{"as_of":"2026-09-22T19:00:00+00:00","country":"IR","country_base_rate":0.008778346744696415,"honest_caveats":["evidence rows cluster at hour 00 because OONI + CensoredPlanet land at daily granularity; treat hourly precision as an UPPER BOUND, not ground truth","incident first_seen is also often midnight UTC; targets inherit that coarseness","30-day temporal holdout, no country leakage. Per-country AUC is graded only on countries whose holdout has BOTH classes AND where the model gave a non-degenerate ranking (excluding the AUC=0.5 ties from single-class or flat predictions). n_countries_meaningful vs n_countries_with_auc shows the split.","calibration is per-horizon isotonic, NOT ACI-online — interval bounds are static","low-evidence countries are filtered out (MIN_EVIDENCE_24H>=1 in features); the model explicitly does NOT score 'no data' countries — endpoint returns 503 in that case","feature importance top driver is typically 'country_base_rate' — the model is largely learning WHICH countries shut things down, with hourly+rolling features modulating risk"],"horizon_hours":6,"interval_method":"static halfwidth ±0.15 (ACI not yet wired)","n_evidence_last_24h":7,"p_high":0.161,"p_low":0,"probability":0.011,"schema":"voidly-hourly-forecast/v1","top_features":[{"feature":"hour_of_day","importance_gain":0.3229773938655853,"live_value":19},{"feature":"hour_sin","importance_gain":0.234930157661438,"live_value":-0.9659258262890684},{"feature":"hour_cos","importance_gain":0.11702489852905273,"live_value":0.2588190451025203},{"feature":"country_base_rate","importance_gain":0.05966554954648018,"live_value":0.008778346744696415},{"feature":"dow_cos","importance_gain":0.055941030383110046,"live_value":0.6234898018587336}]}