API
Reusable Python HTTP transport with bounded retry, throttling, credential-header, and redirect controls.
View repository ↗Systems
Strategy, authorization, runtime behavior, and context have distinct owners and testable interfaces.
Control plane
Coordinates priorities and bounded work across execution channels; the components below retain their own policy, runtime, and context responsibilities.
Core / 01—03
Policy enforcement
Evaluates submitted plans against caller-supplied step limits, tool controls, and impact classifications. Accepted sequences are translated and submitted to APEX without probabilistic replanning.
View repositoryAgentic Process Executor
Combines plan validation, durable execution records, crash recovery and replay across natural-language tasks and exact execution plans.
View repositoryRetrieval and storage
Provides document ingestion, embedding-space identity, semantic retrieval, and grounded generation through library and versioned HTTP interfaces. A dedicated storage owner serves runtime consumers through explicit library and HTTP interfaces.
View repositoryASON evaluates and submits accepted plans to APEX. APEX validates and executes the submitted sequence while durably binding the ASON authorization identity and exact approved-plan digest before dispatch. RAG is accessed through explicit adapters and remains separate from policy enforcement and execution. Run the deterministic core-stack reference ↗
Supporting systems
Reusable Python HTTP transport with bounded retry, throttling, credential-header, and redirect controls.
View repository ↗Local Docker Compose integration and cross-repository validation for the public runtime portfolio.
View repository ↗Operational tooling for encrypted backups, guarded Git synchronization, incremental log monitoring, and deterministic Markdown-to-HTML conversion.
Private repositoryDeclarative scene automation and reproducible headless rendering for Blender workflows.
View repository ↗Bounded integration examples for logistics, telephony, and local voice automation, each with a distinct operating envelope.
View repository ↗Zero-dependency browser encyclopedia with streamed Gemini generation, cancellable navigation, memory-only credentials, and explicit untrusted-output handling.
View repository ↗Model choice
AXIOM uses a model-agnostic architecture: planning connects to policy, execution, and retrieval through explicit interfaces. Choose hosted or self-hosted LLMs without binding your control plane to one vendor. ASON evaluates proposed plans, APEX executes accepted tool sequences, and RAG serves context independently of the model generating an answer.
Integrations are validated against the selected provider, API, model capabilities, data requirements, and acceptance tests. Switching providers preserves the engineering boundaries; each replacement is checked for tool support, output contracts, quality, latency, and cost.