How policy analysts think — and where the work breaks under generic AI.

The options-memo skeleton

• ▸Problem statement — bounded, dated, citing source memos.
• ▸Stakeholder map — Actor instances with role + jurisdiction populated.
• ▸Constraint inventory — Constraint(type=Statutory|Regulatory|Mandate|Norm|Resource|Capacity|Jurisdictional|Procedural).
• ▸Options — each option is a hypothetical PolicyOption extending Commitment; lives in the Scenario graph, not the Analytical graph.
• ▸Trade-offs — typed BLOCKS / ENABLES / CONSTRAINED_BY edges between options and constraints.
• ▸Reception model — StakeholderResponse predictions per Actor; Wind Tunnel surface.
• ▸Recommendation + dissent register — narratives, drift logged.

Where Kingdon and Sabatier sit in the kernel

Kingdon's "three streams" (problem, policy, politics) and Sabatier's advocacy-coalition framework both live above the kernel as analytical lenses. The kernel does not encode them directly — it encodes the substrate they operate on. An analyst writing through a Kingdon lens reads from Analytical and Narrative and writes a memo; the memo cites typed nodes; the typed nodes survive the analyst.

The constraint-stress problem

Most policy options fail not at announcement but at first contact with a binding constraint — usually statutory, sometimes capacity, occasionally a stakeholder reception nobody modelled. A generic LLM cannot reliably tell you which constraint will bind first because it does not maintain a typed constraint inventory across the file. The kernel does. The system can return a deterministic answer: "Option P's first binding constraint is Statutory §A.4; the responsible actor is Agency X; the historical response under similar conditions is documented at edges Y, Z."