Operational anomaly intelligence
See the signal.
Find the shift.
Transparent anomaly screening, change-point discovery and audit-ready incident review on local, reproducible data.
Signal monitor
Alert distribution
3 periodsLabeled example evaluation
Latest flagged observations
Unverified alertsSignal Explorer
Inspect the full chronological record. Scores are calculated using preceding observations only, never future samples.
Observation timeline
Record-level signal inspection
Detector Studio
Adjust a rule and immediately recompute every alert. Different detectors can disagree, particularly near shifts and seasonal peaks.
Rule configuration
Lookback samples form the past-only reference window. The first samples are intentionally excluded from alert scoring.
Detection summary
Preview under current rule
LIVE UPDATEChange Point Lab
Screen for persistent level changes with before/after rolling median comparisons. Change-point candidates use future observations and are therefore retrospective, not real-time alerts.
Potential regime changes
Change candidates
The detector uses two adjacent windows, robust scale and suppression of nearby duplicate candidates. Spikes can also trigger candidates, so human confirmation is essential.
Incident Review
Record manual dispositions for flagged observations. Each Review click cycles: Open → Investigating → Confirmed → Dismissed. Export reports before refreshing.
Investigation queue
Disposition breakdown
Manual decisions are scoped to this browser session. They are never synchronized or permanently stored by the app.
Data & Methods
Import chronological records for a local analysis, or download audit artifacts and documented methods. Nothing is uploaded to a server.
Local CSV importer
36–4000 rows. ISO-8601 timestamps, strictly increasing, numeric values and optional 0/1 ground-truth labels. Maximum file size: 1 MB.
Active data source
Malformed or invalid CSV files are rejected without replacing the dataset you already have loaded.
Audit exports
Data handling
Methods and limitations
Measures deviation from the median of the preceding window, normalized using median absolute deviation (MAD) and a minimum scale floor.
Compares each reading with a smoothed baseline. Large residuals are clipped before updating the baseline to reduce transient contamination.
Retrospectively compares consecutive windows, flags large median differences and suppresses nearby duplicate candidates.
Screening rules have no calibrated false-positive guarantee. The examples are synthetic, and alert scores do not establish operational causes.