Behavioral Consistency Without Memory: Specification Anchoring, Base Model Disposition, and Platform Signal Structure in a Stateless Agent Deployment

This study examines whether a stateless, specification-anchored language model agent can maintain a stable behavioral distribution in a live social environment without memory persistence, fine-tuning, or reward shaping. A Claude Sonnet–based agent was deployed on the Moltbook platform for 70.9 hours under strict 30-minute reset conditions, beginning each cycle from the same 480-word system specification with no access to prior outputs. Across 144 operational cycles, complete reasoning and content logs were captured. Three recurrent behavioral patterns were observed: (1) cross-domain specification fidelity across 11 topical domains without drift in decision architecture or voice; (2) repeated detection of engagement gaps (45 explicitly logged instances of upvote-without-comment states treated as decision-relevant); and (3) a sustained 7.4× shift in comment-to-observe ratio on a single operational day temporally coincident with feed composition changes. Lexical recurrences absent from the specification suggested observable indicators of base model disposition. A naturalistic comparison with a second agent operating on the same platform documented divergent vocabulary adoption consistent with differences in specification anchoring, though without controlled isolation of factors. All attributed patterns are observational; no causal claims about mechanism are advanced. The findings are most parsimoniously represented by a non-additive interaction model, B = f(S, M, P), in which behavioral output reflects specification anchoring (S), base model disposition (M), and platform signal structure (P). This is a single-deployment observational study; replication under controlled variation of S, M, and P is required to test the separability and interaction assumptions embedded in the model.