Accuracy & testing — how Leo, Nox's safety layer, is measured

How Leo, Nox's safety layer, is tested, what the measured numbers actually mean, and how the medical review process works. Every figure on this page is generated by the test harness and committed with the code — none of it is hand-typed marketing copy.

Scope: These numbers describe ONLY Leo's deterministic red-flag detector, measured on Nox's own labeled test set. The test set is authored and maintained by the Nox team alongside the detector — it is not an independent or clinical benchmark. These figures are not a clinical certification, a diagnostic-accuracy measure, a sensitivity/specificity study, or a guarantee for real-world symptoms.

Headline metrics

  • Overall recall: 100.0% (145 of 145 should-fire cases produced a safety note)
  • False-positive rate: 0.0% (0 of 53 benign cases triggered a note)
  • Test set: 198 labeled cases · 102 rules across 72 categories
  • Missed cases: 0 · Misfires: 0 (full lists in the raw report)

Per-category recall

CategoryTest casesCaughtRecall
Headache66100.0%
Eye pain55100.0%
Chest pain55100.0%
Shortness of breath55100.0%
Abdominal pain99100.0%
Fever33100.0%
Allergic reaction55100.0%
Mental-health crisis55100.0%
Pregnancy55100.0%
Child symptoms55100.0%
Stroke (FAST)66100.0%
Seizure44100.0%
Low blood oxygen (SpO₂)44100.0%
Severe dehydration44100.0%
Fever with confusion33100.0%
Overdose / poisoning44100.0%
Severe bleeding11100.0%
Sepsis warning signs11100.0%
Diabetic emergency11100.0%
Testicular torsion11100.0%
Choking11100.0%
Severe burn11100.0%
Heat stroke11100.0%
Aortic dissection22100.0%
Abdominal aortic aneurysm22100.0%
Blood clot in the leg (DVT)22100.0%
Fainting / syncope22100.0%
Dangerous heart rhythm11100.0%
Hypertensive crisis22100.0%
Blocked artery in a limb11100.0%
Head injury22100.0%
Spinal injury11100.0%
Cauda equina syndrome11100.0%
Broken bone / open fracture22100.0%
Compartment syndrome11100.0%
Crush injury11100.0%
Stab / gunshot / impalement11100.0%
Drowning / near-drowning11100.0%
Hypothermia11100.0%
Frostbite11100.0%
Severe altitude sickness11100.0%
Smoke inhalation11100.0%
Snake bite11100.0%
Animal / human bite22100.0%
Airway swelling (stridor)11100.0%
Severe asthma attack11100.0%
Coughing up blood11100.0%
Unable to urinate11100.0%
Kidney stone complications22100.0%
Bowel obstruction11100.0%
Strangulated hernia11100.0%
Ectopic pregnancy11100.0%
Ovarian torsion11100.0%
Postpartum emergency22100.0%
Priapism11100.0%
Sudden hearing loss11100.0%
Thyroid storm11100.0%
Adrenal crisis11100.0%
Sickle cell crisis11100.0%
Fever while immunocompromised11100.0%
Severe alcohol withdrawal11100.0%
Serotonin syndrome11100.0%
Severe drug reaction (SJS)11100.0%
Acute psychosis11100.0%
Eating-disorder complications11100.0%
Spreading dental infection22100.0%
Uncontrolled nosebleed11100.0%
Severe menstrual bleeding11100.0%
Dialysis emergency11100.0%
Newborn jaundice11100.0%
Sexual assault11100.0%
Domestic violence11100.0%

What is being tested

Before any AI model answers, Leo's deterministic detector — plain pattern-matching code, not AI — screens each message for acute red-flag presentations like stroke signs, chest pain, severe breathing difficulty, or a mental-health crisis. Because it is deterministic, it can be tested exactly: the same input always produces the same result.

How it is tested

The team maintains a labeled test set of emergency, urgent, and benign messages. Every case runs through the same detection function the live app uses. Recall is the share of should-fire cases that produced a safety note; the false-positive rate is the share of benign cases that triggered one anyway.

What happens when a rule matches

A match is graded at one of two severity levels. An emergency match (for example stroke or heart-attack signs) puts a prominent banner at the top of Nox's reply that tells you to call your local emergency number — the exact number is chosen from your saved region, then your browser's language, falling back to the universal 911 / 999 / 112 guidance. An urgent match adds a strong 'get seen promptly' note. Either way the banner is added before the AI's answer, and the AI is instructed to lead with seeking care rather than self-treatment.

Why you can trust the numbers on this page

The metrics report is machine-generated by a committed command, and an automated test fails the build if the rules or the test set change without regenerating it. The numbers rendered here are read from that committed report at build time — the page physically cannot show numbers the test suite didn't produce. Missed cases and misfires are published verbatim in the raw report at https://nox.aurenaring.com/api/transparency.

The medical review mechanism

How safety changes are reviewed

Every change to Leo's red-flag rules requires re-running the evaluation harness and committing the regenerated metrics report — an automated test enforces this, so the published numbers can never silently drift from the live detector. The rule categories, severity thresholds, and recommendations are maintained against a written governance record with named data sources and licensing.

Clinician review status: pending (honestly)

Leo's red-flag rules and safety prompts have not yet completed an independent review by a licensed clinician, and Nox claims no clinician endorsement until one is on record. When a real review is completed, the reviewer's name, credentials, scope, and date will be published on the Trust & Transparency page — and re-reviewed whenever the rules change materially, and at least once a year.

Conservative by design

The detector is intentionally tuned to prefer an unnecessary safety note over a missed emergency. It reads the latest message only, covers a fixed list of high-risk categories, and does not diagnose — emergencies outside its categories may not trigger a note, which is exactly why Nox never describes it as catching everything.

A second AI backstop, measured separately

Beyond the deterministic detector these numbers describe, Nox runs a second safety check only when the pattern rules find nothing: a lightweight AI classifier that catches dangerous descriptions worded in slang, another language, an older disease name, or indirectly. Because it uses AI and is not deterministic, it is evaluated on its own separate test set and is deliberately NOT included in the recall and false-positive numbers on this page — those describe only the deterministic detector. The backstop can only add a safety note, never remove one.