Every claim, and the command that makes it again

This project puts correctness that can be demonstrated ahead of speed, and speed that can be quantified ahead of elegance. What follows is that order spent: four pages, each of which states what is claimed, what the claim was measured against, and how to run the measurement.

Nothing on these pages is a figure this repository cannot produce again. Where a number is quoted, the command that produces it is printed beneath it, and where a measurement varies from session to session the spread is printed beside the median rather than left out. That last rule is the one that costs something, and the performance report says what it costs.

• The demonstration

  Two disc galaxies on a bound, grazing encounter: what the configuration says, which of its settings the result is sensitive to, and what the run costs to compute and to draw.

• Validation

  The analytic comparisons, the convergence study that recovers each integrator’s order from a measurement, and the counterexample that decides which integrator is the default.

• Performance

  Every speed as a fraction of a ceiling measured on the same machine in the same session, with the spread and the drift beside each median, and an account of which figures reproduce.

• The solvers

  Four implementations of one interface, what each is fastest at, and the decisions that let a GPU backend be added without becoming a second copy of the physics.

The instrument plays three of these runs: Kepler, the cluster and the collision.
The reference documentation, generated from the headers, is at the documentation site.