What each number measures
and what it does not prove

The stopwatch starts when the radiologist opens the editor for that exam and stops at signature. Left out of the count are image acquisition, the PACS queue, study transfer and any administrative step before or after. The metric covers exactly the stretch where the platform acts, and nothing beyond it.

That boundary is a choice against our own interest. A TAT measured from the exam's arrival to signature would yield a larger number of minutes saved. And it would largely measure the institution's operation, not the product. We prefer the narrow window: it is smaller, it is auditable and it answers only for our part of the workflow.

The available source confirms the window, the sample and the start and end events. It does not publicly document stratification by modality, complexity, institution or a detailed outlier policy. That gap prevents treating the result as representative of any service.

In the historical snapshot, the median is 52 seconds. Among the reports, 54.6% are under one minute, and the reported mean is between 3 and 5 minutes. The three values describe the observed distribution, but do not explain causality or clinical quality.

The estimate of about 10 hours per 100 reports uses a different class of evidence. It combines the rounded internal mean of 4 minutes with an approximate external baseline of 10 minutes. It is not observed time saved in a paired comparison.

We withdrew the claim of a reduction of up to 78% because it mixed statistics and a baseline without a reproducible paired comparison.

The approximate baseline of 10 minutes per exam is attributed to Forsberg et al., Journal of Digital Imaging, 2017. This page classifies it as a literature-based estimate, not as a Laudos.AI measurement.

The external study's modality mix, year and setting may differ from your service. A locally measured baseline should prevail in the institutional pilot.

Time does not describe quality. LaiBench is a controlled internal fidelity benchmark on a published, versioned set. It does not constitute external clinical validation. The AI structures and suggests; the physician reviews, edits, validates and signs, in accordance with Resolution 2.454/2026.

The LaiBench score and editor time measure different things. Neither of them, alone or combined, proves clinical outcome, safety in any context or superiority over competitors.

The honest way to check is to reproduce it. Before changing anything, measure in your service the time between opening the editor and signing, using the same event definition as this page. Record the median and the mean. Then run the pilot and measure again, over the same 30-day window and without discarding any case.

If the difference in your environment is smaller than ours, the number that counts is yours. We want to know: we publish the methodology here so you can challenge it with measurement, not opinion.

The published value corresponds to the fixed period from 2026-04-09 to 2026-05-09, with 5,200 reports. Without access to a more recent verifiable internal source, it remains frozen with its date.

Every page that cites the median must point to this snapshot. Updating the date or calling the number a rolling window without recalculating the source would be fabricating recency.

It is speech recognition WER, not clinical correctness. The number says that the dictated word came out written as it was spoken. It does not say the report is right. LaiBench is what answers for correctness, and the two measurements do not replace each other: a perfect transcription of wrong reasoning is still wrong.

A direct comparison with other systems does not exist, and we are not going to fabricate one. The numbers they publish are in English, on their own test sets, with their own normalization criteria. Stacking everything on the same bar would make a pretty chart and an indefensible claim. What this number supports is narrower. In radiology dictation in Brazilian Portuguese, with medical terms inside sentences, the transcription got 3.06% of the words wrong.

The sentences are controlled, and that is a limitation, not a detail. A shift has room noise, a bad microphone, accents, fatigue and interruptions. The reading-room number is the one that shows up in your pilot, with your audio. If it comes out worse than this one, yours is what counts.

A public sample is not yet documented for the WER. Missing are the number of audio clips, speakers and devices, as well as specialties and the normalization protocol. Until that is published, we treat the result as a controlled internal benchmark, incomplete for generalization.

The measurement starts with the editor open and ends at signature. It includes dictation, structuring and review within that interval. It is not the total time since the exam arrived. And it does not claim that every case takes the same time. The median describes the center of the distribution, while complex cases remain longer and call for separate analysis.

The 54.6% rate under one minute sizes the operational effect, together with the estimate of hours returned. Neither replaces the institution's baseline. Modality, complexity, experience and template quality change the result. That is why the pilot measures the team itself with the same criteria before and after.

A clinical adoption should also look at corrections, addenda, template adherence, disagreements and the perception of those who review. Gaining seconds while increasing rework would not be a gain. The public methodology exists so anyone can question, reproduce and compare the number against quality, not to end the conversation.