Bloom ships the four-stage evaluation pipeline Anthropic ran against 16 frontier models

Bloom's pipeline runs in four named stages: Understanding, Ideation, Rollout and Judgment ^[5]. A seed configuration goes in at the front, the middle stages develop candidate evaluation ideas, Rollout executes them against target models, and Judgment scores the behavior that comes back ^[6]. The generation stages are what a static prompt checklist lacks, and they inherit whatever the seed says.

The report covers 16 frontier models and four behaviors: delusional sycophancy, instructed long-horizon sabotage, self-preservation and self-preferential bias ^[7]. Sixteen models by four behaviors is 64 model-behavior combinations ^[17]. If the behavior you care about is a support agent agreeing with a customer who has misread your refund policy, none of the four is it, and you write the seed configuration yourself. The repository documents seeds and configurations as part of ordinary usage ^[13].

For a published elicitation rate to say anything about your deployment, four things would have to match the report's run: the behavior definition, the configuration, the target model and the judging approach. The dev.to account draws that boundary explicitly, describing an evaluation as covering only the behavior, configuration, target model and judging approach it actually tests ^[10]. It also calls the reported measures evaluation metrics, not a general ranking of a model's safety or suitability for every business application ^[9].

Under MIT the code can be incorporated into internal tooling or adapted for a particular workflow, subject to the license terms ^{[4] [12]}. Installation from GitHub with pip is documented, along with the wiring needed to run evaluations with models ^[13].

The license is the cheap part. A meaningful run needs access to the target models, infrastructure to execute the runs, seed configurations that reflect risks the team actually faces, and a way to connect findings back into the development and review process it already has ^[14]. The published account does not state what a run costs ^[19].

Anthropic dated the announcement December 19, 2025, and linked the code repository alongside the full technical report ^[3]. As of 2026 that repository says Bloom has a new home and is developed and maintained by Meridian Labs ^[15]. The dev.to piece advises reading the repository's current documentation and maintenance information before making Bloom a long-term dependency ^[16]. Either way, the tests specific to a deployment stay in-house: its own prompts, tools, data access, escalation paths and customer-facing outputs ^[11].