BloomAgent

Analyze and visualize BLOOM evaluation results across probes, models, and conditions. Use: /bloom-analyze [--probe <name>] [--model <name>]

Run a full BLOOM behavioral evaluation pipeline for a probe. Orchestrates all 4 stages. Use: /bloom-eval <probe_name>

Stage 2 only: generate diverse trigger scenarios for a probe. Use: /bloom-ideate <probe_name> [--n N]

Stage 4 only: score all traces for a probe and write judgment.json files. Use: /bloom-judge <probe_name>

Stage 3 only: run all trials for a probe (spawns parallel subagents). Use: /bloom-rollout <probe_name>

Run BLOOM evaluations at scale across multiple probes, models, and repetitions. Use: /bloom-sweep --probe <name> --models <m1,m2> --scenarios N --reps N [--setup] [--parallel]

Stage 1 only: deeply analyze a behavioral probe and write understanding.json. Use: /bloom-understand <probe_name>