How the analysis works

Every finding you see comes from fixed arithmetic rules run over the figures on your documents. A language model is never the thing that decides what is true.

The order things happen in

  1. Extraction. The text and figures are read from the document and turned into structured line items and totals, each with a confidence score.
  2. Validation. The structure is checked, and any field that could not be read confidently is marked as such.
  3. The rule engine. Deterministic checks run over the numbers. This is where every finding comes from. Given the same document it produces the same result, every time.
  4. Redaction. If a language model is going to be used at all, names, dates of birth, identification numbers, addresses and contact details are replaced with placeholders first.
  5. Rewording, optionally. A model may rewrite a finding to read more clearly. It is given the finding and its evidence and nothing else. It has no tools, no network access and no access to any database.
  6. Validation of the output. Anything the model writes is checked before you see it. Every figure must appear in the evidence. Legal conclusions, medical advice, accusations, guarantees and invented citations are rejected, and the original wording is shown instead.

Because findings come from the rules rather than from a model, text inside an uploaded document cannot influence what we report, even if that text is written to look like an instruction.

The checks that run

Rule engine version 1.0.0.

On a single statement

  • MISSING_REQUIRED_FIELD
    Notes which details could not be read, and which checks were skipped as a result.
  • LINE_ITEM_SUM_MISMATCH
    Adds the line items and compares the result against the stated subtotal.
  • TOTAL_RECONCILIATION_MISMATCH
    Works the balance through from the subtotal, adjustments and payments to the amount due.
  • DUPLICATE_LINE_ITEM
    Finds rows identical in description, code, date and amount.
  • REPEATED_SERVICE_DESCRIPTION
    Finds the same service listed at different amounts.
  • QUANTITY_PRICE_MISMATCH
    Multiplies quantity by unit price and compares against the line total.
  • MISSING_ITEMIZATION
    Flags a large total supported by very little detail.
  • SERVICE_DATE_AFTER_STATEMENT_DATE
    Finds services dated after the statement was issued.

When comparing a statement with an EOB

  • BILL_EXCEEDS_EOB_PATIENT_RESPONSIBILITY
    Compares the amount due against what the EOB says is your responsibility.
  • EOB_PLAN_PAYMENT_NOT_REFLECTED
    Checks whether an insurance payment shown on the EOB appears on the statement.
  • CHARGE_NOT_ON_EOB
    Lists charges that could not be matched to an EOB line.
  • EOB_LINE_NOT_ON_BILL
    Lists EOB lines that could not be matched to this statement.
  • BILLED_AMOUNT_DIFFERS_FROM_EOB
    Compares the amount for the same procedure code across both documents.

Confidence, and why it is shown

A finding is only as reliable as the figures it rests on. If a line was hard to read, the finding that depends on it is marked lower confidence and phrased as a question rather than a statement. Low-confidence items are never counted in a headline and never used to prompt an upgrade.

Every finding shows the numbers behind it. If what you see on the page does not match what the finding says, the finding is wrong and you should ignore it. Please tell us when that happens.

What this can and cannot tell you

It can

  • Reads the documents you upload and extracts the line items and totals
  • Checks the arithmetic and internal consistency of a statement
  • Compares a bill against an EOB when you provide both
  • Points out entries that may need clarification, and shows you the numbers
  • Prepares request letters for you to review and send yourself
  • Keeps your documents, dates and correspondence organised in one place

It cannot

  • Know what your care should have cost
  • Know your insurance policy terms unless you upload them
  • Know whether a charge was medically appropriate
  • Know what your provider or insurer will decide
  • Give legal or medical advice
  • Represent you, or contact anyone on your behalf

What we do with your documents

Documents are encrypted, stored in private storage, and readable only by you. They are never used to train models. If you disable optional AI processing in your privacy settings, the deterministic analysis still runs in full, because that is the part that actually finds discrepancies.

Read the privacy notice · How long documents are kept · Security

Wintora provides general information and administrative assistance. It does not provide legal or medical advice, and it does not represent you before a provider, insurer, government agency, court or any other organisation.