rk-harness findings

Explicit methods scored in Q15 fixed point, and implicit and adaptive methods measured outside the archive.

An automated search for Runge-Kutta methods that hold up in Q15 fixed-point arithmetic on small microcontrollers. Cost is counted against a modeled Cortex-M0+ cycle budget rather than measured on the chip. This site is rebuilt from the run's data at the end of every cycle, and nobody edits it by hand. The project overview explains why the project exists and how it works.

Built from cycle 3,930 of the run, 141,364 archive records. Newest elite recorded 2026-09-08 22:30 CT, about 35 hours before the last saturation check at 2026-09-10 09:41 CT.

The three classes

A scored record has room for one explicit tableau and nothing else, so the implicit and adaptive classes are measured outside the archive and never ranked against it.

explicit
65 of 96 run
matched-accuracy targets reached by 11e898cbab59 in Q15, rk4 in Q15 and rk4_float64 in float64; 96 of 96 rows ran, from benchmark/results.json

Scored, archived and ranked at an equal cycle budget.

Q15 fixed point, checked by the pinned verifier.

explicit methods

implicit
44 of 44 run
matched-accuracy targets reached by sdirk2_analytic_jac in float64 and sdirk2_fd_jac in float64; 44 of 64 rows ran, the other 20 rows were skipped for one recorded reason (no analytic Jacobian is derived for this problem), and the rows that ran cover all 8 problems and reached every target, from benchmark/results.json

Measured because the explicit class runs out of stability before it runs out of budget: on 1 of 3 stiff validation problems every discovered explicit tableau overflows Q15 (validation/results.json).

float64; not order-verified and not scored.

implicit methods

adaptive
47 of 64 run
matched-accuracy targets reached by bs32_float in float64 and bs32_q15 in Q15; 64 of 64 rows ran, from benchmark/results.json

A record has no room for the second b vector an error estimate needs.

float64, with one fixed pair also run in Q15; not order-verified and not scored.

adaptive methods

Each card counts its own class's rows in benchmark/results.json. The value is the targets reached out of the rows that ran, and the figure beside it is how many of the class's rows ran at all: a row that never ran is not a target the class missed. Ours on these rows means run by this harness through its own solver path, which includes the classical rk4 control, so it does not mean discovered.

All three cards read benchmark/results.json, which runs the tableaus named in validation/results.json. That document was built at 141,364 archive records, so these three numbers stand at that state and not at the archive's current one.

The three classes side by side

explicitimplicitadaptive
method familyExplicit fixed-step Runge-Kutta: a strictly lower triangular A and one b vectorTwo-stage SDIRK over a dyadic grid of gamma and a21, with a Newton iteration count and a Jacobian policyAn embedded pair on the dyadic lattice (order 3 with an order-2 estimate) and a step-size controller
how candidates are foundMAP-Elites cells filled by exhaustive enumeration and CMA-ES islandsDeterministic enumeration, plus the side-track jobsDeterministic enumeration, plus the side-track jobs
the question askedWhat error does a method reach inside a fixed cycle budget?How many modeled cycles does a method need to reach an accuracy target? (cycles to tolerance)How many modeled cycles does a method need to reach an accuracy target? (cycles to tolerance)
problemsThe frozen search set and held-out setThe validation problemsThe validation problems
arithmeticQ15 through the pinned solverfloat64: there is no Q15 LU factorizationfloat64; the Q15 adaptive solver runs one fixed pair
what checks a resultThe pinned verifier, order conditions included, then scoring on the held-out setNothing: not order-verified and not scoredNothing: not order-verified and not scored
cost basisThe pinned cost modelA design estimate that assumes a division costFloat64 attempt counts, each attempt priced by the Q15 cost model
source documentsrk-work/archive, validation/results.json, benchmark/results.jsonimplicit_archive/elites.json, the side-track ledger, validation/results.json, benchmark/results.jsonadaptive_archive/elites.json, the side-track ledger, benchmark/results.json

Numbers from different columns do not compare, so each class tab carries its own.

Python wall clock on a desktop preserves the cycle model's ordering (Pearson r 0.970 over 56 fixed-step Q15 runs), and the cycle counts are modeled rather than measured on a Cortex-M0+. In the benchmark head-to-head the champion tableau (11e898cb) runs in 20.47 µs per Q15 step against 31.99 µs for rk4, each a median over that method's own 7 problems. The per-step speedup of 1.452x is the geometric mean of the per-problem ratios, so it is not the quotient of those two medians. Details are in the measured speed section.

Epoch status

Epoch 1 frozen scored method class: explicit fixed-step Runge-Kutta

frozen at
2026-09-14 23:33 CT
reason
Frozen for a correction to the coefficient cost rule (DECISIONS D45). Every number on these pages was scored under verifier hash de5bec22 and keeps that meaning.
last progress
2026-09-08 22:30 CT: an acceptance at the heldout_verified tier
saturation counter
0 consecutive saturating checks; 6 trigger a freeze
last check
2026-09-10 09:41 CT, verdict CONTINUE
falsification file
present

Progress is a record in an empty cell, an improved elite or a heldout_verified acceptance. With no progress inside the saturation window and the falsification file written, a check counts as saturating, and enough in a row reach the freeze threshold. Starting the next epoch, and re-pinning the verifier for it, is a person's decision. Times are stored in UTC and shown in US Central.