Command line

disp2vlm-validate {gate,run,vlm,summary,demo}

Command	Does
gate [--format text\|json]	Print the release gate and the benchmark frames.
run --config run.yaml [--no-cache] [--verbose]	Validate one candidate on one frame.
vlm --config vlm.yaml [--verbose]	Validate a VLM run.
summary RUN_DIR... [--out release_summary.html] [--title STR]	One page across runs.
demo OUT_DIR [--seed INT]	Write the synthetic demo frame and its configs.

run prints the verdict and the reasons, and exits normally either way: read the verdict from the output or from metrics.json.

gate

$ disp2vlm-validate gate
disp2vlm-validation: CalVal release gate
  sill_must_drop                 True
  sill_rise_max_frac             0.05
  sill_min_pair_km               50.0
  velocity_rmse_rise_max_mm_yr   0.2
  velocity_bias_rise_max_mm_yr   0.2
  bias_statistic                 median
  target_abs_bias_mm_yr          1.0
  bootstrap_ci                   0.95
  velocity_excluded_classes      ('defo_area', 'near_event')
  independent_gnss_only          True
benchmark frames:
  ...

Python

Everything the CLI does is a function call:

from disp2vlm_validation.pipeline import load_run_config, run_validation

metrics = run_validation(load_run_config("run.yaml"))
print(metrics["verdict"], metrics["verdict_reasons"])

See the API reference.