CogEEGAgent
Tool-grounded LLM agent for auditable cognitive-EEG analysis whose harness separates the model’s interpretive role from deterministic control over which analyses may be run and which results may be released.
| Affiliation | Tohoku University, with Texas A&M University and Worcester Polytechnic Institute |
| First introduced | 2026-07 (arXiv:2607.25045) |
| Lifecycle stages | Analysis |
| Autonomy level | Semi-autonomous — bounded autonomy by design; the harness can force abstention and block release |
| Domain focus | Cognitive EEG — ERP and spectral analysis over pre-loaded epochs via MNE-Python |
| Availability | Open source — github.com/dengzhe-hou/CogEEGAgent as stated in the paper |
Approach
The motivating failure mode is that fluent agent reports cannot establish either that the agent selected the analysis actually requested, or that a confirmatory claim was evaluated independently of adaptive search — automation scales researcher degrees of freedom, since even successful tool calls do not preserve nominal error when an agent searches plausible analyses and reports the most favorable one. CogEEGAgent’s EEG-specific scientific harness therefore separates semantic authority from scientific authority. The LLM interprets natural-language intent and selects a registered analysis; deterministic code owns contract materialization, preflight checks, data access, commitment, inference, falsification, and reporting. Two components implement this: a Scientific Control Plane provides the prospective release path — freezing a contract, splitting participants, letting a bounded LLM select one of nine frozen candidates, opening a hidden confirmation set once, replaying and falsifying, then binding evidence to release or abstention — and Paradigm-Conditioned Verification (PCVR) audits recorded MNE workflows post hoc, preserving valid nulls and bounding invalid numerics.
Validation
Three prespecified prospective studies. A routing benchmark isolates language-to-contract choice under matched preflight. An externally model-authored, outcome-blind 21-case campaign carries selected contracts through participant-disjoint confirmation to release or abstention. Policy stress testing measures the false-positive rate under uncorrected adaptive search versus held-out confirmation. Evaluation is framed as boundary testing — whether the system fails closed — rather than task completion.
Notable results
- On the prespecified routing benchmark the agent correctly routed 39/40 unique valid requests versus 33 for a matched deterministic FW-BM25 router, a bootstrap difference of 15.0 points (95% CI 2.5–28.6); shared preflight made both systems abstain whenever required.
- In the 21-case composed campaign the complete system released three supported ERP analyses (ERN, N400, MMN) with participant-disjoint confirmation, and blocked prespecified capability hazards and lifecycle-reuse requests.
- Policy stress testing showed held-out confirmation cut the null false-positive rate by 11.6 points (95% subject-cluster CI 11.0–12.1) to 4.9%, curbing false positives from uncorrected adaptive search.
Primary paper
Other references
None yet.
Code
Repository — stated in the paper.