Event Reasoning System records 12-step dependency chains and 51.6% inference time reduction under frozen rules

The paper formalizes the Event Reasoning System as an explicit state machine in which semantic nomination and declared conditions gate rule execution. Low-level rules execute local transformations; high-level control rules recursively bind subgoals and admissible paths under task constraints, budget, and information gaps. Admissible transitions are defined so that any revision preserves dependencies and rejects paths that violate period, unit, or evidence requirements.

On structured-event graphs the system reaches twelve-step chains, reduces executed irrelevant operations from 32 to 0 at lookahead 16, and passes all 12 focused checks (nine unseen derivations, three justified refusals) plus all six fact-change cases with rules frozen. Dynamic subgoal configuration accounts for the measured 51.6% time saving relative to global single-level control.

Hierarchical composition converts a fixed library into task-specific paths without storing complete solutions, directly addressing premise-change fragility in multi-agent settings. The explicit graph supplies the audit trail required for verifiable refusal and revision, a property absent from implicit chain-of-thought or ReAct traces.

Next steps include scaling the finite rule library to open-ended agent environments and measuring whether the same dependency-preservation guarantees hold when rules are learned rather than declared.