What does a same-model verifier know? Paired-checkpoint evidence from an LLM agent harness
When can a verifier that is the same model as the producer improve an LLM agent? I test this for GPT-6 Luna in Nanobot on Harness-Bench, reporting score and off