Ammonix Specialized AI Agent Technology: A Local, Auditable Architecture for Specialized AI Agents in High-Stakes Domains

Abstract

We introduce Ammonix, a foundation architecture for local, specialized, explainable AI agents in high-stakes operational domains such as clinical diagnosis, industrial and power-grid control rooms, and mission control. In these domains the responsible human operator must remain in charge, and needs an always-available local partner that knows the domain, remembers prior cases, respects hard rules, and can explain a recommendation without inventing facts. Ammonix composes parts with fixed roles: a mechanistic world model that reduces each raw observation to clear-text expert features; a calibrated classifier swarm that places the case in a Knowledge Universe, a memory of all recorded prior cases organized by the action or diagnosis they called for; the Ether, an outcome-derived success field over that universe; regional Skills that turn the local evidence into a ranked, rule-filtered recommendation and are refined offline by Retrospective Harness Optimization with Verifiable Rewards (RHO-VR); and a frozen local language model that voices the recommendation as rationale and dialog. Where a domain has few recorded cases, the universe can be built from simulated experience whose training labels are known by construction, and re-anchored by real cases as they accrue. The language model is never trained; domain competence is engineered into the harness around it and optimized there, which keeps every step auditable and lets the system run on local hardware. This paper sets out the architecture, the conditions a domain must satisfy for it to apply, and the properties that follow from its design. Instantiations sharing the same mathematical grammar, control-room operation of a waste-to-energy plant, 12-lead ECG interpretation, and healthcare claims collection, are developed in companion papers. Ammonix differs from retrieval-augmented and end-to-end learned agents in that nothing on the path from observation to recommendation is generated: every step is expert-written code, a supervised classifier, or a lookup in recorded history, and the language model only explains the result.