that width, and the second free parameter of the 2-stage SDIRK is never varied

Arithmetic: tableau and stability algebra exact over Fractions; the measured order is float64

A = [[gamma, 0], [a21, gamma]], c is the row sum, b is solved from the two order-2 conditions, so b2 = (1/2 - gamma) / a21. The numerator of R(z) works out to (1, 1 - 2 gamma, 1/2 - 2 gamma + gamma^2) whatever a21 is, so |R(inf)|, A-stability and L-stability are functions of gamma alone. The a21 axis moves the order-3 residuals, stiff accuracy, the exactness of b, its CSD cost and the measured order.

pointcyclea21_valuesa_stable_gammasb_exact_pairsgammasl_stable_gammasmin_abs_r_infmin_order3_residualpairsr_inf_a21_invariantstiffly_accurate_pairs
s04_t046032121441500.280.109375480true0
s04_t066065121251500.280.105469975true0
s04_t08606512711500.280.104492975true0
s04_t10606512651500.280.116292975true0
s04_t12606512651500.280.120687975true0
s06_t046032485966300.06371190.1087242016true0
s06_t066065484256300.06371190.1047364095true0
s06_t086065481136300.06371190.1044314095true0
s06_t10606548716300.06371190.1129764095true0
s06_t12606548656300.06371190.1144614095true0
s08_t046032446156500.001244440.1077472080true0
s08_t066065444446500.001244440.1047064225true0
s08_t08606544756500.001244440.1042024225true0
s08_t10606544116500.001244440.1121724225true0
s08_t1260654416500.001244440.113724225true0
s10_t046032656146500.001244440.1075032080true0
s10_t066065654436500.001244440.1046774225true0
s10_t08606565976500.001244440.1044314225true0
s10_t10606565206500.001244440.1142324225true0
s10_t1260656556500.001244440.1183524225true0
s12_t046032656116500.001244440.1082362080true0
s12_t066065654416500.001244440.1060044225true0
s12_t08606565946500.001244440.1056524225true0
s12_t10606565166500.001244440.1183524225true0
s12_t1260656526500.001244440.1230214225true0

sdirk.gamma_dyadic_scan

9 points

Closes: EPOCH3-DESIGN: L-stability is to be checked numerically for the snapped gamma, and the NOT_L_STABLE threshold is a placeholder

Arithmetic: tableau and stability algebra exact over Fractions; the measured order is float64

gamma is snapped to a dyadic, a21 is fixed at 1 - gamma so that c2 = 1, and b is solved from the two order conditions. Stiff accuracy and order 2 can only hold together at the irrational gamma, so order is the one kept.

pointcyclecandidatesa_stablebest_gammabest_measured_orderbest_r_at_infinityl_stable
s046013105/162.0062-0.280
s056017109/322.00660.2098770
s0660171219/642.0069-0.06371190
s0760171419/642.0069-0.06371190
s0860171775/2562.0069-0.001244440
s0960171775/2562.0069-0.001244440
s1060171775/2562.0069-0.001244440
s1160171775/2562.0069-0.001244440
s1260171775/2562.0069-0.001244440

sdirk.jacobian_cost

3 points

Closes: EPOCH3-DESIGN: the optional analytic Jacobian field is unpriced because every side-track SDIRK number is costed with the finite difference

Arithmetic: float64 only; no Q15 effects are included

the analytic Jacobian evaluation is excluded from the cycle count by the same convention costmodel.cycle_count uses for rhs evaluations, since both are application-supplied code. The reported saving is therefore the finite-difference assembly arithmetic plus the n extra rhs evaluations per step that assembly needs.

pointcyclebest_error_analyticbest_error_finite_differencecycles_analyticcycles_finite_differencecycles_savedf_evals_savedmax_abs_diffmax_rel_diff
enzyme_qssa1202.76241e-062.76241e-063483722429.69249e-091.69982e-10
robertson_scaled120n/an/a5986525433.72529e-073.72529e-07
servo_load_step603.46048e-063.46048e-063483722429.93411e-091.02767e-09

sdirk.newton_iters

4 points

Closes: EPOCH3-DESIGN: three Newton iterations is the prototype's setting, not a ruling

