Try it
Would we have caught it? Replay your metric history and see.
Paste or upload a metric history — an eval score, an error rate, a latency. Your browser runs the same backtest the ValidAnytime cloud engine runs at onboarding, and shows whether — and at which exact point — it would have fired on this history. Nothing is uploaded.
Paste the full history including the suspected regression. Optionally mark where you believe it began — the engine then grades both the catch and the quiet stretch before it.
One value per line, or comma / space separated. Anything non-numeric is ignored. Drop a CSV here to load a column.
How many leading points you believe were normal — marks where the incident began, so the quiet stretch before it gets graded too.
Or load a labeled synthetic sample:
Nothing leaves your browser — parsing and the backtest math run entirely on your machine.
Your result appears here.
Paste a history or load a sample. The engine learns “normal” from the stretch before your marked onset, then shows the exact point the evidence crosses the alarm line.
How it reads your history
Learn normal
A split-conformal calibrator learns the spread of your own believed-normal stretch — your baseline, not an assumption.
Accumulate evidence
Every point either lands inside the calibrated interval or misses. Two e-process monitors turn the misses into evidence that stays flat on stable data and climbs when a real shift begins.
Alarm once, honestly
The moment evidence crosses the alarm level, you get the fire — valid however often you looked, with the false-alarm probability capped by the config.
Questions, answered
Want the whole story behind the math? Read the glossary