Your platform lead wants to delete the four probes and...

The proposal trades off two critical components: the self-aware field’s validity and the cost profile. A model’s verbalized confidence is distinct from its internal activation, making the self-aware field better suited for a model-internal probe. Furthermore, replacing lightweight probes with an LLM call for every query significantly increases compute costs, making the gate more expensive than the retrieval it is intended to avoid. The best approach is a hybrid: use an LLM judge for intent and knowledge, but retain a trained probe for self-awareness, validated against held-out labels.