Runtime Controls

Runtime controls are where AI governance becomes real.

Policies, standards, and review checklists are necessary, but they do not control a live AI request by themselves. Orlo brings controls onto the production path so every consequential request can be evaluated, routed, checked, grounded, approved, observed, and improved.

Control Categories

Control Purpose
Task versioning Bind runtime behavior to an approved, reproducible task definition
Model routing Choose an evaluated model or strategy for the task
Fallback Move to a safer option when the primary path fails or confidence is insufficient
Validation Check structure, constraints, policy-sensitive fields, and deterministic rules
Retrieval grounding Require answers to cite approved sources where the workflow depends on knowledge
Tool policy Decide whether an agent step can call a tool or change state
Human approval Pause sensitive actions until a reviewer approves, denies, or times out
Runtime limits Bound retries, tool calls, budgets, latency, and other operational exposure
Containment Stop, redact, abstain, or send work to review when a request exceeds policy
Monitoring Track failures, review pressure, drift, quality, latency, and usage
Feedback Convert reviewed corrections into future evaluation data

Possible Decisions

Runtime controls should produce explicit decisions, not vague warnings.

Decision Meaning
Allow The request or step can proceed
Warn The request can proceed with a visible caution or review marker
Retry The system should attempt a constrained retry
Fallback The system should switch to another model, strategy, or response path
Require approval Human review is needed before the action continues
Abstain The system should decline to answer or act
Reject The request violates a policy or validation rule
Contain The issue should be isolated, logged, and sent to review

Guardrails In Context

Guardrails are most useful when they are connected to the rest of the operating loop.

A blocked response should create evidence. A failed validation should inform future evaluations. A tool denial should be visible in an agent session. A human approval should be linked to the task, deployment, and trace. A repeated failure should become a monitoring signal and then an improvement item.

That is why Orlo treats guardrails as runtime controls inside a broader control plane, not as a standalone checklist.