Reason. Your policies, traceable rationale.
Retrieval finds facts; reasoning takes positions. Stage three runs your policies against the graph and signs every conclusion with its rationale.
STAGE 03 · POLICY MEETS GRAPHWhat happens at Reason
This is where the brain earns the name. Given a task — decide this claim, price this quote, route this ticket — the reasoning layer runs your policies against the graph: eligibility rules, thresholds, precedents, the judgment your experts encoded as skills at capture time.
Every decision carries its rationale. The output of stage 03 is never a bare answer; it is a decision plus the policy version applied, the facts consulted (with lineage back to source records), the confidence the agent assigns itself, and the open questions it couldn't close. In the demo run: policy v2.4 applied — rationale attached.
Rationale isn't a nicety — it's what makes stage 04 possible. A human can only meaningfully approve a decision they can inspect, and an auditor can only certify a system that shows its work.
The reasoning contract
Policy as versioned artifact
Policies are authored, reviewed and gated like code. A decision references the exact policy version it ran under.
Grounded, not generative
Claims about your business resolve against the graph — grounded in your enterprise data, not the public internet.
Confidence & self-assessment
Agents publish confidence scores and open questions with every run, so low-certainty work routes to humans automatically.
Deterministic where it must be
Hybrid reasoning: rule engines handle the invariants, models handle the judgment — the mix that took BIM layouts from 3–4 hours to 15 minutes with zero collision errors.
Model-agnostic engines
The engine layer swaps as the frontier moves. Your policies and skills persist across model generations.
Cost-aware serving
Right-sized models per task keep serving costs in check without sacrificing accuracy on the decisions that matter.
We stopped buying projects and started owning capabilities. Claims that took days now close in minutes — on intelligence that belongs to us.
Pick your function. Own the intelligence behind it.
Discover one opportunity, engineer one capability, deliver one measurable outcome — then scale.