NeuraDock Agent: Boundary-Aware Grounding for Low-Channel EEG

NeuraDock Agent is an open-source architecture that separates a deterministic local EEG engine from a hardware-aware language layer, the authors say in a paper submitted to arXiv on 25 Jun 2026. The system is designed for seven-channel, low-channel electroencephalography and keeps raw EEG and dense per-sample arrays local while sharing a compact, allowlisted summary and a versioned context pack with the language layer.

What is NeuraDock Agent?

NeuraDock Agent splits signal processing and language reasoning: a numerical engine runs locally to parse recordings, perform quality control, execute reviewed spectral workflows, and write machine-readable artifacts, while the language model receives only a compact, allowlisted summary plus a versioned context pack describing hardware and implementation limits. The context pack specifies the seven-channel hardware, reviewed workflows, result fields, implementation boundaries, scientific limits, and reference cases, and raw EEG data remain local rather than being passed to the LLM.

The paper lists authors Zhiyuan Xu, Yueqing Dai, Junling Li, and Junwen Luo and is available as arXiv:2606.26519. The submission is 25 pages with 6 figures, and the authors present the design as a way to make scientific software easier to use while constraining what the language layer can assert about hardware and data it never directly receives.

How was the system evaluated?

The authors evaluated NeuraDock Agent at three levels with concrete reproducibility and robustness checks. First, 12 recordings produced identical structured results over ten numerical repetitions, and a complete Rest/Task run produced identical result, report, and figure hashes over three repetitions. Second, request-capture and failure-injection experiments tested HTTP, malformed-output, and connection failures and confirmed preservation of local artifacts and the tested data boundary. Third, a boundary-awareness benchmark tested 36 ordinary and adversarial questions under four context ablations and two LLMs, yielding 288 results that the paper uses to argue for hardware- and implementation-aware grounding.

Those experiments focus on correctness, boundary preservation, and the language layer s calibration against the local engine s documented limits. The authors caution that these results "do not establish clinical validity or a validated absolute cognitive-load index."

Why it matters

NeuraDock Agent addresses a common failure mode when LLMs interact with sensors: a general model can assert measurements or capabilities a sensor cannot support. Low-channel EEG has sparse spatial coverage and variable signal quality, conditions that make plausible but unsupported interpretations easy to produce. By restricting the LLM to an allowlisted summary and a versioned context pack that documents hardware, workflows, and limits, the architecture reduces the chance that the language layer will overclaim what the system can infer from seven-channel recordings.

This design also separates deterministic numerical computation from the less deterministic language layer, which supports reproducible numerical artifacts and hashes as shown in the authors' repetition tests. That separation makes it possible to keep raw per-sample data local while still using an LLM to generate human-readable reports from reviewed, machine-readable outputs.

What to watch

Look for community uptake of the code and context-pack patterns the authors describe, and for independent validation beyond the paper s reproducibility and robustness tests. The paper s experiments focus on reproducible outputs and boundary-awareness; demonstrating clinical validity or a validated cognitive-load index would require separate clinical studies and benchmarks.