How the two projects are built

The rk run is an autonomous search: no person chooses what it tries or how it scores, though people start it, stop it and deploy changes to its unpinned code. It is built as a container that searches and scores, a host that keeps it running and publishes its data, and two websites that read the results. The 2025 ML project was a training pipeline that generated Butcher tables, integrated generated ODEs with them and scored them against a reference solver. Its generator was never trained, and because its score was clipped, its evolutionary search saved the textbook tables it was seeded with.

The rk run's system

Everything that decides a score sits inside the container behind a pinned hash, and everything that keeps the run alive or publishes it sits on the host. The runner's cycle is the explicit lane, the only one whose candidates are scored into the archive.

Architecture of the rk runThree groups, stacked top to bottom. The host machine holds a stats file written on the host, a start command that a person runs by hand after a reboot, and a watchdog with a pause guard. The watchdog can kill, stop, pause and resume the container, and it pushes the outputs to GitHub. In the container, the harness is mounted read-only with no GitHub credentials inside, the verifier's sha256 over 14 pinned files is checked at start, and 91 golden and canary cases must pass. The runner then repeats one cycle: a directive from a model or a fixed fallback, a search by enumeration or CMA-ES, verification with exact rational order conditions, scoring in Q15 with floor rounding at a 65,536-cycle budget, and an append to the archive. Beside the runner sit the lanes and side tracks, the hypothesis ledger and the literature digests. The outputs are the rk-work archive, the findings site rebuilt from it every cycle, and the overview site built from it by hand.host machinerunner: one cyclecontainer (Docker)outputs (GitHub repos)stats filewritten on the host,so a dead containerstill gets reportedstart commandrun by hand aftera reboot; starts therun and the watchdogwatchdogpause guard: battery or foreground loadpushes to GitHub; resumes its own stopsharnessmounted read-only; no GitHub credentials insideverifiersha256 pinned over 14 files, checkedat every start, stored in every recordgolden gate91 golden and canary cases must passbefore any cycle runsdirective: model or fixed fallbacksearch: enumerate, or CMA-ESverify: exact rational order conditionsscore: Q15 floor, 65,536-cycle budgetappend: each result to the archivelanes andside trackskept offthe archivehypothesisledgerverdicts setby codeliteraturedigestswritten bythe modelrk-work archiveappend-only JSON lines, one file per dayfindings siterebuilt every cycle,no JavaScriptoverview sitebuilt by hand fromthe run's datakill, stop, pause, resumenext cycleevery cyclepush to GitHub
The rk run on 2026-09-23. Arrows show which part starts, controls or feeds which; the three boxes under the runner work beside the explicit lane.

Components

The 2025 ML project's pipeline

Novel-Numerical-Integration-Methods proposed Butcher tables in three ways and scored all of them with one stepper and one composite score. A surrogate model trained beside them, but nothing read its predictions.

Pipeline of the 2025 ML project and where it brokeThree sources propose Butcher tables: a generator (trials 1 to 11) that is never trained and falls back to a random table reseeded to 42; evolution (trials 12 to 15) whose population is seeded with RK4 or Dormand-Prince; and random sampling in trial 16. The candidate tables and the baselines go to a fixed-step stepper that uses only the explicit part of a table, and its results are compared with a reference solver, Dormand-Prince (SciPy RK45), which handles 5 of the 6 test families. The error feeds a composite score clipped at 1, which feeds a best-table tracker and, through selection, the evolution loop. The scores also train a surrogate model by gradient descent, and nothing reads its output. Labels starting with "Flaw:" mark breaks at the generator, the evolution seed, trial 16's fitness, the stepper, the reference solver's coverage, the clipped score, the tracker and the surrogate.generator (trials 1 to 11)an MLP whose optimizer is never steppedfallback: a random table reseeded to 42Flaw: never trained (N4)evolution (trials 12 to 15)population seeded with RK4 orDormand-Prince and perturbed copiesFlaw: saved table stays the seed (N2)random sampling (trial 16)random tables, no mutation or crossoverFlaw: fitness read a missing field (N7)stepperfixed step; usesonly the explicitpart of a tableFlaw: a bug (N5)reference solverDormand-Prince(SciPy RK45),5 of 6 test familiesFlaw: artifact (N6)composite scoreweighted accuracy, efficiency, stabilityclipped at 1: all 10,000 logged scores are 1Flaw: nothing left to select on (N2, N3)best-table trackerreplaces the saved table only on astrictly higher scoreFlaw: ties keep the seeded table (N2)surrogatean MLP trained by gradient descenton the scored candidatesFlaw: nothing reads its output (N4)candidate tableserror vs referencetrains on scoresselection
The pipeline in Novel-Numerical-Integration-Methods at commit d89c2c809ee0. A heavy outline and a line starting with "Flaw:" mark each part that broke, with its claim id in parentheses.

Components

What still holds

The harness runs end to end for explicit tables: the stepper runs any explicit table, the training generator covers 11 ODE families including stiff ones, and the final benchmark ran 12 methods on 10,000 ODEs, 8,113 of them scorable. Trials 12 to 15 reproduce the RK4 and Dormand-Prince errors exactly, 0.14 and 0.0791, so the evaluation gives the same number for the same table.

Most of what broke was in the training code and the score: a generator optimizer that was never stepped, a fitness that read a missing field and a score clipped at 1. Two defects sit inside single parts: the stepper runs only the explicit part of a table, and the reference solver covers 5 of the 6 test families.