React Anomaly Explainer

A control chart tells an operator that a point is out of bounds. The value an assistant adds is the why — but only if it shows its evidence and its confidence, so this pairs the explanation with the specific signals behind it and a suggested next step, never a bare verdict.

EQ-07 chamber pressure

Out of control

Explanation

82%

Chamber pressure broke the upper control limit at 12.4 mTorr and is trending up, not spiking — a sustained shift rather than a one-off transient.

Evidence

Last 3 points10.2 → 12.4, rising
Turbo pump current+18% vs baseline
Pump vibrationabove warning

Likely cause

Turbo pump bearing wear — the rising baseline and elevated pump current point at the pump, not a chamber leak.

Suggested: Place EQ-07 in maintenance and schedule a turbo pump service before the next campaign.

The chart flags the out-of-control point; the panel explains why, cites the signals behind it, and shows its confidence.

Installation

npx shadcn@latest add "http://localhost:3000/registry/anomaly-explainer.json"

Props

Prop	Type	Description
title	string	Chart heading
series	number[]	The data points
mean	number	Centre line
ucl / lcl	number	Upper / lower control limits
anomalyIndex	number	Index of the flagged point
explanation	AnomalyExplanation	{ summary, evidence, likelyCause, confidence, suggestedAction }

The chart flags, the panel explains

Detection and explanation are different jobs, and this shows both. The control chart draws the mean, the control limits, and the series, and rings the out-of-control point in red — that is the what. The panel is the why: a plain-language summary, the evidence it rests on (the rising run, the pump current, the vibration), the likely cause, and — crucially — a confidence, because an 82%-sure explanation of a failure should not read like a certainty.

Evidence, not assertion

A model that says "it is the pump" is worth little; a model that says "it is likely the pump, because current is up 18% and the trend is rising rather than spiking" is worth acting on, and lets the operator disagree if they know something it doesn't. The suggested action closes the loop from anomaly to decision. Reserve this for genuinely ambiguous flags — where the why is non-obvious — and let the Verify Banner and Confidence Meter carry the honesty conventions the rest of the AI set uses.