Platform Evidence Research — Orchestrator

Pipeline skill — turns owned-account analytics (engagement, follower growth, per-post performance) into a sourced per-platform evidence base that social, SEO, short-form, and evaluation skills consume to ground recommendations in measured reality instead of intuition. Operates on the operator's own accounts (not market trends).

Core Question: "What does our own platform evidence actually say — and which recommendations does it support?"

Why this skill exists, evidence-not-intuition doctrine, credential-free posture, distinction from research-shortform, when NOT to use: references/playbook.md [PLAYBOOK].

Critical Gates — Read First

Non-negotiable before dispatching any agent (5 evidence-source types + thresholds: references/evidence-protocol.md [PROCEDURE]):

1. No invented metrics. Every number traces to owned_analytics / public_metrics / manual_export / forum_observation / prior_eval with a measured_at date. Unsourced = fabricated.
2. Source-type labeled per datum. A public-page view count is not owned analytics; an estimate is not a measurement. Never present a guess or benchmark as instrumented owned data.
3. Evidence availability is platform-specific — declare it honestly. Each per-platform section declares RICH / MODERATE / CONSTRAINED. TikTok and Instagram default CONSTRAINED — never invent retention-curve or demographic depth a platform doesn't expose.
4. Missing evidence is a gap, never a fabrication. Platform with no evidence → NO_EVIDENCE flag + "what to export" note. All platforms empty → NEEDS_CONTEXT.
5. Metrics decay — two freshness windows. metrics_window_date (30d refresh / 60d warn) governs performance numbers; algorithm_context_date (90d refresh / 180d warn) governs platform-mechanic context. Stale evidence is flagged, never silently aged.