Validation
Written to be forwarded. Four sections: how ground truth is made and recovered, how runs are pinned, how studies are cleared, and what may be claimed.
01 · Methodology
Validation volumes are synthesised with anatomy and tissue properties specified exactly, so the correct answer is known before the module runs. The module must then recover that answer within a tolerance fixed per measurement family.
Every measurement carries a citation to the primary literature that defines it — in the specification, and again in the delivered dataset.
In the validation report
02 · Provenance
Pipeline version, model weights and parameters are pinned per run. The same scan, re-run on the same pin, returns the same values.
The provenance record travels inside the export, so a reviewer holds the number and the method that produced it in the same file.
03 · De-identification
A study is cleared before analysis begins — nothing is measured until the de-identification check passes.
The de-identification record is included in every delivery, alongside the values it covers. See also Security and Regulatory status.
04 · Claims discipline
What may be claimed about a module is fixed at specification, before code exists. The build fails if the claim drifts — a stronger result cannot quietly become a stronger claim.
All analytics are descriptive: quantification and visualisation, not diagnosis. Research Use Only — CorpX is not approved by any regulatory authority for clinical diagnosis or treatment selection.