Skip to content

Design constraints

These are decisions, not preferences, and they are why the dependency list is short. See CONSTRAINTS.md.

  • Pure Python + NumPy at runtime. No compiled extension to build, no BLAS requirement to satisfy, no solver to license.
  • Seeded and reproducible. Two runs of the benchmark must be identical apart from wall-clock timings, and CI fails if they are not.
  • Coverage is a ratchet, not a report. Every module has a floor; a drop fails the build.
  • No vendored datasets. Reproducibility comes from seeded configs and committed figures, not from committing megabytes of capture.
  • Claims are typed FACT, MEASUREMENT, INTERPRETATION or HYPOTHESIS, and a synthetic result is never presented as a field measurement.