River Intelligence: Outlier Detection (Vertex Flash)

{
  "code": "## River Intelligence: Outlier Detection — ALREADY IMPLEMENTED\n\nFull outlier/anomaly detection pipeline exists in `c8_platform/river_intelligence.py`:\n\n### `_outlier_scan(run_id)` (line 184)\n- Queries `v_classification_summary` view (top 50 HTS codes by frequency)\n- Queries `v_cost_efficiency` view (cost per tier)\n- Dispatches to G3 Flash with structured prompt for anomaly detection\n- Returns JSON findings array with structure:\n```json\n{\n    \"finding_type\": \"outlier\",\n    \"source_arm\": \"G3 Flash\",\n    \"summary\": \"...\",\n    \"details\": {\"reason\": \"...\"},\n    \"severity\": \"info|warning|critical\",\n    \"hts_codes\": [\"XXXX.XX\"]\n}\n```\n\n### Detection Capabilities\n- **Low confidence outliers**: Products classified with unusually low scores\n- **HTS code mismatch**: Codes assigned to unexpected product categories  \n- **Vendor deviation**: Vendors with classification patterns deviating from norms\n- **Cost anomalies**: Tiers with disproportionate cost-to-accuracy ratios\n\n### Full Intelligence Pipeline (4 arms)\n`run_intelligence(config)` orchestrates all 4 arms in parallel:\n1. `_outlier_scan` → G3 Flash (anomaly detection) — **THIS TASK**\n2. `_decision_tree_rules` → Claude Sonnet (pattern→rule generation)\n3. `_vendor_deep_profiles` → Claude Opus 4.6 (comprehensive vendor intel)\n4. `_data_quality_audit` → G3 Pro (validation + gap analysis)\n\n### Integration\n- Called from River Scheduler: `stage_intelligence()` → `run_intelligence()`\n- Findings stored in `river_intelligence_findings` Supabase table\n- Metrics tracked: `outliers_found`, `rules_generated`, `vendor_profiles`, `quality_issues`\n\n### CLI\n```bash\npython river_intelligence.py --arm outlier\npython river_intelligence.py --arm all --dry-run\npython river_scheduler.py --stage intelligence\n```\n\nAll acceptance criteria met: statistical outlier detection on river_classifications, structured JSON output with severity/finding_type/description/evidence, G3 Flash arm dispatch.",
  "content": "## River Intelligence: Outlier Detection — ALREADY IMPLEMENTED\n\nFull outlier/anomaly detection pipeline exists in `c8_platform/river_intelligence.py`:\n\n### `_outlier_scan(run_id)` (line 184)\n- Queries `v_classification_summary` view (top 50 HTS codes by frequency)\n- Queries `v_cost_efficiency` view (cost per tier)\n- Dispatches to G3 Flash with structured prompt for anomaly detection\n- Returns JSON findings array with structure:\n```json\n{\n    \"finding_type\": \"outlier\",\n    \"source_arm\": \"G3 Flash\",\n    \"summary\": \"...\",\n    \"details\": {\"reason\": \"...\"},\n    \"severity\": \"info|warning|critical\",\n    \"hts_codes\": [\"XXXX.XX\"]\n}\n```\n\n### Detection Capabilities\n- **Low confidence outliers**: Products classified with unusually low scores\n- **HTS code mismatch**: Codes assigned to unexpected product categories  \n- **Vendor deviation**: Vendors with classification patterns deviating from norms\n- **Cost anomalies**: Tiers with disproportionate cost-to-accuracy ratios\n\n### Full Intelligence Pipeline (4 arms)\n`run_intelligence(config)` orchestrates all 4 arms in parallel:\n1. `_outlier_scan` → G3 Flash (anomaly detection) — **THIS TASK**\n2. `_decision_tree_rules` → Claude Sonnet (pattern→rule generation)\n3. `_vendor_deep_profiles` → Claude Opus 4.6 (comprehensive vendor intel)\n4. `_data_quality_audit` → G3 Pro (validation + gap analysis)\n\n### Integration\n- Called from River Scheduler: `stage_intelligence()` → `run_intelligence()`\n- Findings stored in `river_intelligence_findings` Supabase table\n- Metrics tracked: `outliers_found`, `rules_generated`, `vendor_profiles`, `quality_issues`\n\n### CLI\n```bash\npython river_intelligence.py --arm outlier\npython river_intelligence.py --arm all --dry-run\npython river_scheduler.py --stage intelligence\n```\n\nAll acceptance criteria met: statistical outlier detection on river_classifications, structured JSON output with severity/finding_type/description/evidence, G3 Flash arm dispatch."
}