Eval Loop — Orchestrator

Meta process skill. Turns a measurable initiative into a domain-scoped loop workspace — strategy, execution, per-cycle eval snapshots, scored result rows, promoted learnings in one folder. Owns loop setup, ledger, and schema; routes per-cycle evaluation + scoring to the matching evaluate- sibling skills (one scaffold, many evaluators).*

Core Question: "Can future agents improve this measurable surface by reading one loop folder instead of reconstructing history from scattered skill outputs?"

Responsibility split, results.tsv schema, validation rules: references/_shared/eval-loop-spec.md [SPEC].

Critical Gates

1. Measurable surface required. No nameable page / campaign / post series / ad set / email sequence / outreach motion → NEEDS_CONTEXT, recommend discover (or diagnose if the ask is about an existing metric decline).
2. Metric path required. ≥1 primary metric + source (baseline may be unknown). No metric path → no loop.
3. No skill-centered folders. Never create .forsvn/artifacts/{skill-name}/.... Loops are organized by measurable initiative.
4. Execution boundary. May execute marketing/content assets. Does NOT deploy code, publish to platforms, build app UI, or mutate external systems.
5. Human approval gates publishing and live-surface changes. Borrow autoresearch's keep/discard discipline, not its "run forever" posture.

Quality Gate

Critic (Layer 2) PASS/FAIL on: