overnight-insight-discoverylisted

# Overnight Insight Discovery ## Problem Clients often want **genuinely surprising insights** from their data — things that make them pause and re-read, not just another dashboard chart. But surfacing true ah-ha findings is hard: - **Pure LLM exploration** is creative but unreliable. Opus 4.7 with 1M context will happily narrate known facts as "surprising", or drift into statistical sloppiness. You can't distinguish a real find from a hallucination after the fact. - **Pure mechanical scans** are rigorous but narrow. A conditional-lift miner surfaces every statistically significant pair, most of which are already known to domain experts or are restatements of the model's top SHAP features. - **Single-round review** rubber-stamps whatever the analyst wrote. No independent voice checks novelty, client relevance, or methodological soundness. Without structure, overnight insight jobs produce low-signal briefs that get one read and then quietly discarded. The morning surprise wears off; the client doesn't act. This skill structures an overnight autonomous insight job as: 1. **Two independent tracks run in parallel** — one LLM-autonomous (creative latitude), one hybrid deterministic-with-LLM-narration (mechanical rigor). Each produces its own brief against the same scoping doc. 2. **Each track runs a review loop** — `agent-review-panel` (6 personas + Supreme Judge) scores the brief, `plan-review-integrator` applies findings, fix-subagents re-query BQ and