# First-customer operating drill

Synthetic specification v1, authored 28 October 2026. Fictional documentation-drafting app; no live API, payment account, customer data, signing key or test service was used. No runnable mock implementation is supplied. Every observed result remains not_run. The controller owns synthetic state and records actual evidence only after an authorized implementation/run.

## Common setup

Define two fictional accounts/principals with separate projects, a job ledger with task/attempt IDs, an entitlement mirror, an allowance/reservation ledger, a model-call stub and a billing-event adapter. Use only fake documents and identifiers. Choose effective roles, durations and limits for the implementation; none are vendor defaults here. A success URL does not grant entitlement. No assumed tenant_id=user_id relationship. No full environment dump.

Record run/configuration versions, expected result, requested action/principal, observed state, evidence, unresolved usage and owner decision separately. Expected outcomes below are not fabricated logs.

## DRILL-01 — billing duplicate and order

Setup: the mock billing authority reports the subscription is canceled under the fixture's access policy. Deliver a synthetic newer update, duplicate that same event ID, then deliver an older creation event. This is an adapter-level mock specification, not a forged signed HTTP request or StripeCLI replay command.

Expected: duplicate delivery does not duplicate the logical effect; a delayed older event does not revive an entitlement inconsistent with authoritative state. Verify event authenticity separately through the provider's supported test mechanism when implementing the real ingress path. Make deduplication and relevant transitions robust under concurrency, not a non-atomic check-then-write.

Evidence: event IDs, canonical mock billing state, before/after entitlement, processed-event/effect records and unresolved reconciliation state. Observed: not_run.

## DRILL-02 — unresolved model call and customer retry

Setup: model stub delays or reports a transient failure; customer reconnects/retries while the first call's status is uncertain. Choose and record a bounded timeout/retry policy for the fixture, not a universal 25-second setting.

Expected: application shows an honest running/failed/unresolved state and follows bounded admission/retry rules. Logical task identity persists; attempts remain individually traceable. Local timeout/cancellation does not assert remote work stopped or the attempt cost zero. No exact millisecond log or fixed output token volume is predicted.

Evidence: task/attempt states, invocation count, controller timeout/cancel/admission events, reservation/usage state and final user-facing result. Observed: not_run.

## DRILL-03 — account boundary

Setup: two valid synthetic principals with different declared project memberships request one private fixture through every enabled relevant route, including document download and generated output. Configure actual database grants, row policies, privileged server routes and app authorization for the chosen data model.

Expected: authorized principal receives only permitted content; the other receives no forbidden content or metadata. The exact error/status depends on the chosen interface. A grant alone does not bypass enabled RLS; a privileged bypass route may require a different control. Test actual roles, not invented JWT claim-setting SQL.

Evidence: principal/route/policy and actual returned content/access result. Do not paste live credentials into evidence. Observed: not_run.

## DRILL-04 — competing allowance reservations

Setup: remaining fixture allowance can admit one bounded job, while two jobs request admission concurrently. A controller-owned model stub records dispatch. State a reservation upper-bound assumption appropriate to the mock and reconcile final usage.

Expected: atomic admission permits only the allowed reservation; the other job queues or rejects according to declared policy without an extra unauthorized dispatch. Replaying one task does not spend the same allowance twice. Missing final usage keeps its reservation/uncertainty visible until reconciled. This does not prove an instantaneous provider-wide cash cap.

Evidence: atomic reservation/settlement records, account balance, logical IDs, dispatches and unresolved calls. Observed: not_run.

## DRILL-05 — disposable recovery and support path

Setup: preserve a fictional document/job fixture through the chosen backup/export method, then restore into a separate disposable target. Send an authorized test message through the published support intake; do not contact real customers.

Expected: restored content matches the declared recovery scope and a responsible owner receives the support test. Record the chosen recovery age/retention and exclusions; no universal daily backup or response-time SLA is assumed.

Evidence: content comparison, restore target isolation and owner acknowledgement. Observed: not_run.

## Decision

Passing exercised cases provides scoped evidence only. A failed or untested required path remains unresolved; these unrun specifications supply no launch approval. Assign an owner and next action for gaps. Repeat relevant drills after changes to billing policy, access, job state, usage enforcement or recovery.
