Console · Clarion

How Clarion works AI that classifies real consumer complaints — and proves every call with evidence.

Clarion reads messy, free-text consumer complaints from the public US CFPB database and sorts each one into a controlled taxonomy of failure modes. Every classification must quote verbatim evidence from the complaint, or the model abstains. Anything low-confidence or self-contradictory is routed to a human reviewer — the system never auto-decides a shaky case.

The labels you'll see

Accepted The model's classification cited verbatim evidence and cleared the confidence bar.

Abstained The model could not support any category with evidence, so it declined to classify — deliberate honesty, not a failure.

In review Low-confidence or contradictory — routed to a human reviewer instead of auto-deciding.

Queued Pulled from the source and awaiting the pipeline.

Grounded Evidence spans have been located in the narrative.

The pipeline rail

Each complaint moves left to right — scrub → embed → classify → ground → decide. When you press Classify, the rail lights up live as each stage completes.

Why the text says XXXX

Names, account numbers and other personal data are redacted — first by the CFPB, then again by Clarion — before storage or any model call. The XXXX blocks are that redaction.

Pick a complaint to inspect it

Choose any complaint on the left to see its redacted narrative, the evidence the model cited, its confidence, and the full audit trail.