Methodology and conflict of interest
Short version: study numbers come from the studies themselves, vendor facts from each vendor's own page on the date shown, sampling bounds from exact binomial math, and the test file and examples are synthetic. Medrecords AI publishes this site and sells 1 of the tools in it, so we follow the 7 rules listed under conflict of interest.
Limits: We can't vouch for any tool's error rate on your files: the studies we cite measured other tasks and other models, and vendor features are as each vendor's site stated on the capture date.
Where the facts come from
| Source | Used for | Captured |
|---|---|---|
| Published studies, read from the journal page, PubMed Central or arXiv | The evidence table and every study number on the site | 2026-11-12, linked per row |
| Each vendor's own website | The tool table: how each tool links to sources and what review it states | 2026-11-12, linked per row |
| Google results for 5 searches about AI record summaries | Which vendors and articles to read, and what this site should add | 2026-11-12 |
| medrecords.ai pricing page and help center | Facts about our own product | 2026-11-12 |
| Our own calculation | Sampling bounds, the calculator and the bound chart | 2026-11-12 |
When a study counts errors differently from another, we report each in its own terms and don't combine them. When a vendor doesn't state something, we write "not stated", which means we didn't find it on the page we read, not that the feature is missing.
How the sampling numbers are computed
The bounds on the citation checks page are exact 1-sided Clopper-Pearson upper bounds for a binomial proportion. With 0 failures in n lines, the bound is 1 minus alpha to the power 1/n; with failures, the calculator solves for the rate at which the chance of seeing that many failures or fewer equals alpha. The math assumes lines are picked at random and errors are independent. Real errors cluster on poor scans and handwriting, so the bounds are a best case.
How the synthetic test file and examples are built
The test file and every example on the site are invented. Names, dates of birth, record numbers, facilities and clinicians are fictional; record numbers start with SYN. Each trap was written to trigger 1 of the 10 failure types, and the answer key was written before any tool saw the file. We haven't published tool results on it, ours included.
Our conflict of interest, and the rules we follow
Medrecords AI publishes this site and sells Medrecords AI, an AI medical record review tool that appears in the tool table. That is a conflict. These are the rules we hold ourselves to:
- A disclosure appears at the top and bottom of every page.
- Our product is labeled as ours wherever it appears.
- Our product gets the same columns in the tool table as every other tool, and the page about it lists the failure types it doesn't catch and who should pick something else.
- Links to medrecords.ai appear only on that page and in marked boxes that say when it fits.
- We don't make claims about other tools that their own sites don't support, and we don't repeat any vendor's accuracy figure, ours included (we don't publish 1).
- The checks on this site apply to every tool's output, ours included, and the test file is built so our product can fail it.
- Vendors can correct their entries. Email aj@medrecords.ai; confirmed fixes go live within 2 business days.
How often pages are checked
We recheck the tool table and the evidence table at least every 45 days, and add studies as they are published. The "Updated" date on each page is the last time its facts were checked.
Questions about how this site is made
Did any vendor pay to be listed?
No. There are no paid placements or affiliate links, and no vendor can pay to be added or moved.
Why don't you rank the tools?
This site is about checking output, not picking a tool. A ranking would need test results on the same files, which we haven't published. The test file lets you produce your own.
Why should I trust a site published by a vendor?
Don't trust it; check it. Every study and vendor fact links to its source with a date, the math is shown, and the methods work on any tool's output.