Predicates resolved by code, and the model's own readings and reading list.
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.
The hypothesis ledger: predicates the planning model committed to before the data could answer them, each resolved by code. Open a row for the statement, the proposed mechanism and the control that would have shown the mechanism wrong. Below the ledger are the model's readings of the archive and its literature digests.
Each row pairs a machine-checkable predicate over per-cell statistics, for example slow.p3s4.heldout < slow.p4s4.heldout, with the mechanism and control the model recorded before the data could answer. p and s name an (order, stages) cell, the leading word picks the cost model, and heldout in a predicate is that cell elite's error on the held-out set, which selects elites and so is not a clean test set.
Verdicts come from code, never from the model. Once every cell a predicate names holds min_samples records, the predicate is evaluated against those cells' running statistics. d is Cohen's d; below 0.2 the verdict is inconclusive whichever way the comparison went, so a weak effect cannot be claimed. A predicate that names an empty cell is inconclusive too, because absence of evidence never refutes. n is the smallest record count among the named cells, and c is the cycle the verdict landed on (c123+ means proposed at cycle 123 and still open).
×N means the model posed that predicate N times over the run. The row shows the newest verdict and lists the id of every posing; mixed beside the verdict means those posings did not all get the same one.
Source: rk-work/hypotheses.jsonl; exact counts.
| verdict group | distinct predicates |
|---|---|
| supported | 110 |
| refuted | 81 |
| inconclusive | 90 |
| open | 87 |
90 older predicates are not shown here. The whole log is in rk-work/hypotheses.jsonl.
61 older predicates are not shown here. The whole log is in rk-work/hypotheses.jsonl.
70 older predicates are not shown here. The whole log is in rk-work/hypotheses.jsonl.
67 older predicates are not shown here. The whole log is in rk-work/hypotheses.jsonl.
Model-written text, in this section and the next. The sources are citations the model collected itself, so check them before relying on them.
Readings of the archive the model wrote during the run, newest at the top. Each one comments on the numbers as they stood at its cycle, so where the text and the tables disagree, trust the tables. Unlike hypotheses these readings carry no predicate, so code cannot catch them being wrong.
At cycle 2574, the archive fills 16 of 40 searchable order-two cells, 9 of 32 order-three cells, and 5 of 24 order-four cells. No searchable order-one cell is filled, though one candidate lies outside its stage range. The lowest held-out error shown is 0.01223 for an order-two, five-stage candidate. The best shown order-three and order-four held-out errors are 0.02286 and 0.02787. These are results for the archived candidates, not a general ranking of method orders.
Search error often understates held-out error. One order-three, five-stage cell has search error 0.003999 but held-out error 0.06314; an order-four, six-stage cell has search error 0.002124 and held-out error 0.02787. More stages also do not consistently improve held-out error within an order. Fixed-point rounding, coefficient arithmetic, and the problems used for search and evaluation could each contribute, but the archive does not isolate their effects. The recent cross-model hypotheses listed in the run are marked refuted by the harness; their combined predicates do not show which individual comparison failed.
The literature digest supports judging error together with execution time under a fixed sampling budget. It also cautions that reduced precision can alter observed convergence and stability. For Cortex-M0+, multiplication can take either 1 or 32 cycles depending on the implementation, while memory access and coefficient arithmetic add costs that stage count alone misses. The archive’s cycle figures are therefore useful comparisons within this run, but they do not establish timing on a particular device. The digest’s discussion of implicit methods reinforces the same limit: stage counts omit iteration and linear-solve work.
The current order-four, four-stage bucket-three cell remains empty, and the stall counter is 13. A useful next check is whether candidates in that cell can meet the archive criteria, followed by repeated held-out evaluations of the strongest occupied cells. Tracing rounding and overflow alongside measured cycles on the intended hardware would help distinguish fixed-point effects from search-set mismatch and clarify whether the apparent error gains fit the available time budget.
