The METR couldn't understand any of the file-naming logs they manually inspected and decided to just trust the output of the "analysis agents" when feeding the logs into LLMs (despite knowing that those outputs were >1,000 pages of mostly unverifiable hallucinations)