{
  "version":"original-network-current-detection-v1-20261105",
  "purpose":"Separate after-close contemporaneous monitoring from the already frozen next-five-session prediction task. Conditional retrospective price-proxy diagnostics only; not independent anomaly certification.",
  "primary_variant":"raw_w7_l5_combined",
  "current_stress":"Either KOSPI or KOSDAQ closes at a daily return <= -2%. Consecutive stress days form one onset. This observable price definition is not the full meaning of an abnormal market regime.",
  "matched_alert":"Stress occurred on the alert day or either of its preceding two original trading sessions. Do not count an alert before the onset as contemporaneous detection.",
  "event_detection":"An onset is detected if a rule alerts on that session or within the next two original trading sessions. Report earliest delay, missed onsets and held windows; censor onset recall when the complete two-session deadline is outside the test interval.",
  "train_end":"2022-05-31",
  "test_start":"2022-06-01",
  "test_end":"2023-05-30",
  "fit_policy":"Reuse the frozen v2 models, transformations, training 95th-percentile cuts and five-session cooldown. Do not select thresholds, variants or inputs from test performance.",
  "rate_policies":["original_frozen_alerts","equal_token_capacity"],
  "equal_token_capacity":"Apply a one-token bucket to each rule's original chronological alerts. Initially one token, maximum capacity one, refill 1/20 per original trading session, consume one for an accepted alert. Held sessions remain held and still advance the trading calendar. Unused capacity never accumulates above one. This is the same maximum opportunity rate, not equal realized alert counts; report those counts explicitly.",
  "baseline_names":["index_only_combined","index_and_sector_proxy"],
  "direct_stress_reference":"Also report daily -2% stress itself with five-session cooldown and the same capacity rule. This is a definition reference that observes the target directly; its perfect same-day classification does not validate forecasting or structural anomaly truth.",
  "uncertainty":{"block_sessions":10,"resamples":1000,"seed":20261105},
  "comparison_calendar":"For each pair, require both competing alert series to be observable. Do not drop missing sessions or shorten onset windows. Report pair-specific precision and event counts on the same evaluable onsets.",
  "exploratory_gate":"Under equal_token_capacity, the fixed primary rule must detect more common evaluable onsets and have the paired precision-difference 95% interval strictly above zero against both index_only_combined and index_and_sector_proxy. Report delays and the direct-stress reference separately. A pass does not certify a true anomaly detector.",
  "required_reports":["All variants and price/volume/combined ablations","Both capacity policies and realized alert counts","False alerts, missed and held onsets, detection delays and right censoring","Paired common-calendar comparisons and uncertainty","Separate future-prediction results without conflating lead time with current detection"],
  "network_comparison_outcomes_inspected_before_freezing":false,
  "limitations":["Fixed historical stock sample and previously seen test period","Unverified original price adjustment and release vintages","Seven sector ETFs are proxies and do not cover all sectors","Price stress is not independently adjudicated structural anomaly truth","Equal capacity does not force equal realized alert counts","A historical monitor is not a live production service"]
}