Autonomy should increase only when evidence and controls increase.

NOEPRAX separates recommendation, drafting, reversible automation and high-impact actions instead of treating every AI task as equally safe.

◎

AI can analyze and suggest. A person decides.

LOW RISK

◇

AI prepares content or actions for review.

REVIEW

↗

Only reversible, bounded actions with monitoring.

LIMITED AUTONOMY

◉

Sensitive or consequential actions require explicit human approval.

HIGHER RISK

Actions outside policy or legal/safety boundaries are blocked.

NO EXECUTION

MODEL POLICY

Cheaper models handle suitable work; stronger models are escalation paths, not defaults. Paid inference belongs to paid entitlements.

MEMORY POLICY

Long raw histories should not be sent to every model call. Structured memory and relevant retrieval reduce exposure and token cost.

EVALUATION

Model quality is judged by successful outcomes, tool success, policy compliance, latency and cost — not by model brand alone.