Triton ULTRA Vape Detection Accuracy Report

Key findings

Triton ULTRA correctly classified 8,315 of 8,400 logged test events, producing an observed aggregate accuracy of 98.988%. Triton reports this conservatively as 98.9%, rounding down to one decimal place. This figure describes the study corpus; performance can vary with the product, distance, airflow, room geometry, and installation.

The testing program evaluated genuine vape events alongside non-vape aerosols, vapors, and fumes that can challenge sensor classification. It included 23 non-vape sources and 7 vape-event sources representing different products and test conditions.

How the accuracy figure was calculated

Aggregate accuracy incorporates both sides of the classification task: detecting genuine vape events and correctly rejecting the non-vape sources included in the test set.

Study design and test coverage

Testing took place at intervals from May to July 2026 in actual school restrooms, a commercial environment, and a residential environment. Triton ULTRA used default sensitivity settings and the firmware released on May 25, 2026.

Each event has a unique Testing ID and a Source ID linking it to the product and source category in the workbook’s Summary sheet. The log records the expected result, sensor output, binary assessment, test distance, and amount used.

Non-vape sources included aerosol body sprays, antiperspirants, hairsprays, dry shampoo, disinfectants, fabric and air refreshers, surface cleaners, thermal styling, water vapor, and soldering fumes. Vape sources included disposable and refillable nicotine devices, high-VG and high-PG e-liquids, zero-nicotine e-liquid, a Delta-8 cartridge device, and exhalation into clothing.

Test methodology

Event-level scoring

Each trial used a two-minute observation window, beginning with the first puff, spray, or other introduction of the source. A vape alert within that window counted as a true positive for a genuine vape trial and a false positive for a non-vape trial. No alert counted as a false negative for a vape trial and a true negative for a non-vape trial.

Ventilation between trials

Rooms were ventilated between trials until residual material had dissipated and a sufficiently clean-air baseline returned: PM2.5 below 10 µg/m³ and TVOC below 400 ppb. This procedure was intended to minimize carryover between events.

Distances and quantities

Source-to-sensor distances ranged from 0 to 6 feet, with a median of 3 feet. Amounts ranged from 1 to 60 units, depending on the source. The log includes 4,200 spray-based events, 2,900 puff-based vape events, 800 events measured in seconds, and 500 events measured in squirts.

Complete results

The tables below reproduce the report’s results. On smaller screens, scroll within a table to see every column.

Overall classification performance

Results by source category

Results by individual product / source

Results by test distance

Results by measurement type

Error analysis

The dataset contains 46 false positives and 39 false negatives, totaling 85 classification errors.

False positives were recorded for aerosol antiperspirants (20), combustion/soldering fume (9), aerosol body sprays (9), dry shampoo (4), thermal styling (2), and hairspray (2). The other non-vape categories recorded no false positives in this study.

False negatives totaled 20 nicotine-vape events, 9 THC/Delta-8 events, 7 filtered or exhalation-into-clothing events, and 3 nicotine-free vape events.

What the results mean in practice

What does 98.9% accuracy mean?

The 98.9% figure describes performance in this dataset. Results can vary with airflow, room layout, installation, source characteristics, and distance. Individual products and test conditions can perform above or below the aggregate result.

The study evaluated whether events were classified as vape or non-vape. The Delta-8 results apply to the tested source; they do not establish performance across all THC products or show that the sensor distinguishes THC from nicotine. The dataset does not measure notification speed, verify the full recommended installation coverage area, or establish competitor performance.

How should a positive alert be interpreted?

The observed false-positive rate was 0.836%: 46 of 5,500 non-vape trials produced an alert. This is the proportion of non-vape trials incorrectly classified as vaping, not the proportion of all alerts that were false.

In this corpus, 2,861 of 2,907 positive classifications were true positives, giving 98.418% precision. The remaining 46 positive classifications, or 1.582%, were false positives. Genuine vape events represented 34.52% of the corpus. Field precision depends on how often vaping actually occurs and on the background sources and operating conditions; the study’s precision should not be assumed to apply unchanged in deployed schools.

Response time, aerosol travel, and an empty room on arrival

Alert timing includes both the time aerosol takes to reach the intake and the device’s processing time once enough aerosol is present. Distance, plume size, exhalation direction, and airflow affect the first component. A small puff directed downward or into clothing may take longer to reach the device or may never reach a detectable concentration.

A person can leave before aerosol reaches the sensor or before staff arrive. An empty room on arrival does not by itself establish a false positive. Response protocols, sensor placement, and supported camera or video-management integrations can help teams review the event.

Coverage and deployment guidance

Triton recommends approximately 160 square feet per ULTRA sensor, with ceilings up to 8.5 feet. Internal testing found that multiple-puff events produced sufficient aerosol for detection at each distance tested within that envelope. Larger spaces, higher ceilings, strong exhaust, or small puffs filtered through clothing can lengthen response time or contribute to a missed event.

The 160-square-foot recommendation is installation guidance, not a guarantee of uniform 98.9% accuracy. The accuracy dataset covers 0–6 feet and does not establish performance beyond that range. Spaces that exceed recommended coverage should use additional sensors. Triton can review placement, unit count, sensitivity, airflow, and response protocols for individual buildings.

Study limitations

• Testing distances were limited to 0–6 feet. No measured accuracy rate beyond 6 feet can be inferred from this dataset.
• The non-vape interference set is broad but does not cover every substance or activity that may occur in a deployed environment.
• The THC category contains one source: a Delta-8 cartridge device (P29), detected in 491 of 500 trials. This binary vape/non-vape study does not establish performance across all THC products or differentiation of THC from nicotine.
• Room layout, HVAC, airflow, sensor placement, background aerosol, source characteristics, and installation conditions can affect performance.
• The event log does not contain enough information to independently assess alert latency, airflow, temperature, humidity, or placement relative to HVAC equipment.
• This study evaluates Triton ULTRA. It does not establish competitor accuracy or superiority over other sensors.

Report, calculations, and testing logs

The underlying workbook is the primary evidence record. Unique Testing IDs and Source IDs let reviewers trace the aggregate result to source-level records. The workbook contains 23 non-vape source IDs (P01–P23) and 7 vape source IDs (P24–P30).

Appendix A · Calculation

Total events: 8,400 · True positives: 2,861 · True negatives: 5,454 · False positives: 46 · False negatives: 39 · Correct classifications: 8,315.
Observed aggregate accuracy: 98.988095%.
Published accuracy: 98.9%, reported conservatively by rounding down to one decimal place.

Appendix B · Testing logs / summary

The report identifies an underlying event-level testing log. The source-level table above reproduces all 30 products and sources and their associated results.