01 / FREEZE
Generate one public mask
Odd dimensions, a secondary component, bridge, hole, and deterministic speckle exercise the cleanup operation without using a scan.
DICOM segmentation cleanup primitive
The parity lab, before the first receipt
Public input mask
129 × 95 central plane
input foreground
CPU reference foreground
Metal candidate foreground
non-zero residual
02 / EXECUTE
Run both routes
The CPU reference and requested Metal route apply the same 3 × 3 × 3 clamp-to-edge opening: two erosions and two dilations.
03 / COMPARE
Inspect every voxel
Binary outputs use exact equality. The residual cannot be reduced to a visual impression or a tolerance chosen after the fact.
04 / RECORD
Save the evidence
The app exports a human-readable lab and JSON receipt with fixture identity, output arrays, backend state, and trial timings.
What a completed receipt must establish
One narrow claim, checked visibly.
The report is intentionally more demanding than “Metal was requested.” A completed candidate route must state whether every morphology pass actually ran on Metal before it can receive a timing or parity verdict.
| Evidence | Recorded in the app | Purpose |
|---|---|---|
| Fixture identity | Deterministic mask dimensions and checksum | Lets an independent reviewer recognise the input without exposing a DICOM study. |
| Backend result | Fully completed Metal, CPU fallback, or mixed execution | Separates an attempted GPU route from a completed GPU result. |
| Parity contract | Exact binary-mask equality | Makes every changed voxel visible in the residual view. |
| Timing scope | Preparation, dispatch, synchronization, and readback | Keeps UI rendering and report generation outside the measurement. |