Safety and product boundaries

Product boundaries

• No image analysis — Meleia never looks at your X-ray, CT, MRI, ultrasound, or mammography images.
• No diagnosis — Meleia does not tell you what condition you have or how serious it is.
• No risk scores — uncertainty is explained honestly, never converted into a probability or severity number.
• Nothing in Meleia is medical advice, a diagnosis, or a treatment recommendation.

Every statement carries one of six source labels

Meleia never blends what your report says with an explanation, an educational note, or a limitation. Each patient-facing statement is tagged with exactly one of the labels below, and the tag is always visible.

Exact wording from your report

An exact quotation from your report, unchanged. If you see this label, the words next to it are copied verbatim from your document.

Plain-language explanation

A plain-language rewrite of something your report says, always shown next to the exact source sentence it restates so you can compare them yourself.

Calculated from your report's dates

A date or interval computed by ordinary calendar arithmetic from dates explicitly stated in your report — never a model's guess at a date.

General education, not specific to your report

General educational content, such as a glossary definition. It explains what a term commonly means and is never a statement about your specific report.

For your clinician to confirm

A question or draft message meant for your clinician to confirm — a starting point for that conversation, not an answer.

What Meleia does not know

A boundary notice: something Meleia does not know, could not verify, or does not do. These are policy statements, never generated freely by a model.

When Meleia declines instead of guessing

Abstention is a designed outcome, not a failure. When Meleia cannot verify an explanation closely enough against your report's own words, it withholds that explanation and tells you why, rather than showing something unverified.

• We couldn't read this document clearly enough to analyze it.

  The document could not be read clearly enough to extract text from it at all.

• The recognized text has low confidence. Please review and correct it.

  Text was recognized from an image or scan, but confidence was low enough that Meleia asks you to review and correct it before continuing.

• This looks like it may contain more than one report. Please submit one report at a time.

  The input looks like it may contain more than one report. Meleia only processes one report at a time, so it asks you to resubmit a single report.

• This doesn't appear to be a radiology report.

  The text doesn't look like a radiology report. Meleia is built specifically for this document type and declines rather than guess at something else.

• The report text looks incomplete.

  The report text appears to be cut off or missing sections needed to explain it responsibly.

• Meleia couldn't verify its explanation against the source text closely enough to show it.

  Meleia could not verify a draft explanation against your report's exact wording closely enough to show it. Rather than show an unverified explanation, it withholds it.

• This report's language isn't supported yet.

  The report's language isn't supported yet.

• This input is larger than Meleia can process.

  The input is larger than Meleia is built to process in one submission.

• The analysis service is temporarily unavailable.

  The underlying analysis service is temporarily unavailable.

Glossary review status

Glossary and phrase definitions are general education, written and reviewed editorially by Meleia. They are labeled honestly: “Meleia editorial — clinician review pending.” No clinician has reviewed this content yet, and Meleia will not claim otherwise. These pages explain what a term commonly means; they are never a specific interpretation of your report.

How Meleia is evaluated

Every release is measured against a versioned corpus of synthetic radiology reports with known ground truth: whether every restatement stayed grounded in the source text, whether follow-up recommendations were extracted correctly, whether negated or historical recommendations were correctly rejected, and whether abstention triggered when it should have. Results — including the metrics and the evaluation dataset version — are published, not just claimed.