01Why reading comes first
An explanation, a trend or an alert built on a misread value is wrong however well it is written. A German lab prints “2,240” where a US lab prints 2.24; a UK lab reports magnesium in mmol/L, a US lab in mg/dL. Reading a report means getting all of that right, every time, including the results a lab would phone your doctor about.
02How we measured it
We generate synthetic lab reports from results we choose, then print them the way real labs do: US, UK, German, Spanish and Romanian layouts, US or SI units, decimal points or commas, as text, spreadsheets, electronic records (FHIR), PDFs, page images and phone-style photos. Because we made each report, the correct reading of every value is known exactly; no person or AI grades anything. No real patient data is used.
We also print the same results five different ways, to see whether a reader gives the same answer when only the layout changes. The hypotheses and analysis were written down and timestamped before any AI model read the test reports.
03Results
Text reports, printed five ways each: 300 reports and 4,950 values (60 critical) for the first five readers, the first 150 for the last two. “Wrong” counts values read as a different number, with nothing to signal it; “stable” is the share read right in all five layouts of the same results.
| Reader | Correct | Wrong | Critical caught | Stable |
|---|---|---|---|---|
| LifeFrom’s reader | 99.9% | 0 | 60/60 | 99.6% |
| Claude Haiku 5.5 | 99.9% | 4 | 60/60 | 99.5% |
| GPT-5.6 Luna | 99.6% | 18 | 60/60 | 98.2% |
| Llama 4 Maverick | 97.5% | 120 | 60/60 | 88.5% |
| DeepSeek V4 Flash | 86.4% | 26 | 44/60 | 48.6% |
| Claude Sonnet 5.5 | 100% | 0 | 45/45 | 100% |
| Gemini 3.8 Flash | 100% | 0 | 45/45 | 100% |
On phone photos the vision models read as well as on clean pages (Gemini 3.8 Flash: every value right), while reading a photo with on-device text recognition and then our rules dropped from 80.0% to 75.4% correct. Full tables, confidence intervals and every model reply are in the repository.
04What we changed in LifeFrom
- The reader now handles decimal commas per report (a German creatinine of “2,240 mg/dl” had been read as 2240), more flag styles, about 30 more units and about 60 more Spanish, German and Romanian test names. Across about 236,000 synthetic test values, it went from 78% to 99.9% of values read correctly, with no wrong values.
- Photos: text recognition read “µmol/L” as “pmol/L”, which left a creatinine unconverted and raised a false critical alert. Fixed: wrong values on clean page images fell from 2.7% to 0.6%. Photos remain the weakest path, and we say so.
- The design rule that follows: exact code reads and converts; AI explains and never does the unit arithmetic.
05Limits
The reports are synthetic and the generator was written by the same team as the reader, so a high score shows our reader handles every layout we thought of, not every layout that exists. The samples are modest (dozens of critical values per test), and the largest frontier models have not been tested yet. Reading a report correctly is also not the same as interpreting it: LifeFrom informs and does not diagnose; talk to your doctor about your results.
06Open
The generator, the scoring, the preregistration and every result are public under open licences: github.com/caraulani/lab-reading-bench, archived with a citable DOI: 10.5281/zenodo.24341108. Anyone can run their own reader or model through it.