The archive is sparse. Order 2 fills 16 of 40 searchable cells, order 3 fills 9 of 32, and order 4 fills 5 of 24; order 1 has no filled searchable cells, though one result sits outside the stage range. The lowest recorded held-out error is 0.01291 in an order-2, six-stage cell. The best recorded values for orders 3 and 4 are 0.02286 and 0.02787, both at six stages. These are results from the occupied cells, not a ranking of what each order could achieve across the empty cells.
Search error often understates held-out error substantially. For example, the order-4, six-stage bucket-3 cell has search error 0.002124 and held-out error 0.02787. That gap is consistent with candidates fitting the search cases more closely than the held-out cases, although the archive alone does not identify the cause. Fixed-point rounding, shifts, coefficient representation, and overflow behavior are possible contributors. Cycle counts also vary within a stage count and overlap across stage counts, so stage count by itself does not explain the measured cost.
The literature digest gives reasons to check those mechanisms directly. On a Cortex-M0+, multiplication can take either 1 or 32 cycles depending on the implementation; shifts, memory access, and register use also matter. Arithmetic shifts round negative values downward, and multiplication retains only the low 32 bits, both of which can affect fixed-point results. The digest also notes that implicit Runge, Kutta methods may incur stage solves, Jacobian work, and linear solves. The archive does not say whether its candidates use those structures, so those costs should only be applied where relevant.
Several recent cross-cost-model hypotheses are recorded as refuted, but the supplied grid does not show the model-specific measurements needed to explain each verdict. With many empty cells and a stall counter of 42, it remains unclear how much of the pattern reflects attainable methods versus search coverage. Useful next checks are the held-out cases with the largest search gaps, the instruction and arithmetic traces behind cycle counts, and targeted searches in empty neighboring cells.
At cycle 2499, the archive fills 16 of 40 second-order cells, 9 of 32 third-order cells, and 5 of 24 fourth-order cells. No earliest-order cell is filled within the searchable stage range, though one entry lies outside it. The lowest displayed held-out errors are 0.01291 for a six-stage second-order cell, 0.02286 for a six-stage third-order cell, and 0.02787 for a six-stage fourth-order cell. In this archive, a higher order label has not consistently meant a lower held-out error.
Search error and held-out error often separate. For example, the five-stage second-order cell in bucket 3 has search error 0.001701 and held-out error 0.03179, while its six-stage bucket 2 counterpart has higher search error, 0.008088, but lower held-out error, 0.01291. Q15 coefficient quantization, floor rounding, and a fixed cycle budget can change which tableau performs well beyond the search cases. The recent cross-model ranking hypotheses listed here were refuted, which further limits what can be inferred from a ranking under one cost model.
The literature digest explains why comparing tableaus within the same order and stage count is useful: order conditions can leave free coefficients that affect error. Published coefficient objectives do not directly predict held-out error under this fixed-point evaluation. It also cautions that stage count alone cannot represent cost, particularly for implicit methods, where solver iterations, Jacobian work, and linear solves matter.
Many cells remain empty, and the run is stalled at the five-stage, bucket 3 cell, so the observed rankings may change as coverage grows. The next useful checks are repeated held-out evaluations of close candidates, comparisons across the cost models using the same tableaus, and measurements of the work included in each cycle count. Those checks would help distinguish persistent accuracy differences from sensitivity to the test set or cost assumptions.
The archive fills 30 searchable cells: 16 for order 2, nine for order 3, and five for order 4. No order-1 cell is filled within the searchable stage range, though one entry lies outside it. The lowest recorded held-out error is 0.01291 for an order-2, six-stage method in bucket 2. The best recorded values for orders 3 and 4 are 0.02286 and 0.02787, both at six stages. These are results for the tableaus found so far, not a general ranking of the orders.
Search error often understates held-out error. For example, the order-2, five-stage entry in bucket 3 has search error 0.001701 and held-out error 0.03179, while its bucket-2 counterpart has higher search error, 0.002731, but lower held-out error, 0.01356. Quantized coefficients, Q15 multiplication with floor rounding, and errors accumulated across stages and steps can all affect performance on cases outside the search set. The occupied cells also show a cycle and error tradeoff: the lowest held-out errors occur among methods using more stages and, in these entries, more cycles.
