Project Gantt

🔒 Internal · Updated Aug 22, 2026 (v5.1 · **🔁 AGENT LEARNING PROGRAM — PHASES 0–8 BUILT (DEV).** Every agent prompt is a database row; memory is assembled once with a hard budget; prospects carry closed persona classes; Mia writes a different campaign per audience and records which angle she used; MarIA labels the objection she met and the strategy she answered with; results feed back into what Mia writes next. The fleet layer pools learnings across tenants behind a k-anonymity gate and currently **refuses everything at 3 real tenants — the control working, not a failure**. Automatic self-improvement is built and deliberately switched off, and §4.7's kill switch now has an hourly runner, which was the last engineering precondition for enabling it. Suite 1011 green, DEV only. 🔴 **The binding constraint is data, not code: nothing has sent, so the loop is correct and unfed.**)