Watch the evidence-led demo

These two statements describe the same loan, currency, and reporting date. One says the interest is one hundred dollars. The other says one hundred twenty-five. Paper Trail Agent finds that twenty-five-dollar contradiction and shows where both numbers came from.

This is a real browser session with fictional PDFs. The public demo uses deterministic checks. Paid Meta calls and arbitrary uploads are blocked for guests.

Open the interest finding and select each source. The original statement page highlights the cited line. Principal and interest remain separate fields, and statements from different periods are sent for review instead of being called contradictory.

Now try the wrong-total case. The components add to three thousand three hundred dollars. The statement claims three thousand three hundred fifty. Python Decimal recomputes the total and confirms a fifty-dollar difference. The model does not get to invent the answer.

The internal trace separates extraction, verification, and an independent critic. Before a finding is confirmed, the critic checks the amount, calculation coverage, and evidence. After checking the source, a person can approve or reject the finding and save a note. That decision is preserved with the report version.

The scanned case takes a different path. Its image-only page has no searchable text. The app requests review and confirms no discrepancy. Missing evidence is never a clean result. Authenticated OCR is implemented and tested, but its separate worker is not enabled in the public demo.

Here is a completed local Muse Spark investigation on the same fictional statements. These are actual tool calls through a case-scoped MCP server: evidence retrieval and an exact calculator. Inputs, outputs, validation, and the bounded tool budget are visible. This is recorded live-model evidence, separate from the public guest flow.

The evaluation panels make the limits inspectable. Six baseline cases and nineteen varied-layout cases have known answers. The agent scorecard shows field accuracy, controlled critic faults, tool contracts, and retained failed runs. Retrieval precision is fifty percent in this small probe, and that weakness stays visible.

Paper Trail Agent turns conflicting statements into cited, calculator-checked findings and a human review checklist. Try the synthetic cases and inspect the public code.