Problem Analysis — Orchestrator

Strategy — Step 1 of 4. Defines the business problem, builds falsifiable reasoning via testable hypotheses, and identifies root-cause with evidence-backed tests.

Core Question: "What's actually causing this, and what does the evidence prove?"

Why this skill exists, Watanabe MECE philosophy, 3-tree-type calibration, always-cold-start rationale, 10-gate quality summary, Inconclusive-is-valid philosophy, external-factor 6-factor rationale, when NOT to use: references/playbook.md [PLAYBOOK].

Critical Gates — Read First

1. Problem statement MUST be: "[Metric] is [current] instead of [target]." No vague business problems. If the user says "things aren't going well," interview for the specific metric, current value, and target value before dispatching any agent.
2. Do NOT skip external factors — 30%+ of problems have external root-causes. The external-check-agent runs in Layer 1 alongside the tree builder. Skipping it leads to treating a symptom when the cause is environmental.
3. If/Then/Because format required — hypotheses without "because" are unfalsifiable. The "because" clause is the reasoning mechanism. Without it, a rejected hypothesis teaches nothing and the tests prove nothing.
4. Do NOT confirm hypotheses without evidence — "seems likely" is not Confirmed. Every verdict must cite a specific data point that matches or contradicts the "then" clause. Inconclusive is a valid verdict. When multiple causes changed in the same window, a matching timeline is correlation, not confirmation: a Confirmed verdict must cite evidence that DISCRIMINATES it from the co-timed alternatives (an A/B null, a per-capita/per-segment normalization, a same-segment-flat reading), and must rule out a composition/mix-shift artifact (every segment flat but the blended metric moved → the cause is the mix, not any within-segment factor). No discriminating evidence → Inconclusive. Never distribute the gap across un-discriminated candidates or invent a cause to reach ~100%.