Bloom
Explain how Bloom auto-generates evaluation scenarios around a single behaviour: the four-stage scoring pipeline, the seed mechanism for reproducibility, and how it divides work with Petri.
An open-source framework from Anthropic that automates behavioural evaluation by expanding scenarios around one behaviour and quantifying its frequency and intensity. Bloom runs a four-stage pipeline of understanding, ideation, execution and judgement, then has a judge model score each rollout. Recorded seeds keep the experiments reproducible and comparable across models.
