Second Shift is a long-horizon NeuroAI harness that keeps an agent pursuing one research goal across context resets, crashes, and changing constraints: improve EEG classification of imagined hand versus foot movements. It directly addresses three parts of Long Horizon Engineering: - Sustained, coherent memory: MongoDB Atlas preserves goals, experiments, and measured results. Each decision rebuilds its context from that evidence. With 10,000 synthetic distractor notes, the evidence packet remained approximately 1,200 tokens. - Persistent goal pursuit: The campaign survives worker failure, reuses completed experiments, and adapts when the electrode limit drops from 64 to 9. All eight recovery checks and five constraint checks passed. - Learning from hard metrics: Claude selects experiments using prior numerical results. Code processes EEG and measures performance. Our selected configuration achieved 73.2% sealed test balanced accuracy across five participants and 75 test trials. We demonstrated these mechanisms on a small, real EEG research campaign. Billions of tokens and weeks of continuous operation remain scaling targets, not measured results.
Team
Placement
Participant
