# Proposed failure-injection pilot

Use a named test environment and disposable records. This is a starting plan, not a validated sample size. Test each category during active work and after restart. Define expected behavior before running; retain all failed results.

| Category                         | Active-work result | After-restart result | Evidence and unresolved questions |
| -------------------------------- | ------------------ | -------------------- | --------------------------------- |
| Worker or connection dies        |                    |                      |                                   |
| Repeated tool failure            |                    |                      |                                   |
| Unknown write result             |                    |                      |                                   |
| Expired permission or approval   |                    |                      |                                   |
| Changed source                   |                    |                      |                                   |
| Cancellation or exhausted budget |                    |                      |                                   |

Record the same run identity, preserved budget, artifact version, operation identifiers, destination receipts, authorization decision and final status. Measure usable outcomes, review time, recovery cost, duplicate effects and total job cost.