Validation docket

Everything we have measured, everything we have not, kept in three separate sections so the two can never blur. Artifact names are the committed files behind each statement; ask system@apertureapi.com for any of them.

A. Synthetic benchmarks (public data or closed-loop simulation)

Public dataset comparison. We reconstruct a synthetic in-line hologram from a published dataset (Zenodo record 8059636, a USAF resolution target) and compare our result against the dataset's own reference reconstructions. This is a reconstruction-fidelity comparison on a synthetic fixture, not a measurement of a physical object. The run's own quality band is fail, and we publish it that way. We do not quote the agreement figure in sales material. Artifacts: utirnet_ncc_usaf_synthetic.json, utirnet_evidence_usaf_synthetic.json / .html.

Calibration closed loop. Software-defined calibration fits the camera's true pixel pitch from a photo of a known target. In committed closed-loop proofs on two simulated targets (a hex nut and a salt crystal), the optimizer recovered a deliberately mis-specified pixel pitch to within 0.0005% and 0.0007%. Wavelength and distance recover only to about 1–2%, because the two trade off against each other in the physics; we claim pitch recovery and nothing else. The hologram was generated by the same model the optimizer fits, so this proves the optimizer and the endpoint plumbing — not a physical rig. Artifacts: sdc_end_to_end_hex_nut.json, sdc_end_to_end_salt_crystal.json.

Autofocus diagnostics. On in-line holograms the previous default focus measure was biased; the cause was found and the default was changed on 2026-10-12. On the committed synthetic target, autofocus lands inside the search range, and a warning fires whenever focus lands on a search bound. Artifact: focus_diagnostics.json.

B. Physical-target results

None yet. No physical target has been measured against an independent reference on a rig we do not own. Every result we produce says so: accuracy and repeatability are not-yet-established and traceability is false.

Design partnerships exist to produce exactly this: your frame, your reference measurement and our reconstruction, all three committed. The first such result will be listed here, including if it disagrees.

C. Known unknowns and limits

• Accuracy: not established for any physical setup.
• Repeatability: not established.
• Run-to-run output reproducibility (in-line path): not established. On 2026-10-12 two runs of the same sample with the same settings on the same build returned the same provenance ID but different output arrays, quality metrics and evidence report IDs. The provenance ID proves the same input, settings and software were used; it does not prove the outputs match. Keep the result package you relied on.
• Twin-image suppression on in-line holograms: not established. The committed report shows a small effect and no conformity verdict. Artifact: twin_suppression_report_usaf.json.
• In-line quality bands: often fail. The bands are published ceilings; a fail describes the image and is reported, not hidden.
• Speed: 10–40 seconds per request on the current shared host. We publish no throughput rate.
• Large fields: requests over the memory ceiling are refused with HTTP 507 and an estimate, rather than failing partway.

Next step

To turn section B from "none" into a result on your rig, read Design partnership or write to system@apertureapi.com.