Humanmax & Compress — Orchestrator

Polish-pipeline for existing AI-drafted copy: strip AI tells, inject brand voice, compress with zero idea loss so the text reads human-written. 6 sub-agents. Capability metadata: routing.yaml. Agent table + routes + calibration + detector thresholds: references/agent-manifest.md. Methodology: references/playbook.md.

Triggers: "sounds like AI", "too robotic", "humanize this", "ai slop", or terminal call from write-copy/write-ad/write-outreach (Route C). Core question: Would a human editor believe a human wrote this, and cut nothing?

Critical Gates — load first

Full text + --fast carve-outs: references/procedures/critical-gates.md.

1. Pattern scan first — strip without diagnosis is guessing.
2. ZERO em dashes in final output.
3. Strip before voice — soul-injection on AI prose is polishing a turd.
4. Content type matters — short-outbound caps compression at 0-10% to protect tokens.
5. Detector resistance is structural — run detector pass for high-stakes content.

Absolute Prohibitions — zero tolerance

A single instance = critic auto-FAIL. Full list + enforcement: references/procedures/absolute-prohibitions.md.