Run the arithmetic yourself

The Q15 integrator running in your browser, checked against the evaluator that scored the archive.

Everything on this page runs in your browser. The Q15 primitives, the explicit Runge-Kutta step and the error metric are a line-for-line port of rk_harness.fixedpoint and rk_harness.simulate, and every chart is computed from the coefficients when you click.

checking this page's arithmetic against the Python evaluator…

Rank the whole field at one cycle budget

11 methods, one budget of 65,536 cycles, one problem. Inside that budget a cheap method takes more, smaller steps than an expensive one, so this compares methods rather than step sizes. The thin mark on each bar is where that method lands under the other rounding mode.

Click or tap a bar to load that method into the trajectory below and see its numbers under the bar. Tapping anywhere else hides the numbers and keeps the method loaded. From the keyboard, Tab to the chart and walk it with the arrow keys; Enter loads a method and Escape hides the numbers. Orange is classical, blue was found by the search.

Watch one method integrate

Dashed is the float64 reference solution; solid is the int16 state, converted back to physical units.

Cost against error, recomputed live

The same 11 methods, placed by what they cost and what they achieve on the selected problem. The dashed line is the Pareto frontier: the methods no other method improves on for cost and error at once. It moves when you change the problem or the rounding mode.

Tap a dot to load its method and see its numbers beside it. The keyboard works here the same way it does on the ranking above.

How this page is checked

tools/demo_data.py runs every method, problem and rounding case (154 in all) through the pinned Python evaluator and stores the final int16 state and step count. The badge at the top recomputes all of them here and reports how many match to the last bit. The round-to-nearest rule was pinned the same way: it reproduces all 56 errors in tools/floor_round.json, the file the floor-bias finding is computed from, with a largest relative difference of 0.

Two limits. The cycle counts come from the analytic cost model, not silicon; the measured wall clock is where that model meets a real clock. And the findings site ranks methods by RMS error over four held-out problems, while this page shows one problem at a time so the mechanism stays visible.