Blog
Field notes
On trustworthy alarms, the peeking problem, and catching regressions before your customers do.
- Benchmark7 min read
Benchmarking drift detectors, honestly
We replayed every detector we ship against labeled synthetic fleets: where CUSUM genuinely wins, where it cries wolf 756 times, and what the budgeted tier trades for its guarantee.
Read - Calibration6 min read
CUSUM's false-alarm promise, measured
A CUSUM tuned for one false alarm per 2,000 points delivered one per 189 on textbook data and one per 13 on realistic texture. The ~154× gap decomposes cleanly — none of it is exotic.
Read - The peeking problem6 min read
Your monitoring dashboard is lying to you
Every fixed-threshold dashboard false-alarms by design — watch a metric long enough and it will trip. Here's why, and the fix that lets you check as often as you want.
Read