Evidence
Engineering validation
This page is the verification ledger. I record what the software already does, and what my automated tests actually checked. A passing test is not diagnostic accuracy.
Evidence status
I added engineering tests as the software grew. The public log below is one snapshot, dated 2026-10-19, commit 3354f22: seventeen tests across geometry, quality, fusion, and screening labels. I am not silently redating that snapshot.
| Capability | Status | Evidence |
|---|---|---|
| Deterministic geometry | Engineering tests exist | Known-coordinate unit tests: finite unsigned Lovibond, DPD/IPD arithmetic, injected Myers bands, missing hyponychial returns none. Geometry excerpt. |
| Image-quality behavior | Engineering tests exist | Blur, exposure, framing, finger-too-small fixtures fire documented reject codes. A pass returns empty reason codes. Quality excerpt. |
| Fusion / one-image rule | Engineering tests exist | Protocol rejects a one-image patient conclusion. One labeled finger cannot be Marked at patient level. |
| Screening-label titles | Engineering tests exist | Visit titles and same-day ordinals. This suite is not the stub-release guard. Stub morphology withholding is shown in the release excerpt. |
| Landmark detection | Experimental; stubbed by default | Placeholder points can still produce numbers. Not validated anatomy detection. |
| Abstention behavior | Implemented; calibration not established | Unable to assess below a 0.55 research cutoff. Uncertainty excerpt. Code behavior is not a validated clinical rule. |
| Device and lighting robustness | Not established | Future study. |
| Skin-tone performance | Not established | I do not claim demographic fairness from these tests. |
| Clinical diagnostic accuracy | Not established | Requires a clinical study. Software tests are not that study. |
How to interpret these results
The 17-test log is a record of what those four suites did on 2026-10-19, after tests had been written through the project. It is engineering verification of math and policy on fixtures. It does not prove sensitivity, specificity, or fairness, and it is not a substitute for a later, separately dated test run.
- test_geometry proves finite unsigned angles, ratio arithmetic, and label mapping of injected numbers, including 181° that the angle function cannot compute.
- test_quality_parity proves documented reject and pass codes on fixtures, not that a pass is a valid measurement.
- test_fusion_policy proves the protocol forbids a one-image patient conclusion.
- test_screening_labels proves visit title formatting. It does not prove research-only release safety.
Summary: test summary. Full log:
LungLens engineering tests (isolated pytest) Date: 2026-10-19 Suite: test_geometry, test_quality_parity, test_fusion_policy, test_screening_labels Result: 17 passed Commit tip audited: 3354f22 These tests check geometry math, quality reject/pass codes, fusion policy, and screening labels. They do not establish clinical accuracy, fairness across pigmentation, or validated AI localization.
What is not established
Passing engineering tests does not prove clinical performance. Classification bands in code are provisional. Landmark AI is not the default path. Cameras and lighting still matter after a quality pass.