my best guess at why it perfectly remembers this document is not RL (famously only ~1 bit of information per trajectory), but context distillation (probably during post training), one of the earliest mentions of which is an early anthropic paper
quotes: 1994973338448662858
same thread: 1995340506138874123
cited on: observations
Reproduced against link rot, credited and linked to its original. Yours and you’d rather it weren’t here? Open an issue.