A vape detector claims high accuracy: separate missed events from false alerts

Define what one tested item means

A detector might classify a short time window, a hand gesture, or an entire person as having a specified behavior. The same word accuracy can accompany different units. Locate the unit, the positive label, and how the reference was determined. A human annotation or device signal has its own limits; do not automatically call it an error-free account of real behavior.

Google’s first-party machine-learning lesson defines accuracy, precision, and recall using a binary confusion table. We use these mathematical definitions for reading, without adopting a model or reporting actual vaping performance. The useful question is which denominator the claimed percentage uses and whether it matches the practical detection task someone says the system performs.

Google for Developers: Classification: Accuracy, recall, precision, and related metrics

Build a fictional four-box table

Imagine one hundred invented test windows. Ten have the reference-positive label. A toy system flags eight of those ten and misses two. Of the ninety reference-negative windows, it falsely flags nine and correctly leaves eighty-one unflagged. The four boxes are eight correct flags, two missed reference-positive windows, nine false flags, and eighty-one correct non-flags.

Its accuracy is eighty-nine of one hundred. Its recall is eight of ten reference-positive windows. Its precision is eight of seventeen total flags. The percentages answer different questions, despite describing the same table. These numbers are fictional arithmetic, not a vaping study, clinical detector, or product benchmark. Keep both the unit and reference definition beside the figures rather than assigning them to people.

Connect the error to the proposed use

If a tool would populate a log automatically, false flags could create entries that did not match the reference and misses could leave events out. Ask how the actual system handles review and corrections, without assuming a product offers either. If its stated task is only exploratory research classification, do not promote it to verified personal monitoring because the score sounds impressive.

A different mixture of positive and negative windows can change some reported metrics even with similar class-specific behavior. Look for the tested setting and whether the report supplies counts. Avoid comparing unlike accuracy percentages from two demonstrations. Data separation and external testing remain necessary questions, but this exercise specifically explains the types of error after you locate the reported test table.

Write a bounded technology claim

Your note can say the system classified this kind of item against this reference, with these reported missed and false classifications. If only overall accuracy is available, say the error breakdown was not verified. A lack of detail does not prove the technology is fraudulent; it limits what you can repeat about its suitability for the claimed task.

Do not use a score to diagnose dependence, infer chemical exposure, or assume that automatic recording improves quitting. Measurement performance and intervention benefit need separate evidence. Nor should you upload private histories to a demonstration to settle a marketing question. A complete reading output preserves the reference, unit, denominators, and unanswered error question rather than turning one percentage into a promise.