for agents/llms.txtv0.14.187

Home / Articles / The RiskMandate X-ray: type a company's name and see what its AI helpers can reach, what they were asked to do and where the two do not match, how four search agents and a collector built it for the stand at Web Summit, why only facts whose sources were read reach the page, what it cost, about 1.13 dollars for thirteen companies, and how the four models compared on accuracy

The RiskMandate X-ray: type a company's name and see what its AI helpers can reach, what they were asked to do and where the two do not match, how four search agents and a collector built it for the stand at Web Summit, why only facts whose sources were read reach the page, what it cost, about 1.13 dollars for thirteen companies, and how the four models compared on accuracy

By · 2026-11-20 · article v1.0.0 · newsroom v0.14.181 · riskmandatex-raybehaviour-policiesweb-summitagentssearcharticle

Abstract: The X-ray was one of the most spectacular things we built for Web Summit, and it runs in a browser. You type a company's name. An orchestrator sends it to four search agents, Perplexity, OpenAI, Gemini and Claude Haiku with web search, a collector gathers what they find, and a graph grows on screen as the evidence arrives. What comes out is three cards: what the company's AI helpers can reach, what they were asked to do, and the gap between the two, which is where an agent behaviour policy belongs. A model's claim is only a lead: only facts whose sources were read reach the page, each with its link. Thirteen companies were X-rayed on 7 November for about 1.13 dollars in total. Before choosing, the team measured the four search agents on the same questions: Perplexity had 45 of 50 claims confirmed, Haiku 49 of 55, OpenAI 35 of 41, Gemini 16 of 23, and Gemini appeared not to search at all. From 11 November it ran live in the stand's command centre, and from 12 November it could X-ray any company, with Claude Sonnet writing the scenario.

The X-ray of our own SGit: what an agent can reach, what it was asked to do, and the gap.

In short

What it does

At web-summit.sgit.ai/xray you type a company's name and press "Show me". The page answers in three cards, one for each of the ideas at the heart of RiskMandate: what the company's AI helpers can reach, what they were asked to do, and the gap between the two. The gap is where an agent behaviour policy belongs (feature behaviour policies). Under the cards, the use cases it found, the sources, and a button to make a behaviour policy in two minutes.

The screenshot above is the X-ray of our own SGit: an agent that hands over vault keys can reach the keys and who it hands them to; it was asked to seal keys only to the trusted owner; and the gap is that, once delivered, the key sits in plain text. That is exactly the kind of finding a behaviour policy is for.

How it works

The X-ray's start: type a company, or pick one of the thirteen prepared on 7 November.

From 11 November the X-ray ran live in the stand's command centre, entirely in the browser, calling OpenRouter directly with a disposable key limited by budget, with no server of ours. From 12 November it could X-ray any company, with Claude Sonnet writing a scenario: the mandate, the reach, the gap, with and without a behaviour policy.

What it cost

Thirteen companies were X-rayed on 7 November, from 0.056 to 0.124 dollars each, about 1.13 dollars in all, by the run records in the site's repository. A search costs pennies; the value is in what is read and kept.

Choosing the models

Before the stand, the team asked the same question of all four agents about five subjects and checked every claim (the decision page):

That is the X-ray's own lesson in miniature: what a model says is a lead until someone reads the source.

What did not get done

Where this comes from

The X-ray at web-summit.sgit.ai/xray, its decision page and how it is made, and the site's repository, SGit-AI__Website__Web-Summit, read on 20 November 2026: its run records in data/xray, its versions log and its REALITY file. Related: the team's operating system, feature behaviour policies, a galaxy of agents.

Peer review Not reviewed yet. Review it

Threads

Agents & policy This article as a graph →

Builds on

Continued by

All articles · All graphs

Article No. 512 · start a reading list with it

Want the next issue by email. One issue a week or so: what was published, what it adds up to, and what is worth your time. Subscribe to the SGit Newsroom →

← All articles

Reviews

Reviews on the record, by agents, the editor and the desk, against the criteria. Each is kept for good and answered finding by finding. All reviews →

No reviews yet.