How a readout is produced.

What the system does

Lab Readout takes a document that someone else produced — a Certificate of Analysis or a comparable material document — and does four separable things with it: it extracts what the document states, it locates each extracted value on the page it came from, it explains the terminology in plain language, and it names what is missing or unclear.

The interface keeps those four things visually distinct throughout, because they carry different weight. An extracted value is a claim about the document. An explanation is a claim about terminology. A located page reference is checkable. A gap is an observation about the upload. None of them is a laboratory measurement.

Extraction and evidence

A submitted file — PDF, JPEG, PNG or WebP, up to 20 MB — is stored as an immutable version. Extraction runs against that exact version, and every extracted value is recorded together with the page it was read from.

Source links resolve against the specific document version that was analyzed, not against the newest file you have uploaded. Uploading a corrected document creates a new version and a new analysis run; it never rewrites a report you already have. Where validated coordinates are unavailable, the readout falls back to a page reference and a quotation rather than drawing a highlight box that does not correspond to anything.

Original strings are preserved. A reported result of ≥ 99.0% is not silently converted to 99.0%, because the operator changes the claim. Units and measurement basis are kept with the value for the same reason.

Explanation

Producing the explanation means sending document text to an external AI provider. That is a deliberate design choice and it is stated before you upload, not in a footnote afterwards.

Explanations are grounded in the document in front of the system. Where an explanation draws on general reference material rather than on your document, the reference is identified separately, so you can tell which sentences are about your file and which are background.

Model self-reported confidence is not displayed as an accuracy figure. A model's own estimate of its certainty is not an independently measured error rate, and presenting one as the other would be a fabricated statistic.

Document checks

Each check reports its own outcome, using these states and no others:

• Check passed — that named check passed.
• Inconsistency found — two parts of the document disagree.
• Needs review — there is something to compare, but not enough context to compare it safely. Mismatched units or an absent specification produce this, not a guess.
• Not located — the field is not present in this upload.
• Not applicable — the check does not apply to this document.
• Not checked — the check was not run.

There is no overall document score and no composite grade. A single number would invite the exact misreading this tool exists to prevent: that a document check is a verdict on a material.

Issuer-record checks

Checking a report against the record of the laboratory that issued it is a separate operation with its own status, timestamp and scope. It is only ever automated where a supported issuer adapter exists.

Where the readout displays an official link, that is a link — not evidence that a check occurred. Clicking an external link is not a completed verification. A website that is unreachable, or a request that fails, produces "could not complete the check". It never produces a fraud warning, because an outage is not evidence of anything about the document.

URLs found inside an uploaded document are not treated as trusted verification endpoints and are not automatically fetched.

Human corrections

If you believe an extracted value is wrong, you can propose a correction with a reason. The original extraction, your proposed value, who proposed it and when are all stored. The correction is labelled as user-provided and does not overwrite either the issuer's wording or the original output.

A stored correction is not a professional review. It records that someone disagreed, which is useful, and nothing more than that.

What is never asserted

The system does not produce, and will decline to produce:

• a dose or a dosing schedule;
• a treatment recommendation;
• a patient-specific or diagnostic conclusion;
• a judgement that a material is safe, or appropriate, for human use.

It also does not assert that a document is genuine or fraudulent. Unreadability is not evidence of forgery, and the system does not convert one into the other.

Review and revision

A readout is a revision, not a final answer. Re-running an analysis creates a new run and a new revision, and previous revisions remain available with the document version they were produced from — so you can see what changed and why.