Cognitive State as Behavior Signal: Synapse Parameter Predictive Validity in a 1,743-Hour Multi-Agent Panel

This paper reports the first systematic empirical test of cross-time predictive validity for the synapse parameter framework underpinning a personal cognition layer. Over a 127-hour continuous window (2026-05-02 to 2026-05-08) in the Lobster Observatory Substrate—a Mandarin-language multi-agent LLM substrate of 20 agents grounded in a live consumer prediction ecosystem—14 internal cognitive parameters were recorded at hour t paired with 10 behavioral outcome metrics at hour t+1 for every agent–hour pair, yielding 1,743 panel observations and 140 parameter–outcome correlation tests.

Every one of the 14 parameters showed at least one statistically meaningful cross-time correlation with at least one outcome. The strongest correlations approach causal-strength magnitudes: threat_body_signal predicts next-hour message volume at r = 0.835 (n = 1,743); arousal predicts the same outcome at r = 0.832; metacognitive_calibration predicts next-hour money-pain at r = 0.634, while model_revision_pressure predicts it inversely at r = −0.566. These results constitute evidence that the synapse parameter framework satisfies the predictive-validity criterion required by the Afu Brain architectural specification: a parameter is real if and only if its current value changes future decisions.

The dataset is released in full alongside this manuscript under CC BY 4.0. The paper discusses methodological caveats including multicollinearity between somatic-pressure parameters (partial-correlation decomposition shows r = 0.969 between the two leading predictors, with each retaining a small but statistically reliable independent contribution after partialling), the inclusion of one human participant in the panel, and the limits of single-substrate, single-language evidence.