- Cheap diagnostics
- Five always-on summary stats computed instantly for every run: the longest run of frozen quotes, how many quotes are crossed, how many prints are exact duplicates, how many prints arrived out of timestamp order, and the size of the largest single-print price jump. Each is bucketed none / mild / notable / severe.
- The trap: genuine discontinuities
- 5 of the 15 clean runs carry a real, explained price jump — an earnings move or a halt/resume — with no other feed symptom. To the cheap "price jump" diagnostic, this looks identical to an unadjusted split. Jev cannot see the recorded event, only the bucketed jump size, so it will often guess wrong on these by design. Only the deterministic check, which looks up the recorded event, can tell the two apart.
- Jev's guess
- Jev never sees the raw run — only the five bucketed diagnostics. It picks the most likely incident family (or "clean") from a fixed list and returns a probability for each.
- Run diagnostic
- Executes the one deterministic check that can actually confirm Jev's guess — a run-length scan, a bounds check, a timestamp-order scan, or a recorded-event lookup. This is the only step that establishes a fact.
- Accuracy / Missed / False alarms / Brier
- Accuracy: share of the 45 runs where Jev's top guess matches the truth. Missed: a real incident Jev called clean. False alarm: a clean run Jev flagged (this is where genuine discontinuities usually land). Brier: average squared error between Jev's full probability vector and the true one-hot label across all six classes — lower is better.