ARF Specification¶

Agentic Reliability Framework

Provably safe, mathematically grounded governance for AI‑powered infrastructure.

The Framework at a Glance¶

ARF is organized as a layered specification, each building on the one before:

ARF Layered Specification

Core Concepts – reliability, observability, and traceability for AI agents.
Mathematics – Bayesian risk scoring, HMC, and expected loss minimisation.
Psychology – trust calibration, human‑in‑the‑loop design, and explainability.
Governance – policy evaluation, cost estimation, and the full governance loop.
Enterprise – execution ladder, audit trails, and deployment architectures.
Temporal Reliability – optional extension for time‑series anomaly detection and forecasting.

Why ARF?¶

Challenge	ARF Solution
Risky AI actions	Bayesian risk scoring with dynamic fusion of online and offline models
Brittle policy rules	Expected loss minimisation that balances risk, cost, and uncertainty to select the optimal action (approve, deny, or escalate).
Lack of auditability	Full traceability in every decision, with optional enterprise audit trails
Complex decision context	Governance loop integrates cost, policy, risk, epistemic uncertainty, and memory
Scaling from prototype to production	Clear boundaries between core engine (proprietary) and enterprise enforcement

Specification Sections¶

Roadmap – future direction and milestones
Design – architectural decisions and trade‑offs
Mathematics – the Bayesian engine behind risk scoring
Psychology – building trust through transparency
Governance – the decision‑making loop in action

Public vs. Proprietary¶

Component	Status
This specification (`arf-spec`)	✅ Public (Apache 2.0)
Core engine (advisory logic)	🔒 Proprietary – pilot access only
Enterprise enforcement layer	🔒 Proprietary – outcome‑based pricing
Public demo UI (`arf-frontend`)	✅ Public (Apache 2.0)

The core engine is not open source. It is access‑controlled and available under outcome‑based pricing to qualified pilots and enterprise customers.