The literature digest gives reasons to compare individual tableaus rather than rely on order labels. Order conditions can leave coefficients free, and earlier optimization work uses that freedom to reduce error or meet other constraints. The fixed-point studies show that rounding rules and intermediate precision can change integration results even when the method and step size stay fixed. Those findings are consistent with the archive’s variation within each order, but they do not predict its Q15 held-out rankings or cycle costs.
Many cells remain empty, and a 105-cycle stall limits what can be inferred from the current archive. The recent cross-cost-model hypotheses listed in the run state were refuted; 83 hypotheses remain open. A useful next examination would compare repeat evaluations of the leading tableaus on held-out cases, then inspect how coefficient quantization, floor-rounded products, and cost-model assumptions change their error and cycle rankings.
At cycle 2424, the archive fills 30 of 136 searchable cells: 16 for order 2, nine for order 3, and five for order 4. No order 1 cells are filled within the searchable stage range, though one lies outside it. The lowest held-out error shown for each filled order is 0.01291 for order 2, 0.02286 for order 3, and 0.02787 for order 4. These are the best entries in different cells, so they do not establish an overall ranking of the methods.
Search error often understates held-out error. In the order 4, six-stage, bucket 3 cell, search error is 0.002124 while held-out error is 0.02787. The order 3, five-stage, bucket 2 cell shows a wider gap: 0.003999 versus 0.06314. This suggests that optimizing against the search cases can select coefficients whose accuracy does not carry over to held-out cases. Stage count alone does not explain the results; the occupied buckets and reported cycle costs also vary.
The literature digest gives plausible sources for this sensitivity. Floor rounding of fixed-point products can add directional error at each stage, while quantized coefficients and step sizes can accumulate further error. Wider intermediate values and more accurate constants improved the cited ODE experiments, indicating that the implementation can matter as much as the tableau. The recent cross-cost hypotheses were recorded as refuted, but their predicates combine several comparisons; those outcomes do not isolate which comparison failed or establish a general rule across cost models.
The archive does not yet show whether the empty cells are infeasible or simply unfilled, and the stall counter of 72 leaves the rate of further coverage uncertain. Useful next checks are to examine the individual comparisons behind the recent hypothesis outcomes, repeat the large search-to-held-out gaps across more test cases, and vary rounding, coefficient precision, and intermediate width. Measured cycle costs on the intended processor would also help interpret the cost-model comparisons, since the digest notes that multiplication timing can differ substantially across Cortex-M0+ implementations.
47 older entries are not shown here. The whole log is in rk-work/interpretation/interpretations.jsonl.
Digests of published work the model looked up on the web during the run, newest at the top. They feed the directive and hypothesis prompts. Nothing here is a measurement from this project, and no measured page depends on it.
Cycle numbers restart at each epoch, so every entry names the epoch its cycle counts in.
A 2024 study of numerical optimal control reports that fixed-grid integration can produce large errors while adaptive grids can exceed the available sampling time. This supports evaluating solver error alongside a fixed execution budget, rather than using formal order alone to judge suitability for real-time control.
sources:
Arm documents two Cortex-M0+ multiplier implementations: MULS takes either 1 or 32 cycles and returns the low 32 bits. With zero wait states, shifts, adds, and subtracts each take 1 cycle; ordinary loads typically take 2. These figures make the chosen multiplier implementation central to a fixed cycle budget, while memory accesses and register use also affect the cost of a Runge, Kutta stage.
sources:
Implicit families have different costs even at the same stage count. A diagonally implicit Runge, Kutta method solves its stages in sequence; an SDIRK method uses the same diagonal coefficient, allowing a linear-system factorization to be reused when the relevant Jacobian is held fixed. Radau IIA is fully implicit, so its stages are coupled. These structures make stage count alone an incomplete measure of per-step work.
sources:
Runge, Kutta order conditions constrain tableau coefficients but can leave free parameters. Ralston used that freedom to derive second-, third-, and fourth-order methods with minimum truncation-error bounds of a specified type. For two-stage, second-order methods, the choice of stage location distinguishes methods such as midpoint, Heun, and Ralston’s method; the order label alone does not determine the leading error.
