Origin Part 20: Gradient Pollution

Short summary

Part 20 of the Origin series details debugging a vocabulary expansion tool that worked for 2 concepts but failed at 14, with each mitigation shifting breakage rather than fixing it. The author discovered three stacked architectural problems: many 'dead' encoder concepts were never needed by the brain layer, making them unrescuable nominal slots rather than trainable concepts. An audit joining encoder health reports against reasoning data revealed 5,430 of 5,493 dead concepts had zero pressure from the brain layer.

• •Vocabulary topup tool worked for 2 concepts but failed at 14 with cascading battery breakage
• •Each mitigation moved breakage around, indicating a structural problem not a tuning issue
• •Audit revealed 5,430 dead concepts had zero brain-layer pressure — unrescuable nominal slots

Generated with AI, which can make mistakes.