Skip to content

Validation status and limitations

Software verification, numerical stability, seasonal equilibrium and independent observational agreement are separate requirements.

Software verification

At revision 198433d, local package/CLI/calibration validation produced 411 passed, three skipped, no failures, including all 11 slow tests. The UI build passed. This records that revision rather than promising results for later code.

rtk proxy .venv/bin/python -m pytest -c package/pyproject.toml \
  package/tests cli/tests apps/mars-calibration/tests -m 'not slow' -q -rs
rtk proxy .venv/bin/python -m pytest -c package/pyproject.toml \
  package/tests cli/tests -m slow -q -rs

Asset-dependent tests can skip when data are absent. The explicit configuration registers the slow marker across suites. Slow CLI tests include a four-year intervention with GCM diagnostic snapshots and can take several minutes on CPU. Tests cover dry-dycore benchmarks, orbit, energy exchange, radiation closure, CO₂ inventory/escape, positivity, restart, gradients and experiment integrity.

Climate equilibrium

The recorded corrected baseline completed two additional Mars years with finite fields and maximum combined atmosphere/frost CO₂ relative drift about 1.81e-9. Phase-matched pressure, frost and surface-temperature changes passed 5 Pa, 5 Pa and 1 K criteria. Deep-soil changes of 0.442–0.634 K exceeded 0.25 K, so the full equilibrium gate failed. These are recorded results in the progress report, not new documentation-build runs.

MarsMapFields.is_transient uses duration below 668 sols as a heuristic. Passing that duration is not an equilibrium diagnostic. Longer runs still need matched seasonal repeatability, bounded budgets and appropriate averaging.

Reference diagnostics

Ames FV3 is an independent model reference, MCD is model-derived climatology and ARCO-MACDA is reanalysis. Their existing seasonal comparisons show useful spatial structure and remaining wind/frost deficiencies. Sampling, dust scenarios, wind heights and frost definitions differ. See reference comparisons. These diagnostics do not establish independent Viking/TES/MCS observational skill.

The saved ablation audit verifies output integrity and paired protocols but records a physics source-hash mismatch. Saved-run results belong to their original source hashes; they do not validate a later code revision merely because they are committed.

Remaining limitations

  • Prescribed dust opacity; no lifting, transport or interactive dust weather.
  • Dry convection and simplified surface/PBL closures; no water/cloud cycle.
  • Homogeneous regolith properties and zero-flux bottom boundary; slow soil spin-up.
  • No photochemistry or composition-dependent high-pressure validation.
  • Global CO₂ projection does not prove local transport accuracy or conservation of every energy/momentum term.
  • Bounded neural heating can add net energy; it is not a closed energy source.
  • GCM intervention snapshots are independent spin-ups from a torch global-mean trajectory, not a continuous multi-year 3-D intervention integration.
  • Higher resolution presets do not themselves establish convergence or skill.

The physics ledger retains detailed requirements and references. Date-specific design notes are historical context; this section describes the current implementation.

The torch global-mean model is deprecated. Its remaining intervention dependency is documented for compatibility; use GCM workflows for new spatial-model research.