Is Transect free to run?

Transect is open source under the MIT licence and free to download, but the judged scanners call an LLM provider, so those passes cost money. AISI ships a structural pass that needs no API key and makes no model calls, which extracts recorded events such as token use, context compactions and sub-agent launches at no cost.

What transcripts does Transect accept?

Transect reads Inspect AI .eval logs and OpenClaw .jsonl exports, and runs on Python 3.12 or newer. Transect is built on Inspect Scout, and its output is an HTML report, pandas dataframes and a Scout viewer, so a reviewer can read results interactively or analyse them in code. Bundled coding-agent skills can be installed separately.

Does Transect decide whether an agent did something unsafe?

Transect labels behaviour against a vocabulary the reviewer defines, rather than shipping a fixed safety verdict. AISI frames the tool as a way to keep transcript analysis reproducible and auditable, and links it to incident investigation and to research on evaluation cheating. AISI stresses that reliable conclusions still rest on evaluation design and expert human judgement.

INFOTool Research

Transect: UK AISI releases its agent transcript analysis tool

Transect turns million-token agent evaluation transcripts into labelled timelines a reviewer can audit and reopen without new model calls.

October 6, 2026 · UK AI Security Institute · Research by Toby D. Pilditch, Konstantinos Voudouris, Alexandra Abbas, Cozmin Ududec, UK AI Security Institute, Meridian Labs

Inspect AI

INFOTool

Transect is an open-source Python tool from the UK AI Security Institute and Meridian Labs that turns long agent evaluation transcripts into labelled, turn-indexed timeline reports a reviewer can audit.

01What happened

The UK AI Security Institute published Transect with a paper posted to arXiv on October 6, 2026 and an announcement the following day. AISI built the Python package in collaboration with Meridian Labs and released it on GitHub under the MIT licence. Toby D. Pilditch, Konstantinos Voudouris, Alexandra Abbas and Cozmin Ududec are named as its authors.

02How it works

Structural scanners read recorded events out of a transcript, including token use, context compactions and sub-agent launches, at no model cost. Judged scanners then use LLM judges to label behaviour against a vocabulary the reviewer defines in a spec file. Transect aligns both on a single turn-indexed timeline and stores the labels, so an analysis can be reopened without new model calls, and every label carries provenance and reliability information.

03Who is affected

Teams running long-horizon agent evaluations are the audience. Transect reads Inspect AI .eval logs and OpenClaw .jsonl exports, and needs Python 3.12 or newer. Judged passes need an API key for a model provider and incur model charges, while a structural-only pass runs without either.

04What to do

Evaluation and red teams can install Transect from its GitHub repository and start with the structural pass to map an existing transcript before paying for judged labels. AISI cautions that reliable conclusions still depend on evaluation design and expert human judgement.

Questions people ask

Is Transect free to run?: Transect is open source under the MIT licence and free to download, but the judged scanners call an LLM provider, so those passes cost money. AISI ships a structural pass that needs no API key and makes no model calls, which extracts recorded events such as token use, context compactions and sub-agent launches at no cost.
What transcripts does Transect accept?: Transect reads Inspect AI .eval logs and OpenClaw .jsonl exports, and runs on Python 3.12 or newer. Transect is built on Inspect Scout, and its output is an HTML report, pandas dataframes and a Scout viewer, so a reviewer can read results interactively or analyse them in code. Bundled coding-agent skills can be installed separately.
Does Transect decide whether an agent did something unsafe?: Transect labels behaviour against a vocabulary the reviewer defines, rather than shipping a fixed safety verdict. AISI frames the tool as a way to keep transcript analysis reproducible and auditable, and links it to incident investigation and to research on evaluation cheating. AISI stresses that reliable conclusions still rest on evaluation design and expert human judgement.

Sources 3 links, all fetched

1aisi.gov.ukAISI Announcement
2github.comTransect Repo
3arxiv.orgTransect Paper