Numerical searches extend this approach. RK-Opt searches coefficients for a prescribed order and stage count and can minimize a principal error norm while imposing stability or storage constraints. Constraint-programming work likewise solves order conditions and optimizes remaining freedom. These results support comparing tableaus within each order and stage count. They do not establish which tableau has the lowest held-out error after Q15 quantization, floor rounding, and a fixed cycle budget; that comparison requires evaluation under those conditions.
sources:
Fixed-point ODE experiments show that arithmetic error can materially change a solver’s result even when its Runge, Kutta method and step size stay fixed. In tests of RK2 and RK3 neuron solvers, rounding down multiplication results produced substantial accumulated timing errors; rounding constants more accurately and retaining wider intermediate values also improved results. The authors define right-shift truncation of signed two’s-complement values as rounding down, including for negative values.
sources:
Bogacki, Shampine 3(2) and Dormand, Prince 5(4) use shared stages to produce solutions of two orders. Their difference supplies a local error estimate. SciPy’s implementations control error using the lower-order estimate and advance with the higher-order result. These are useful reference methods for assessing whether the accuracy gained from extra stages justifies their cost under a fixed cycle budget.
PI step-size controllers use the current and previous scaled error estimates to choose the next step size. Published implementations bound the change in step size and retry rejected steps. Those retries make execution time variable, so a fixed-budget comparison must account for the controller and its worst-case work, as well as the RK stages.
Fixed-point ODE research reports that reducing the step size can eventually increase error when quantization dominates truncation error. For Q15 with floor rounding, a small difference between embedded results can also disappear on the integer grid; this is an inference from the embedded error calculation and fixed-point quantization, not a measured Cortex-M0+ result. Arm’s Q15 DSP operations document saturation, which must be distinguished from the specified floor-rounding arithmetic. The consulted sources do not establish a particular pair or controller as best for this target.
sources:
SciPy’s default solve_ivp method is the adaptive RK45 5(4) pair, with default tolerances rtol=1e-3 and atol=1e-6. Its explicit Runge, Kutta implementation stores stage derivatives in an array of shape (stages+1, state size). It accepts a step when the RMS norm of the estimated error, scaled componentwise by atol + rtol × max(|old state|, |new state|), is below 1. Its step-size controller uses a 0.9 safety factor, limits reductions to a factor of 0.2 and growth to a factor of 10, and can retry rejected steps.
sources:
For a fixed integration interval, reducing the step size lowers an order-p method’s global truncation error, but increases the number of rounding events. Kehlet and Logg derive an accumulated roundoff bound proportional to the inverse step size when errors do not cancel, and an inverse-square-root scaling under a model of unbiased random errors. The latter model cannot be assumed for floor-rounded Q15 arithmetic: each downward quantization has a directional error, and correlations across steps may matter. Fixed-point ODE experiments also show that rounding choice can materially change accuracy.
sources:
A 2025 motor-control study compared Euler, backward Euler, Heun, and RK4 for an embedded predictive controller. In that application, two-stage Heun closely matched the RK4 reference at a 1 ms step and fit the available computation budget; Euler needed much smaller steps, while RK4 was too costly. This supports testing error per cycle rather than assuming that either fewer stages or higher formal order will win. The study used a dSPACE platform, so its timing results do not establish Cortex-M0+ cycle counts.
sources:
An explicit Runge, Kutta method has a bounded stability region. For a stiff, decaying mode, its step size must keep the scaled mode inside that region even when accuracy would permit a larger step. Studies of optimized stability polynomials show that extra stages can extend the stable step size, especially along the negative real axis; the benefit depends on the problem’s spectrum and the method’s order. [Ketcheson and Ahmadia](https://projecteuclid.org/journals/communications-in-applied-mathematics-and-computational-science/volume-7/issue-2/Optimal-stability-polynomials-for-numerical-integration-of-initial-value-problems/10.2140/camcos.2012.7.247.pdf) [Doehring and colleagues](https://link.springer.com/article/10.1007/s10915-024-02478-5)
sources:
132 older digests are not shown here. The whole log is in rk-work/literature/digests.jsonl.