Origin Part 15: The Wall Behind the Vocabulary

Short summary

The author discovered that Origin's vocabulary integration tool failed because its per-slot gates sampled only 120 random negatives—sufficient for local validation but insufficient for production robustness. Fixing domain assignment (brain→biology, anger→emotion) improved metrics, but full-encoder testing exposed the core structural issue: new slots fired top-1 on 15% of unrelated sentences. The key insight is that gate validation works for narrow same-domain growth but fails when concepts land in semantically isolated feature space without inherited discrimination.

• •Per-slot gates validated new vocabulary concepts against only 120 random negatives, missing false-positive patterns in production
• •Correcting domain assignment improved metrics but full testing revealed slots were firing on 15% of unrelated sentences
• •Root cause: gates work for same-domain growth but structurally fail for semantically isolated concepts lacking discrimination inheritance

Generated with AI, which can make mistakes.