Agents that act inside jurisdiction.

Perception converts raw signal to declared state: collectors fetch, parsers extract dates, money, entities and events, and every observation keeps its source URI and capture timestamp. Cognition plans with replaceable models: the same actor survives a model swap because identity, jurisdiction and memory live outside the weights. Learning refines from recorded outcomes: dedup keys collapse repeats, SRank grades specificity, timeliness, verifiability and actionability, world-state tracks entity facts with decay, and contradiction scans surface same-entity conflicts instead of averaging them away. Action executes inside authority ceilings: tool calls carry actor, scope and approval state, writes outside jurisdiction fail closed into HOLD, and Blackbox receipts record what ran, under which credential, with what result. Traditional automation follows scripts; generative systems draft content; agentic systems close loops under oversight with a reconstructible trail.