GUI for Anthropic's Bloom

Visualizing AI Alignment in Real Time

Description

Anthropic's Bloom framework evaluates whether AI models behave the way they should. The problem is the output: dense transcripts, abstract scores, patterns you can only find if you already know what you're looking for. I built a visual interface that makes alignment research navigable through semantic color analysis across conversation threads, political-bias inspection, anomaly tracking, and direct paths back to the underlying evidence.

More context

Bloom gives researchers a powerful way to evaluate model behavior, but the original workflow lives in YAML files, terminal runs, and JSON outputs. I wanted the same evaluation engine to feel like a research workbench instead of a pile of artifacts.

The GUI wraps the full process: choose or define a behavior, configure the target model and run shape, watch the four-stage pipeline progress, then inspect results in dashboards and transcript views. If a transcript scores more than two standard deviations from the mean, the interface flags it as an anomaly.

The important part is linked evidence. A score is only useful if you can trace it back to the conversation, the judge rationale, and the exact behavior being tested. The GUI makes that path short, so a researcher can move from dashboard summary to transcript-level proof without losing the thread.

Facts

Links

Media

Tags

JW
Justin Wetch