Research context. This benchmark uses publicly available naturalistic human–AI conversations and is included as ecological corroboration that trajectory-level effects also appear outside the controlled synthetic corpora. It is secondary to the site's primary commercial evidence, which comes from the controlled longitudinal evaluation.
72.8%Recurrence MAE reduction vs causal persistence at +10 turns
62.2%Historical Organization MAE reduction at +10 turns
53.5%Interactional Organization MAE reduction at +10 turns
Replication. Recurrence improved in 678/680 eligible conversations, Historical Organization in 679/680, and Interactional Organization in 672/680. These results support naturalistic generalization, but they are presented here as research validation rather than the core commercial benchmark.
Scope. The strongest current commercial-facing evidence remains the controlled longitudinal 1,000-conversation evaluation, which directly tests long-horizon trajectory prediction, motif structure, delayed influence, and +15/+20 forecasting. The naturalistic benchmark is retained to show that the signal is not confined to synthetic data.