At each decision
- Hold the first pick for 60 s after the first nonzero evaluation (timed gate), so early score mass can settle
- Take the evaluation scores as Pi; look up remaining procedure times Ti from the knowledge-graph catalog
- Work the top-ranked hypothesis's not-yet-done procedures in series
- The experimenter removes the fault only when the required procedures are complete — not by anomaly name
- Wait 1 min: any alarm that drops or clears → success; otherwise eliminate that hypothesis (and any sharing the same procedures) and re-rank
Orderings compared
P/T — descending priority
Highest P first
Shortest T first
Random (5 seeds)
Expected recovery time for a fixed order π: E[T] = Σk Pπk Σj≤k Tπj. P/T is optimal when probabilities are calibrated and procedures do not overlap.