Interventions

An intervention changes how the environment delivers one or more events to a role. A tuple of interventions is a fault schedule. Applying a schedule to the root world produces a branch.

The seven operators

Operator	Effect at delivery	Parameters
delay	Shifts the receipt time of the targeted delivery later.	by_ns
drop	Suppresses the delivery.	none
duplicate	Delivers extra copies (<id>#dup<n>) with provenance back to the original.	count, spacing_ns
reorder	Pushes a delivery past the next event of the same kind.	by_ns, or margin_ns
corrupt	Sets a payload field, lowers quality, and marks provenance.	field, value, quality
revoke	Cuts a role's feed from a point in time on.	role, from_ns
clock_skew	Offsets receipt times for everything from one source.	source_id, offset_ns

A target is either concrete ({"event_id": "correction-001"}, optionally with a role) or abstract (kind, role, position). Abstract targets are what a failure genome carries.

Where interventions act

sequenceDiagram
  participant T as Truth
  participant H as Delivery hook (schedule)
  participant R as Role view
  participant A as Actor
  T->>H: correction-001 at 1407 s
  Note over H: delay past the decision
  H->>R: received after 1500 s
  A->>R: AS_KNOWN_AT(1500 s)
  Note over A: correction not yet known, so escalate

Interventions are applied when an event is delivered to a role, never to truth. A schedule that would make an event arrive before it happened breaks the time rule. Such a schedule is recorded as rejected and is never silently skipped.

Exploration: guided versus random

worldline explore samples fault schedules of a fixed size within a budget, using two strategies:

• Random draws uniformly. It is the baseline.
• Guided prioritizes interventions that touch events the decision rules depend on. It reads only the scenario, never runtime truth, so the comparison is fair.

Both strategies run on the same budget and are reported side by side: failing branches per 100, the index of the first decision FAIL, and wall time. Duplicate schedules are deduplicated by hash and counted against the strategy that produced them.

Related: CLI: explore, Minimization.