Quality assurance
A reporting tool that changes behaviour between releases is worse than one that never improves. GigHz Precision AI is developed under a test regime designed so that a fix, once made, cannot quietly come undone.
The build gate
A release cannot be packaged unless every deterministic check passes against the source tree. The checks are not unit tests of functions in isolation; they are invariants about the product's behaviour, extracted from the real code and run against the real configuration:
- Routing: each study phrase in the corpus resolves to the template it should, contrast is bucketed correctly, no routing row points at a template that does not exist, and a one-letter misrecognition does not cost the template.
- Placement engine: a dictated finding can never be silently destroyed; a second finding on the same line accumulates; a 5 mm nodule and a 5 cm mass are distinct; hand-edited lines are protected; named vessels are rescued from the parking section through every door.
- Impression: emergencies lead; chronic findings are demoted; incidental findings are suppressed at every door after the backstop; the impression never repeats itself; dictated impressions are never rewritten.
- Clinical scoring: one implementation per system; feature regexes cannot match their own negation; the grids agree with the published tables on worked cases; the live panel and the generated report derive the same category.
- Boundaries: the feedback payload cannot carry report text; no script loads from a public CDN; administrative features are not rendered to clinical users; secrets never ship to the browser.
- Deployment health: the live site answers, reports its version, has its model configured, serves its vendored libraries, and matches the packaged build.
At the time of writing the suite runs several hundred such checks on every build and every night.
Testing the tests
A check that cannot fail is worse than no check, because it reads as coverage. For every family of check, a mutation test plants the defect the check claims to detect and confirms the check goes red. A check that stays green on its planted defect is reported as blind and the build fails.
The nightly corpus replay
A corpus of dictations, most of them harvested from real field failures, is replayed every week against the live site through the same placement and composition paths a radiologist uses, and the output is scored: did each finding land on its line, were measurements kept verbatim, does the impression mention what it must and exclude what it must not, is the numbering sound. On the other nights the free deterministic checks run against the live site, together with a replay of only the cases tied to open issues, so a regression is still seen the next morning.
The issue registry
Every failure is fingerprinted so that the same defect, seen on a different patient weeks later, collapses to one identity. The registry tracks each identity from open, to fixed, to verified after consecutive clean runs, and flags any identity that fails again after being fixed as a regression, which is the loudest thing in the report. Several distinct identities sharing a subsystem are flagged as a theme: a sign that the mechanism, not the symptom, needs changing.
From field report to permanent check
When a radiologist reports a problem, the dictation that caused it becomes a corpus case with explicit expectations, the fix is made, the check that would have caught it is added to the suite, and the mutation test proves the new check can fail. The report in this documentation that explained a behaviour was, in most cases, written after exactly that cycle.
What is not tested automatically
Clinical correctness of a composed impression on a novel case is judged by the reviewing radiologist, not by the suite; the suite guarantees the rules were applied, not that the rules are medically complete. That is why the Terms require review, and why feedback from working radiologists is the input the suite grows from.
This page is generated from documentation that ships with the application (docs v1.2.2). Version numbers, prices and trial lengths on this page are read from the running system, not typed by hand.