Skip to main content

Module episode

dsfb_debug

Module episode 

Source
Expand description

DSFB-Debug: episode aggregation — Trace Event Collapse implementation.

§Trace Event Collapse — paper §7’s “primary developer-facing delta”

This module is the structural-aggregation core. It collapses per-(window, signal) anomaly events into a small number of typed structural episodes. The collapse is the operator-visible delta of DSFB-Debug versus flat alerting: 11 raw cell-level alerts on F-11 collapse into 3 typed episodes (RSCR 3.67×); 52 raw alerts on the AIOps Challenge KPI fixture collapse into 1 episode (RSCR 52×).

§Algorithm

aggregate_episodes is a deterministic run-length aggregator over the per-(window, signal) PolicyState grid. An episode opens when any signal transitions to Review or Escalate; an episode closes when all signals return to Silent or Watch for correlation_window consecutive windows. The closed episode carries:

  • episode_id — sequential id from 0
  • start_window / end_window — inclusive range
  • peak_grammar_state — max grammar state observed
  • primary_reason_code — most-frequent reason code
  • policy_state — peak policy state (Watch / Review / Escalate)
  • contributing_signal_count — distinct signals in non-Admissible state during the episode
  • structural_signature — (dominant_drift_direction, peak_slew_magnitude, duration_windows, signal_correlation)
  • matched_motif — left as Unknown; populated by the bank’s match_episode_with_consensus after fusion.

§compute_metrics — RSCR + fault-recall + clean-window FP rate

Computes the paper §13 headline metrics from a closed episode list:

MetricFormula
RSCRraw_alerts / max(1, dsfb_episode_count)
episode_precisionepisodes_overlapping_labeled_fault / dsfb_episode_count
fault_recalllabeled_faults_captured_by_at_least_one_episode / total_labeled_faults
investigation_load_reduction_pct(1 - dsfb / raw) × 100
clean_window_false_episode_rateepisodes_in_clean_window / clean_window_count

All formulas are repeated inline here so a reader doesn’t need to cross-reference the paper.

§Determinism (Theorem 9)

The aggregator is a pure function of the policy-state grid. Iteration order is deterministic (row-major over windows then signals). Tie-breakers (most-frequent-reason-code resolves ties by lower enum index) preserve byte-identical output across replays.

Functions§

aggregate_episodes
Aggregate per-window policy evaluations into episodes.
compute_metrics
Compute benchmark metrics from episodes and fault labels. Paper §7.4