Public data
Analyser change explorer
34 method comparisons at 14 data sets, from 12 open-access studies. Pick one to see its bias and crossings at each decision limit.
Research
Research
Our method run end to end on published data: every comparison, figure and score is open to inspect, each with its source.
- paired results analysed
- 10,579
- method comparisons
- 34
- published studies
- 12
- MedEval-1 scores re-derived
- 48 of 48
Open now
Sample
Sample report: CD4
A full analyser-change report on a published platform replacement of 1,885 paired CD4 counts.
Public data
Public data
The studies behind the explorer, each with its authors, source, licence and what we changed.
Live demo
MedEval-1
A versioned evaluation standard for medical AI, built by NakedSignal, with a sealed and rotating held-out test set, demonstrated on public data.
Analyser-Change Registry
A shared record of what analyser changes do to patient results, for laboratories that choose to contribute. Open for founding laboratories; none has contributed yet.
Guides
Practical guides to bias at decision limits, reclassification, lot-to-lot checks and comparability across sites.
Our own checks
Public data
Checks and results
6 of 6 passed
Each check was fixed before our run of 16 October 2026 on the 34 public method comparisons in the explorer. These are our own checks, not an outside validation.
Reproduces reference results (41 of 41): passed
The program reproduces its reference comparison lines and crossing counts exactly.
Re-test budget held on the fitted data: passed
On each comparison's own pairs, the share of results in the re-test band stays within the budget fixed in advance.
Re-test band holds on unseen data: passed
For every comparison large enough for a full band, the re-test share on held-out data stays within the limit fixed in advance.
Small-study safeguard: passed
A comparison too small for a full band gets an indicative grade and keeps its re-test share within the limit on held-out data.
Handles messy exports (15 of 15): passed
Awkward export files with missing columns, decimal commas, text results, duplicate IDs and similar problems each give the expected outcome.
Same result in any row order: passed
The same pairs written in two different row orders give identical results. Added by an amendment after our first public run, when a re-exported file was found to give different out-of-sample re-test shares.
Go deeper
A technical validation kit rebuilds our headline result from public data in an offline container and prints pass or fail. Our evidence library and analysis console are available on written request.