These are architecture acceptance probes, not enabled product features. They run app-owned code in the isolated spell project against the independently built backend, using public APIs, synthetic roles and a controlled real worker. No upstream workflow code is changed to make a probe pass. Template authoring uses a fixture maintainer; ordinary requesters do not gain that grant.
| Intended capability | Demonstrated | Remaining implementation / unproven guarantees |
|---|---|---|
| New intake question / app metadata | Public template API retains an added summary and application metadata; answers survive submission | Approved wording, migration/history and UI decisions |
| App-defined graph change | An additional conditional checkpoint is authored through the public template API and traversed by the real worker | Not arbitrary new step-discriminant support or workflow version migration |
| Review / Q&A | App-owned projection and read/complete commands; 422 validation and already-completed-step denial remain authoritative | Actual review UI, nested rows and full viewer context. Immutable question history / atomic reviewed-version submission are not proven. |
| Move cycle date | Existing spell:write routing writer can update a past date in a running loose process without replaying steps | spell:manage alone is denied by current routing permissions; approve scoped grants and move/scheduling/artifact policy |
| Change routing answers / go back | Ordinary context update does not replay completed steps | A safe generic reroute/replay API and side-effect policy; past editing is not a rewind capability |
| EPL-private notes | App-owned memory store, real identity + readable-process checks, management-only direct-request guard, notes absent from core reads/exports | Mounted route/UI, durable/history/concurrency/retention/deletion behavior, final impersonation and organizational policy |
| External artifacts / automatic application | App server ownership now has a home; matching/proposals/application policy stays app-owned | Retrieval adapters, evidence association, deduplication and approval/automatic-application semantics still unproven |
| cBEAM lifecycle | Generic field/definition tools exist | Product lifecycle, scheduling and placement remain unresolved; no acceptance claim |
| Strategic package after later steps | App owns the screens/projections/document assets | The specific historical-navigation behavior has not been implemented or probed in this slice |
The current manager-only cycle-update denial is a useful finding, not a reason to broaden permissions silently. Likewise, a generic conditional-write/version contract is needed before promising “the exact answers reviewed were submitted” under concurrent edits. The ordinary-update no-replay assertion protects against accidental side effects; it does not prohibit a future explicit safe-reroute command. Keep those capabilities incomplete in the plan.
Private notes never enter process context. Their prototype denies impersonated identities and uses the backend's current process-read scope; that is not new tenant isolation or an approved final impersonation policy. The UI remains visibly unwired. These probes replace the old claim that private notes necessarily require platform field-level read ACLs, not the remaining product work.
Unproven stronger guarantees are not automatically pre-refactor blockers. Basic app-owned review UX does not become impossible merely because the platform lacks an atomic reviewed-version contract. Follow the agreed scope rule: preserve actual guarantees, and plan additional guarantees as explicit component work unless a specific feature or safety boundary requires them.