Precision
0.988
of alarms raised, share correct
No per-domain configuration. No retraining. No oracle thresholds. These are deployed numbers — what an integrator gets after install, power-on, walk-away.
CWRU bearing dataset with 27 distinct fault pairs spanning inner race, ball, and outer race faults at three severity levels. The recommended operating point produces the deployed numbers below — without an oracle threshold and without per-fault tuning.
An industry-standard run-to-failure dataset spans more than 30 days of continuous vibration from four bearings, two of which failed and two of which survived. Flynn ran on all eight accelerometer channels simultaneously — zero configuration changes between them.
336 hours of ambient time-series covering temperature, environmental, and process-control signals — slow-changing data with strong day-night cycles. Flynn maintains a false-positive rate that fits the cadence of normal operator review.
10,000 instances across 12 features of electrical-grid stability data — a signal domain that shares nothing with rotating-machinery vibration. Same source, same flags, no per-domain configuration.
The dominant operational concern with anomaly detectors is the false-alarm rate. Across 120 simulated hours of synthetic vibration — 5 random seeds × 24 hours each — Flynn produced zero false positives, bounded above by 0.025 per hour at 95% confidence. On bearing-data normal-operation traces, zero false positives across 1.2 million samples.
Zero false positives on bearing data. Operationally low on diurnal. The numbers operators actually care about.
Note on completeness: an accelerated-life bearing-degradation dataset compresses years of wear into hours, producing a signal profile well outside Flynn's standard deployment envelope. F1 = 0.268 on this test indicates useful signal is still extracted even when the operational regime is structurally hostile to Flynn's architectural assumptions. Included for completeness; not a recommended deployment scenario.
Flynn's empirical claims are reproducible from committed benchmarks and committed source code. Artifacts and reproduction instructions are available to evaluation licensees.