A refund request can commit even when its acknowledgement disappears. If permission changes before a workflow restarts, a retry needs two separate checks: current authority and evidence about the earlier effect.
Arya.ai studies 336 synthetic histories across LangGraph and Temporal. Four application-written policies face fourteen interruption schedules and three provider contracts. Actions and approvals are scripted, isolating recovery from model behavior.
- All 168 matched host comparisons have identical checked outcomes. The differences arise from application policies.
- Per host, retained-ledger reconciliation passes all 42 supplied histories: 21 checked completions, fifteen correct denials and six unresolved reservations.
- Cached retries pass 30 histories, rechecking authority passes 36 and stable-key retries pass 31. Passing is not the same as completing a refund.
- An empty lookup can precede a delayed commitment. Unresolved actions therefore keep their reserved allowance rather than declaring the earlier request absent.
A saved receipt can also permit settlement without a new write, even when the provider offers no lookup. Recovery depends on where evidence was retained before interruption.
Caveat: The policies are application-written controls, not framework defaults. These finite deterministic fixtures neither rank production reliability nor show that LangGraph or Temporal defaults are broken.
The study applies a strict rule requiring current write authority on repeated dispatch. Other protocols must be judged against their own contracts. A separate proprietary InfinityOS component study is excluded from the native-host counts.
The paper supplies derived outcome tables, but no public execution repository. Inspecting aggregates is possible; independently rerunning the experiments requires unavailable implementation code.