Reusable evaluation tool · Recorded EEG
Who receives a command when the decoder is uncertain?
Compare accepted errors, coverage and differences between participants, then use the same audit with your own saved predictions
Loading the archived predictions audit…
Inspect an operating point
Coverage is the fraction of trials accepted; selective risk is the fraction of accepted predictions that are wrong. The curve is descriptive, and inspecting test results does not validate a newly chosen threshold.
Intervals resample whole participants with the model and threshold fixed. Accepted-accuracy intervals condition on resamples containing accepted trials; undefined resamples are counted below. A 100–100% interval when every observed accepted trial was correct cannot account for unseen errors and is not evidence of perfect reliability.
All operating points and bootstrap counts
| Threshold | Accepted / total | Errors | Coverage | Accepted accuracy | People with commands | Valid accuracy resamples |
|---|
Participant-level results
Every participant is included, including people with no accepted commands. Balanced accuracy uses all their trials and is undefined if a participant has only one class; accepted accuracy is undefined when no trials are accepted.
| Participant | Accepted / total | Errors | Coverage | Accepted accuracy | All-trial balanced accuracy |
|---|
Audit your own predictions
Run the local CSV tool described in the reuse guide, then open its JSON report here. File contents are read in your browser and are not uploaded; the report must use the same documented schema.
What would make the evidence stronger?
This audit does not improve the decoder or establish live control. The next research priorities are a protocol evaluated on an independent dataset, preprocessing that only uses samples already received, and explicit assessment of unintended commands during rest. The research roadmap defines what each improvement would need to demonstrate.
Feedback is particularly useful on the treatment of participants who receive no commands, the limits of the bootstrap intervals and the design of an independent evaluation. A small reproduction or a concrete counterexample would help improve the project more than a general endorsement.