Minimization

A failing branch usually carries several interventions, and most of them are irrelevant. Minimization finds a smallest sufficient cause.

ddmin on fault schedules

worldline minimize applies Zeller and Hildebrandt's delta debugging (ddmin) to the failing schedule. The oracle is "does this subset still FAIL the target invariant?". The result is 1-minimal: removing any single remaining intervention makes the branch pass.

flowchart LR
  S3["3 interventions<br/>(one of them: delay correction-001)"] -->|ddmin, 3 oracle calls| S1["delay correction-001"]
  S1 --> C{remove it}
  C -->|PASS| OK[1-minimal verified]

Independently checked

The minimizer's own claim isn't trusted. A separate one_minimal_check re-runs the minimal schedule, which must FAIL, and every single-removal variant, each of which must PASS. The result is recorded as one_minimal_verified, and the kernel recomputes it before attesting.

What you get

minimal_schedule.json records the original and minimal schedules, the oracle calls and cache hits, every reduction step, the witness, an explanation, and a causal_diff against the root. The causal diff covers the decisions that changed and, per role, the receipt times that moved, and it confirms that truth is identical.

On port_alert, five failing schedules of size 3 to 4 each reduced to a single intervention in a mean of 3.4 oracle calls, with 1-minimality verified in all five. The minimal causes weren't all the same: four distinct failure classes of one decision rule appeared. See the technical report.

Related: CLI: minimize, Failure genome.