Arithmetic: float64 only; no Q15 effects are included

EPOCH3-DESIGN.md: three iterations is the prototype's setting, not a ruling; the count enters the pinned evaluator config at the epoch boundary, so it should be chosen on data

pointcyclecycles_per_step_2_statesnewton_itersorder_estimatestiff_best_errors
it16023612.0069enzyme_qssa 3.21588e-06; robertson_scaled 4.06153; servo_load_step 3.46048e-06
it26030422.0069enzyme_qssa 2.76307e-06; robertson_scaled n/a; servo_load_step 3.46048e-06
it36037232.0069enzyme_qssa 2.76241e-06; robertson_scaled n/a; servo_load_step 3.46048e-06
it46044042.0069enzyme_qssa 2.76241e-06; robertson_scaled n/a; servo_load_step 3.46048e-06

sdirk.stiff_suite

3 points

Closes: EPOCH3-DESIGN: the stiff case rests on rc_thermal and a synthetic two-rate system, while the suite now has three purpose-built stiff problems

Arithmetic: float64 only; no Q15 effects are included

cycle estimates exclude rhs and Jacobian evaluations, matching costmodel.cycle_count; SDIRK2 is priced with a finite-difference Jacobian because the validation problems supply no analytic one

pointcycleexplicit_finishersfinishersrk4_min_stable_nsdirk2_best_errorsdirk2_min_stable_n
enzyme_qssa60nonesdirk2n/a6.89217e-078
robertson_scaled60nonesdirk2n/a1.0198e-05256
servo_load_step60nonesdirk2n/a8.63721e-078

sdirk.stiff_suite_budget

3 points

Closes: EPOCH3-DESIGN: the stiff ladder in J3 stops at 256 steps while the harness scores at a matched cycle budget, where the cheap explicit methods take thousands of steps

Arithmetic: float64 only; no Q15 effects are included

cycle estimates exclude rhs and Jacobian evaluations, matching costmodel.cycle_count; SDIRK2 is priced with a finite-difference Jacobian because that is how every other side-track SDIRK number is costed. The budget is not read from any results file: an artifact that depended on host state would stop being reproducible.

pointcyclebudget_cycleserror_at_budgetfinishers_at_budgetsdirk2_error_at_budgetsteps_at_budget
enzyme_qssa6065536euler 1.89045e-05; heun2 3.42268e-08; midpoint 2.09683e-08; rk4 2.72208e-10; sdirk2 1.45953e-06euler, heun2, midpoint, rk4, sdirk21.45953e-06euler 6553; heun2 2520; midpoint 2978; rk4 618; sdirk2 176
robertson_scaled6065536euler 8.80946e-05; heun2 3.81289e-07; midpoint 3.16523e-07; rk4 n/a; sdirk2 n/aeuler, heun2, midpointn/aeuler 4369; heun2 1680; midpoint 1985; rk4 412; sdirk2 100
servo_load_step6065536euler 2.3216e-05; heun2 3.67262e-08; midpoint 2.62928e-08; rk4 n/a; sdirk2 1.82872e-06euler, heun2, midpoint, sdirk21.82872e-06euler 6553; heun2 2520; midpoint 2978; rk4 618; sdirk2 176
Further evidence, not published here
documentstatewhat it holds
rk-work/validation/axes.jsonnot written yetthe cycles-to-tolerance rows for this class, next to the explicit class on the same accuracy targets
rk-work/implicit_archive/YYYY-MM-DD.jsonlwrittenthe lane's per-day search log: one line per candidate the lane enumerated, ranked against nothing
rk-work/schedule/shares.jsonwrittenthe per-cycle lane log summarised per method class

The traceability rule lists key_findings.json, validation/results.json, benchmark/results.json, the side-track ledger with its artifacts, and the two lane elites documents, adaptive_archive/elites.json and implicit_archive/elites.json. Each document named above sits outside that list and says so in its own schema, so this page names it and reports whether it exists without publishing a number from it. Only the owner can add one to the list.