How an X-ray is made
An X-ray is made by agents, in nine steps. A model's answer is only ever a lead; a fact reaches the page only once a source has been read. Below: the pipeline, then what happened across all 13 X-rays made so far.
- A name is typed: One line: a company, a use case or an industry.
- The orchestrator: The search tool sends the same question to four search agents at once.
- Four search agents: Each model, called through OpenRouter, is asked to search the web and answer in JSON, with a URL for every item.
- Leads: Every claim in the answers is a lead, tagged with the model that said it. Not a fact yet.
- The reading agent: A Claude Code sub-agent opens the sources and reads them.
- Sources read: Each with its URL and the date read.
- The verified record: xray.json: only what a source we read says, each lead marked confirmed, not found or unverified.
- The review: xray@web-summit.dinis checks the record and changes what does not hold.
- Page, deck and run log: The X-ray page, the deck for the stand, and the run page that shows how it was made.
Across every X-ray
13X-rays
52search agent calls
$1.1316search cost in all
830,877 / 102,696search tokens in / out
733leads checked
88%confirmed
366sources read
489reading tool calls
82 minreading time in all
2341kreading tokens
7 minsearch to reading finished (median)
Each model, against what we read
The real result of the experiment: every lead each model gave, across all 13 X-rays, and what reading the sources found. Each bar is one model's leads, split by what happened to them.
confirmednot foundunverified
| Agent (model) | Runs | Leads | Confirmed | Not found | Unverified | Cost in all | Average time | Tokens in / out | URLs given per run |
|---|---|---|---|---|---|---|---|---|---|
Perplexity searchperplexity/sonar | 13 | 204 | 185 (91%) | 9 (4%) | 10 (5%) | $0.0812 | 7.5 s | 2,778 / 14,300 | 15.5 |
OpenAI searchopenai/gpt-5.4-nano:online | 13 | 156 | 138 (88%) | 12 (8%) | 6 (4%) | $0.3644 | 14.4 s | 191,807 / 20,419 | 5.6 |
Gemini searchgoogle/gemini-3.1-flash-lite:online | 13 | 94 | 67 (71%) | 22 (23%) | 5 (5%) | $0.2773 | 3.3 s | 3,009 / 8,866 | 3.2 |
OpenRouter web searchanthropic/claude-haiku-5.5 | 13 | 279 | 253 (91%) | 9 (3%) | 17 (6%) | $0.4088 | 21.9 s | 633,283 / 59,111 | 9.8 |
Sources read, by kind
Each X-ray, and how it was made
| Subject | Searched | Search cost | Leads | Confirmed | Not found | Unverified | Sources | To read |
|---|---|---|---|---|---|---|---|---|
| Amex GBT | 7 November 2026 at 21:16 UTC | $0.0568 | 52 | 45 | 3 | 4 | 19 | 6 min |
| Atlassian | 7 November 2026 at 19:08 UTC | $0.1243 | 68 | 64 | 2 | 2 | 39 | 9 min |
| Booking.com | 7 November 2026 at 21:16 UTC | $0.1243 | 48 | 30 | 1 | 17 | 19 | 9 min |
| Cloudflare | 7 November 2026 at 19:08 UTC | $0.0679 | 71 | 63 | 8 | 0 | 42 | 6 min |
| Feedzai | 7 November 2026 at 21:16 UTC | $0.0929 | 54 | 48 | 5 | 1 | 28 | 7 min |
| Fidelidade | 7 November 2026 at 21:16 UTC | $0.0738 | 44 | 42 | 2 | 0 | 19 | 8 min |
| Klarna | 7 November 2026 at 19:08 UTC | $0.0563 | 39 | 34 | 3 | 2 | 21 | 5 min |
| Manus AI | 7 November 2026 at 19:08 UTC | $0.1201 | 71 | 67 | 4 | 0 | 43 | 7 min |
| Primitive | 7 November 2026 at 19:08 UTC | $0.1193 | 56 | 54 | 1 | 1 | 25 | 5 min |
| Replit | 7 November 2026 at 18:30 UTC | $0.0558 | 49 | 37 | 8 | 4 | 39 | 17 min |
| RiskMandate | 7 November 2026 at 18:30 UTC | $0.0840 | 63 | 55 | 8 | 0 | 22 | 15 min |
| Salesforce | 7 November 2026 at 19:08 UTC | $0.0691 | 61 | 51 | 3 | 7 | 23 | 7 min |
| SGit | 7 November 2026 at 18:30 UTC | $0.0870 | 57 | 53 | 4 | 0 | 27 | 16 min |
What this cannot do yet
- A name typed at the booth is not run live yet. The request path is being built (WS-194). Every X-ray here was run by xray@web-summit.dinis before the event.
- Each search is one call per agent, once. Run again on another day, the same models may give other leads.
- The reading agent's cost is not recorded in dollars, only in tokens. The search cost is what OpenRouter reported.
- The mandates, and the simulation steps placed against them, are illustrative: our judgement of what a deployer would plausibly authorise, not the company's own policy.
- A site that refuses a reading agent leaves its leads unverified. They are listed, not dropped.
- The review is by one agent, xray@web-summit.dinis.
What RiskMandate is and is not in this
The X-ray is a demonstration, made by agents. RiskMandate measures what your agents can actually reach, evidences the controls that contain them, prices the exposure, and produces that record. It is never in the request path: read-only by design (riskmandate.ai/llms.txt). Nothing on this page claims more for it.