Validation
CubeDynamics treats examples as scientific outputs. The publication gate runs five independent modules covering checksum-controlled real inputs, grammar, cube rendering, notebook execution, and deliberate failure controls.
Published PRISM baseline: PASS — 5 of 5 modules. New source candidates have separate production review gates; successful fixture checks do not make them production-certified.
| Module | What it validates | Evidence |
|---|---|---|
| Real data | source, checksum, dimensions, dates, coordinates, units, finite values, physical relationships, and 60 source archive records | Data report |
| Grammar | direct xarray calculations agree with equivalent pipe(cube) \| verb() expressions |
Methods |
| Cube / HTML | six unique faces, complete uncropped textures, declared direction on every axis, and exact RGBA pixels | Cube report |
| Vignettes | core and source lessons name an explicitly supported real fixture, verify its provenance/checksums, execute offline, emit plots, and keep generated examples out of public learning routes | Methods |
| Contrast | known reversals, transpositions, duplicate faces, and cropping behavior are rejected | Contrast report |
The default suite is offline: CI validates the checked-in fixtures rather than silently substituting generated data when a service is unavailable. Rebuilding the fixture is a separate, explicit, checksum-verified acquisition step.
Run it
python -m pip install -e ".[dev]"
python scripts/run_validation.py --run-vignettes
The command writes one result.json and one PNG per module, a suite manifest,
and a collated PDF under artifacts/validation/. Any failed acceptance check or
notebook exits nonzero.
This design follows the evidence-oriented pattern used by the Fire VASE validation suite: modular checks, visual evidence, machine-readable artifacts, and expected- failure controls.