Components · Trust

Confidence Meter

Decimal confidence is theater: a model that says 87.3% invites trust it hasn't earned. The meter renders confidence as three honest bands, keeps the claim and its basis together, and turns into a verification request below the threshold, so human attention is spent exactly where doubt is warranted.

When to use

Any agent output that will be used downstream: extractions, matches, classifications, summaries feeding decisions. Skip it for chitchat; a confidence band on small talk is noise.

Behavior

Confidence maps to low / medium / high bands, never a decimal. Below the threshold (default 0.75), a verification affordance appears with a plain-language nudge. After human verification, re-render with the updated claim; trust earned by a person outranks any model score.

Usage

tsx

import { ConfidenceMeter } from "@/components/confidence-meter";

<ConfidenceMeter
  claim="This invoice matches PO #4412"
  basis="3 matching fields, 1 OCR gap on the total"
  confidence={0.55}
  threshold={0.75}
  onVerify={openSideBySideReview}
/>

Install

shell

npx shadcn@latest add https://agent-pitstop.vercel.app/r/confidence-meter.json

Or copy the source from registry/confidence-meter. MIT licensed.

This invoice matches PO #4412

Based on: 3 matching fields, 1 OCR gap on the total

Medium confidence

Worth a human look before this is used downstream.

Model confidence55%