Model quality, in plain English
You’re currently viewing metrics for AURA classifier v0.1. On this snapshot it was evaluated on 10,095 labeled Windows files. About 76% were harmless and 24% were malware.
This page turns “91.6% accuracy” into things humans care about: how often malware is caught, how noisy alerts are, and how often something bad might slip by.
Click any original or quantized version to see its metrics. All numbers come from the encrypted training dataset snapshot used for that model.
- • Accuracy snapshot: 91.6%
- • Finds most malware: ~93% caught
- • Some false alarms on safe files
- • A small number of misses on malware (see slider below)
Traceix should be one signal in your pipeline, not the only one. For high-risk decisions, pair it with other detections and human review.
See it like a human, not a statistic
Drag the slider. We’ll estimate what happens if this model scans that many files, based on its benchmark dataset.
Out of 1,000 files:
- • ~760 are harmless
- • ~240 are malware
This version correctly flags about 220 of those malware files.
Driven by a malware recall of ~93.4%.
About 15 malware files are treated as safe.
These are false negatives — the small risk that a bad file slips by.
About 70 harmless files are flagged as suspicious.
These false positives are often acceptable when a human or a second system reviews alerts.
Numbers are approximate and based on the dataset used to evaluate each model version. Real-world performance will vary depending on what you scan.
On this model’s test set, about 9 out of 10 files get the correct “safe vs malware” verdict.
When the model says “malware,” this is how often it’s right on that dataset.
Out of all real malware files, this is the share the model actually catches.
False positive rate (safe flagged as malware) and false negative rate (malware treated as safe).