Clarion · evidence-grounded complaint intelligence

The complaint-intelligence layer that shows its work.

Built for teams drowning in free-text complaints who need to prove why a classification happened, not just that it did — regulated-industry quality and compliance teams, public-sector complaint intake, and anyone who needs more than a tag cloud.


On the record

CASE #14595278 · Incorrect information on your report Verified — grounded in the original text
“…that is the result of identity theft and fraud…”

Verbatim from the filed complaint — the exact text that grounds this classification. If it hadn't matched, the whole result would have been downgraded to abstained and sent to a human instead.


decision_log · written 1mo ago confidence 97%

How it works

Four steps. Every complaint. Every time.

1

Ingest

Your complaint data comes in — through the live CFPB feed, a CSV drop, or your own source adapter.

2

Scrub & classify

Personal details are redacted before a complaint is persisted to the corpus or sent to any model; an LLM proposes a taxonomy code and the exact quotes that support it.

3

Ground or abstain

Every quote is checked, in code, against the real text. Anything that doesn't match exactly gets sent to a human instead of guessed.

4

Review & log

A person confirms or corrects, giving a written reason for any change. Every step lands in a permanent record nothing can silently overwrite.

See the full seven-step version →

Traceability & compliance

Every answer survives an audit.

Append-only decision log
Every decision — model, confidence, evidence, reviewer, reason — lands in a log a database trigger makes impossible to update or delete.
Evidence or abstain
Every category must cite a verbatim quote that exact-substring-matches the text. A quote that fails the check downgrades the whole classification to abstained.
Mandatory human review
Low-confidence and contradictory cases route to a triage queue. Every correction requires a written reason and feeds the evaluation suite.
Deploy inside your boundary
Swappable LLM and source adapters (Bedrock, Ollama, your own endpoint; CFPB API and CSV today) keep regulated text inside a boundary you control. ROADMAP SAP QM source adapter specced against the same interface; not yet built.

Mechanisms only — no deployment of Clarion has been through a regulatory audit, and no certification is claimed. What is described here is what the running system does.

Point Clarion at your own data

Runs today on 13,000+ real CFPB complaints. Points at yours next.

LIVE

CFPB public data this site

Real ingestion, classification, and human review, running in production.

LIVE

CSV / file-drop — your own data /import

Upload a CSV of your complaint data, preview exactly what will and won't import, then commit it into the identical pipeline. No code.

LIVE

Read-only GraphQL API /api/graphiql

Pull classified data into your own tooling with the same evidence-grounding guarantees as the UI (three fields today).

ROADMAP

SAP QM adapter — your own ERP

Manufacturing quality-management complaint/defect notifications, following the same source interface. Design sketch only — not built.

Architecture, the source adapter interface, and the API →