Audited 2,275 learning artifacts across 14 rigs spanning Nov 2025 to
Apr 2026. Key findings: 0% citation rate (learnings written but never
retrieved), 100% auto-extraction garbage in primary rig, inconsistent
metadata across fleet, 1,612 uncurated files in global store. Intra-
session flywheel works (80% pre-mortem accuracy). Cross-session flywheel
does not compound. Proposes 6 ordered fixes with measurable success
criteria.
Boden Fuller