Brain–computer interfaces and behavior after errors
This agenda examines published approaches to decoding motor intent and tests whether error-related scalp EEG helps explain attentional lapses and retrying after failure. Intracortical recordings and scalp EEG capture signals at different sites. The agenda focuses on recording neural signals.
Inspect the public EEG benchmark
Scope of the evidence
This is an independent NOEMATHEON research agenda based on public sources. Interpret Neuralink demonstrations, registry entries, and peer-reviewed papers according to what each source shows. This phase reports task-specific decoding performance and reproducibility. Psychiatric diagnosis, clinical outcomes, stimulation, and implant development require separate validation.
Published evidence
Primary outcomes in the Neuralink PRIME record
The U.S. PRIME study is an early feasibility trial in people with tetraplegia or tetraparesis. Its registered primary outcomes are device-related and procedure-related adverse events at 12 months after implantation. The official record checked on 14 November 2026 listed 9 January 2026 as the last posted update, showed recruiting status, and contained no posted results.
BrainGate decoded attempted handwriting
Willett and colleagues converted intracortical recordings into text as one BrainGate participant with hand paralysis from a spinal cord injury attempted to write. That participant reached 90 characters per minute with 94.1% online accuracy before autocorrection. This was a result from one participant on a specific communication task.
Public EEG for comparing movement and motor imagery
The 64-channel scalp EEG recordings on PhysioNet include task markers for actual movement, imagined movement, and rest. Each marker must be interpreted in the context of its run type. The annotations identify movement tasks and rest. Error awareness and retrying require separate data.
Scalp EEG after an interface error
Ferrez and Millán studied scalp EEG error-related potentials (ErrPs) when a simulated interface executed a command that differed from the participant's intention. The separate public BNCI 2015-013 dataset supports analysis of EEG recorded while people watched an external agent make errors. A participant's own incorrect response and an external system's mistake are distinct events and are evaluated against their own annotations.
Behavior, EEG, reward prediction error, and dopamine
Incorrect responses, reaction times, and retries are observed behaviors. An ErrP is a change in EEG aligned to a specific event. Reward prediction error is the difference between a model's expected reward and the received reward. Kishida and colleagues directly measured striatal dopamine during an investment task in people with Parkinson's disease undergoing surgery. In that task, the signal fit a model combining the gap between actual and expected reward with possible outcomes of other choices. Scalp EEG and behavioral records provide no direct measurement of that concentration change.
Open questions
Does motor imagery decoding hold up in unseen participants?
Hold out whole participants to test whether a motor imagery classifier beats a simple baseline. Also evaluate across recording runs within people and test how much performance visual cues, eye movements, or muscle activity could explain.
Does an error-related signal transfer to another task?
Test EEG features learned from external interface errors in held-out participants and recording sessions. Report transfer to self-made errors as a separate outcome. Use conditions that can distinguish reported awareness of an error from surprise at a sensory event.
Does EEG add information about retrying after failure?
In a follow-up task, record failures, missed responses, time to resume, and voluntary retries. On the same test data, compare a model with EEG against one using only behavioral information available at the prediction time. Check whether any added value carries over to a separate everyday task.
Proposed study
Start with separate public reanalyses of PhysioNet motor imagery and NEMAR external-agent errors. Hold out participants and, where independent sessions are documented, hold out sessions. Repeated runs do not count as independent sessions. Fit preprocessing, feature selection, and models on training data alone. Report eye and muscle artifacts, drift in equipment or electrode conditions, false alarm rates, balanced accuracy, and uncertainty. Release the preregistration, data versions and hashes, split rules, analysis code, and aggregate results including null findings. Study actual retries and everyday transfer in a consented follow-up that measures those behaviors.