Walkthrough
Watch one model lose your trust — and earn it back
This runs against the real engine. It trains a baseline, then feeds it four batches: one healthy,
one slightly off, one badly drifted, and one that comes back. The input is a fixed, verified dataset —
the same one every time, so the four outcomes are the same for everyone.
Baseline
300 rows of healthy data. Kirelta learns this shape and picks a monitor for it.
1 · A normal day
Fresh rows from the same distribution. Nothing has changed — the verdict should say so.
2 · Something shifts
The inputs move a little. Not enough to fire the alarm, but enough that more rows are
falling outside the trained range. This is the state most tools miss entirely.
3 · It breaks
A hard shift. The sequential alarm fires, and the engine says stop.
4 · It comes back
Inputs return to normal. The verdict recovers — and because Kirelta remembers the
recent alarm, it also tells you this reading is a recovery, not just a calm day.