Transect helps researchers inspect long AI runs

When an artificial intelligence (AI) system spends days writing code or running experiments, a final score cannot show everything it did along the way. Britain’s AI Security Institute has released a tool to help researchers look more closely.

Announced on 7 October, Transect turns the record of an AI’s work into an interactive timeline. Reviewers supply that record and categories such as coding or data analysis. AI models label the activities, with links back to the original passages so people can check the interpretation.

The team’s accompanying preprint demonstrates the approach on an AI research evaluation containing almost 13 million tokens — the pieces of text processed by language models. The timeline helps expose how work is distributed, including tasks delegated to other AI agents.

The labels can be wrong, and the demonstration does not establish that Transect prevents harm. Human judgement remains essential.

By releasing the software openly, the institute is giving other researchers a way to inspect lengthy AI runs, challenge automated conclusions and improve the tools together.

Sources

• AI Security Institute — Transect: Making large-scale agentic evaluations easier to understand
• arXiv — Transect: Retaining Observability for Long-Horizon LLM Agent Evaluations (preprint)