Before you adopt CMMN — the checklist

1. Exhaust the BPMN escape hatches (map each “dynamic” claim)

[ ] Interruptions (“X can happen anytime”) → event subprocesses (7.05)
[ ] Exceptions (“when Y fails, humans take over”) → error boundaries (4.04)
[ ] Rework (“it can go back”) → loops with exit conditions (1.02)
[ ] Variants (“depends on segment/product”) → DMN-driven routing (Phase 5)
[ ] Unordered inputs (“documents arrive whenever”) → message catches per input; the work may still be prescribed

Anything left after this table is candidate discretion. Usually: nothing is.

[ ] Practitioner authority over sequencing is real, defended, and desirable — their judgment is the product (investigators, adjusters, clinicians).
[ ] The discretionary surface is large (≈ a dozen+ optional/repeatable activities; below that, one BPMN stage + hatches wins).
[ ] Guardrails still matter (sentries/milestones/required items) — otherwise build a task list, not a case model.

[ ] Quarantine: CMMN only for the genuinely discretionary stage; prescribed neighbours stay BPMN, bridged per lesson 03 (process task / case service task).
[ ] Completion semantics decided explicitly (required items, exit criteria) — no immortal cases.
[ ] Test budget acknowledges state-space cost: scenario tests per sentry, not per path.
[ ] Review capability exists: at least two people who can read sentries in XML.
[ ] Ops signals redefined per case model (“stuck” ≠ oldest instance when nothing is supposed to happen next — Phase 9.04 adapted).

BPMN by default; CMMN only where a named human’s judgment over ordering is the feature — and then only for that stage.