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.
Components
- Watchdog. A PowerShell script on the host kills the container on a stale heartbeat, stops it on language-model API spend or low disk, pauses it on battery power or foreground CPU load, and pushes the run's repositories to GitHub. Epoch 2 lost five days to a stop that nothing restarted, so since decision D46 the watchdog resumes the stops it made itself.
- Stats file. The host writes
stats.txt, not the container, so the file can still report a dead container. It never carries a value forward, and it claims nothing about the container unless Docker answered. - Start command. Nothing starts the run at sign-in. After a reboot a person starts it by hand with one command,
start-integration-harness, which brings up the container and the watchdog. - Container. The harness is mounted read-only, so the search cannot change its own scorer. No GitHub credential enters the container: it commits, and the host pushes. The one credential inside is the model's sign-in file, mounted read-only, and the model runs in a read-only sandbox.
- Verifier. A sha256 over 14 pinned files is checked at every start and stored in every record, so changing a pinned file opens a new epoch instead of editing the old one. Epoch 1's records carry
de5bec22, over 10 files, apart from 21 written minutes before that pin was set. Epoch 2 runs under2db0816cover 14; it started after a trace of the compiled step showed that the old cost model hadrk4andrk38in the wrong order. - Golden gate. Before any cycle runs, 91 golden and canary cases must pass. The full suite collected 1,908 tests on 2026-09-23.
- Runner. Each cycle starts by rebuilding its state from the archive files, so a crash costs at most the cycle in progress. It then runs the directive, search, verify, score and append steps, updates the hypothesis ledger, rebuilds the findings site and commits.
- Directive. A language model writes each cycle's direction as a JSON directive that can only narrow the search, and an unknown key sends the runner to a fixed fallback. The model sees held-out errors, and elites are picked on them.
- Search. Early phases enumerate a lattice of dyadic coefficients exhaustively, and later phases run CMA-ES over the stage matrix with the weights solved exactly.
- Verify. The pinned verifier checks the order conditions in exact rational arithmetic, along with explicitness, row sums, coefficient range, Q15 overflow and stability.
- Score. Each surviving candidate runs in Q15 with floor rounding at a 65,536-cycle budget, on the search problems and then on the held-out problems
pendulum,dc_motor,rc_thermalandquaternion. Cycle costs come from a model of the Cortex-M0+, not from a chip. - Archive. Scored candidates are appended to per-day JSON-lines files in
rk-work. For each order, a MAP-Elites grid keeps the lowest held-out error in each cell of stage count and cost band. Epoch 1 wrote 141,364 records. - Lanes and side tracks. A schedule rotates the explicit lane with adaptive and implicit lanes. Those lanes measure float64 cycles-to-tolerance, are not order-verified, and are never ranked against archive errors. Side tracks answer a fixed catalogue of design questions off the archive.
- Hypothesis ledger. Each hypothesis is a predicate in a closed grammar, read by a hand-written parser. Code assigns the verdicts, not the model.
- Literature digests. The model writes web digests that feed later prompts, and the findings site publishes them labelled as model-written.
- Findings site. The container rebuilds the findings site every cycle. The build is deterministic and uses no JavaScript, and a single banned word blocks it.
- Overview site. The overview site is built by hand from a snapshot of the run's data, so it goes stale until someone rebuilds it.
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.
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
- Generator. An MLP maps noise to table entries, but its optimizer is created and never stepped, and its raw outputs never pass validation. Every candidate slot therefore falls back to a random generator that reseeds to 42, so trials 1 to 4, 6, 8 and 9 saved the same four-stage A and b and trials 5 and 7 the same six-stage A and b.
- Surrogate. An MLP is trained by gradient descent on the scored candidates, at start-up and periodically after, and the checkpoints of trials 12 to 15 record its loss in 98 of their 100 epochs. Nothing in the training loop reads its predictions.
- Evolution. Trials 12 to 15 ran a genetic search whose population starts with
RK4or Dormand-Prince plus perturbed copies of it. Selection, crossover and mutation ran and the population moved off the seed, but every score tied at the clip, so the saved table stayed the seed: trials 12, 14 and 15 savedRK4exactly, and trial 13 saved Dormand-Prince exactly. - Random sampling. Trial 16 drew every candidate at random, with no mutation or crossover, and saved the top scorer. Its fitness read an accuracy field that does not exist, so accuracy carried no weight.
- Stepper. A fixed-step stepper reads the diagonal only while that stage is still zero and never reads the upper triangle, so an implicit table runs as its strictly lower part. Every candidate was explicit, so the gap hit only the two Gauss-Legendre baselines, which ran as truncated explicit methods of order one.
- Reference solver. Dormand-Prince (SciPy RK45) supplies each reference solution but handles only 5 of the 6 ODE families in the final test and 5 of the 11 in training, so
robertsonproblems fail for every method in the final test. Those failures, plus polynomial ODEs that blow up, give every method the same 81.1% success rate. - Composite score. Weighted accuracy, efficiency and stability terms are summed, and the total is clipped at 1. All 10,000 rows of the one training log score 1, and the checkpoints of trials 12 to 15 record a best and a mean score of 1 in all 100 of their epochs, so the score cannot tell
RK4from a random table. - Best-table tracker. The saved table changes only on a strictly higher score, so once scores tie at the ceiling the earliest valid candidate stays best. In the evolution trials that candidate is the seeded baseline.
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.