The five architectural concerns

Five concerns. Five non-negotiable boundaries. Policy governs. The Decision Policy is the single source of behavioral truth — risk posture, objective weights, autonomy thresholds, hard constraints, engine selection. Every behavioral choice the system makes flows from here. Decision Twin computes. The compute engine that produces decisions under the policy — the joint stochastic solve, the lookahead, the frontier construction. Change Interpretation gates. The layer that decides which events trigger a replan, which coalesce, which absorb as state updates. SAGE interfaces. The conversational layer for humans — translation, reasoning, bounded orchestration. State persists. Decision provenance, memory across cycles, the audit substrate finance defends against.

The discipline matters because of one specific failure mode. AI-augmented systems tend to accrete authority over time. AI is introduced as interpretation, earns trust, gets extended to recommendation, then to action, then to authority over decisions nobody can audit. The drift happens slowly enough that no one meeting decides it. VYAN's architecture makes the drift impossible by construction: SAGE never decides. Every decision goes through the Decision Twin under the Decision Policy's authority. The five concerns are five glowing boundaries that the system refuses to cross.

Concretely: when a planner asks SAGE to "go ahead and release that PO," SAGE doesn't release the PO. SAGE composes the request to the Decision Twin under the active Policy; the Twin runs the solve (or rejects the request if it violates a hard constraint); the result surfaces to the planner; the planner commits or defers. The conversational ease doesn't shortcut the policy audit. The same discipline applies to the in-product SAGE and to the website's SAGE you're talking with right now.