Saga and replay assessment with calibration
Attempt before calibration. Score 0 absent/incorrect, 1 prompted/incomplete, 2 independently correct with evidence and limits, 3 fresh transfer. Every category requires at least 2. A passing coordinator model does not establish durable workflow behavior.
Required categories
- EV1 — Contract: identify commands, immutable outcomes, workflow/step identities, ownership, ordering and retention. Preserve original schemas and compatibility rules.
- EV2 — State reasoning: predict timeout/late-success/compensation traces; distinguish unknown, cancelling, cancelled and completed. Supply three independent traces and an irreversible-step policy.
- EV3 — Implementation: implement the archive, demonstrate duplicate no-ops and non-mutating rejection of contradictions, and explain finite model limits.
- EV4 — Durable workflow: integrate the coordinator with original atomic outbox/consumer work across three services. Crash at persistence, dispatch and confirmation boundaries; retain compensation obligation and prove safe retry behavior.
- EV5 — Poison operations: quarantine, diagnose, repair and replay an invalid/unsupported message with provenance, stable business identity, controlled rate and stop conditions. Exercise duplicates and later state changes; retain operator evidence.
- EV6 — Recovery judgment: demonstrate failed compensation and a permanently unavailable participant. Assign escalation ownership and explain why no unsupported terminal outcome is claimed. Defend orchestration versus choreography for the original workflow.
Calibration — after attempting
Timeout with both outcomes unknown gives cancelling with no known compensation due. Payment success adds a refund obligation. Refund confirmation removes that obligation but does not resolve unknown reservation outcome. A definitive reservation failure permits cancelled. If both operations succeed before cancellation, the model completes and rejects a subsequent timeout; post-completion cancellation requires a separate business workflow.
Repeated definitive outcomes and compensation confirmations are harmless. Conflicting definitive results are rejected, not silently overwritten. Compensation without a known successful step and cancellation intent is invalid. Failed compensation supplies no successful confirmation, so the obligation stays open.
The due tuple represents intent, not a durable dispatch or exactly-once effect. Quarantine is unresolved work; replay with a fresh broker identity must retain business identity. Finite tests do not establish eventual participant response or liveness under permanent failure.
Delayed transfer
After at least 48 hours, independently analyze a reordered trace with a late success and an unavailable compensation endpoint. Add an irreversible shipment step and explain how the business recovery policy changes. Preserve attempts, repairs and assistance disclosures; author reference tests are not learner service execution.