A Schema-Unified Pipeline for Producing Cooperation-Behavior Data Consumable by Physical AI, Social Simulation, and Behavioral Digital Twins

Data that captures human cooperative behavior — who helps whom, in what context, and whether help propagates through a social network — is a missing input for several downstream systems: imitation/offline reinforcement learning for Physical AI, agent-based social simulation, and behavioral digital twins (BDTs). Existing sources are inadequate: social-media logs record only actions (not the much larger population of non-actions) and are contaminated by bots and engagement manipulation; interview- or survey-based methods capture stated rather than revealed behavior and do not compound over time. We present a schema-unified data pipeline that (i) defines a single standard behavior-event schema, (ii) ingests either real platform logs or a context-dependent synthetic generator into that schema under an explicit anti-leakage invariant, and (iii) exports, without modification to downstream code, three artifacts: RL transition tuples with a Gymnasium- compatible environment, agent policies with an initial network for agent-based simulation, and per-individual digital-twin models with a counterfactual query API. We formalize three properties that behavior data must satisfy to be usable by these consumers — a non-action majority, context-dependent behavior, and label independence from the candidate- construction rule — and provide an automated validation suite that checks them plus per- exporter sanity. On a fully synthetic world (≈1.7×10⁵ events) the pipeline satisfies all three properties (non-action 68.6%; same-cluster-only AUC 0.58, i.e. label not recoverable from the construction rule) while a leakage-free predictor reaches AUC ≈ 0.90. We further demonstrate end-to-end ingestion of real YOROSIX platform logs, in which empirical Ripple structures up to depth 4 across four sharing channels are present, producing all three downstream artifacts. We make no claim that models trained on the current real data areperformant; the contribution is a scale-independent pipeline and the evidence that it produces correctly-shaped, leakage-free data for three distinct consumers from a single source.