Numbers are never guessed: where the model stops and code takes over
A language model is good at reading and explaining. It is not a calculator. In Vitrubo, every value, unit, range and flag is computed by deterministic code before any model sees it.
The most common failure of general-purpose chatbots on lab results is not bad medical reasoning. It is bad arithmetic and bad transcription: a decimal point moved, a unit silently swapped, a reference range invented because the report did not print one. Each is small. Each can reverse the meaning of a result.
Why language models are the wrong tool for numbers
A language model produces the most plausible next piece of text. For explanation that is a strength. For a value that has to be exactly right it is a liability, because the plausible number and the correct number are usually close — close enough to look right, and occasionally different enough to matter. Glucose of 5.5 mmol/L and 5.5 mg/dL differ by a factor of about eighteen; a model that confuses the two will explain the wrong one fluently.
The split
Vitrubo divides the work along that line:
- Deterministic code extracts values, normalises units, attaches reference ranges and computes abnormality flags. The same input always gives the same output, and every step can be tested.
- The model reasons only over data that has already been validated: what the results mean together, how they relate to the person's history, what is worth asking next.
The model is allowed to explain a number. It is never allowed to produce one.
What this buys
It makes the failure modes separable. If a value is wrong, it is a parsing bug with a test that reproduces it. If an explanation is weak, it is a reasoning problem, and the numbers underneath it are still right. Nobody has to wonder whether a flag was computed or imagined.
More on the pipeline is on Under the hood.
This article is general information, not medical advice. Talk to a clinician about your own results.
