Code keeps the rules.

States, permissions, gates, leases, retries and calculations. The defined structure decides what can happen next.

Judgment finds the right path.

Route intent, rank a skill, assess risk or check an output. Focused assessments make confidence and uncertainty explicit.

Your model does the deep work.

Reasoning, writing, coding and conversation run in your host. The OS supplies the organization and methods around that intelligence.

Eight methods behind the intelligence.

Intent routing

Interpret the request into a typed intent, project and relevant OS. When confidence is too low, clarification is more useful than silently choosing a direction.

Skill suggestion

Retrieve a shortlist of relevant methods, then judge which ones fit the task. A useful selector can also conclude that none of the available skills is appropriate.

Confidence-gated routing

Combine a judgment’s confidence with the risk and reversibility of the proposed action. Defined rules determine whether to continue, ask for review or stop.

Verification cascade

Use focused checks for evidence, citations, relevance and completeness. A failed check can request revision or escalate the question instead of hiding uncertainty.

Speculative fan-out

Explore independent questions in parallel when they do not depend on each other. Bring their outputs together before making a consequential choice.

Composite scoring

Ask for narrow assessments with explicit criteria. Let code apply the agreed weights and calculate the result so the arithmetic remains inspectable.

Re-ranked memory

Retrieve potentially useful context, then judge its relevance to the current question. Keep the origin of that context available for review.

Input guardrails

Treat retrieved content as evidence to examine. Check for conflicting instructions or attempts to redirect the task before incorporating it into the workflow.