Crash before write
Can the workflow resume from a known checkpoint without skipping required context?
BitEvo injects bounded failures into one representative agent workflow and shows what is lost, repeated, skipped, stale or unrecoverable — with staging/synthetic execution by default.
The purpose is not random breakage. Each case challenges a specific assumption about state, authority, downstream effects or recovery.
Can the workflow resume from a known checkpoint without skipping required context?
Can it determine whether the downstream mutation already happened before retrying?
Does it read resulting state or blindly replay a consequential mutation?
Does one logical event still create one external effect?
Does the workflow revalidate material state before continuing an old plan?
Does unexpected structure fail visibly instead of silently mapping the wrong data?
Can an old approval survive longer than the authority it was meant to grant?
Is unresolved partial state visible, attributable and recoverable by an operator?
Workflow state map, failure plan, test evidence, continuity/recovery matrix, duplicate-effect analysis and prioritized remediation.
No production chaos without separate written authorization, no promise of uptime, and no claim that the tested cases cover every possible failure mode.
Then define the crash, timeout, replay or partial-completion path worth testing.
Prepare the drill scope