NeuraDock Agent Grounds LLMs for Low-Channel EEG Analysis.

While large language models (LLMs) can simplify scientific software use, they often lack specific domain knowledge, such as sensor capabilities or algorithm limitations. This is particularly problematic for low-channel electroencephalography (EEG), where sparse data can lead to plausible but unsupported interpretations. To address this, the NeuraDock Agent, an open-source architecture, has been developed. NeuraDock Agent separates a deterministic local EEG engine from a hardware-aware language layer. The EEG engine handles raw data, quality control, spectral workflows, and generates machine-readable artifacts. Crucially, the LLM only receives a concise, allowlisted summary and a detailed context pack. This context describes the specific seven-channel hardware, reviewed workflows, result fields, implementation boundaries, scientific limits, and reference cases. Raw EEG data remains local. The system was evaluated across three levels: numerical consistency, data boundary preservation under failures, and boundary-awareness through adversarial questions. Results confirmed identical structured outputs, robust data handling, and the LLM's ability to accept, qualify, or refuse queries based on its grounded context. This demonstrates a practical mechanism for calibrating EEG agents to ensure accurate and scientifically justified interpretations.