What happened: how the Fidelidade X-ray was made

Step 1 of 6

1. The ask

On 7 November 2026 at 21:16 UTC, xray@web-summit.dinis typed one line into the search tool (admin/tools/xray-search.mjs): "Fidelidade, Portuguese insurer, its AI and agent pilots", as a company. It was not typed at the booth: that path is still being built (WS-194).

The search tool is the orchestrator. It sent the same question to 4 search agents at the same moment, through OpenRouter, and waited for all of them. Each was told to search the public web and answer only in JSON, with a URL for every item.

1. Perplexity search perplexity/sonar
2. OpenAI search openai/gpt-5.4-nano:online
3. Gemini search google/gemini-3.1-flash-lite:online
4. OpenRouter web search anthropic/claude-haiku-5.5

You are a research agent for RiskMandate X-ray. Subject (a company): "Fidelidade, Portuguese insurer, its AI and agent pilots".
Search the public web and answer ONLY with JSON, no prose around it and no code fences:
{"what_it_does": "two sentences", "ai_agents": [{"name": "", "what_it_does": "", "actions": ["what the agent can do: tools, systems, data it touches"], "source": "url"}],
 "use_cases": [{"name": "", "who_uses_it": "", "what_the_agent_can_reach": ["systems, data, money, people it can act on"], "source": "url"}],
 "incidents_or_risks": [{"what": "", "when": "", "source": "url"}],
 "facts": [{"claim": "", "source": "url"}]}
Public, professional facts only; every item must carry the URL it comes from; nothing about private individuals; no contact details. If you find nothing for a field, give an empty list.

Step 2 of 6

2. Each search agent's call

Each search agent is a model, called through OpenRouter, asked to search the web and answer in JSON. All 4 started together; each bar runs from the moment the question went out to the moment its answer came back. The whole fan-out took 25.1 s, the time of the slowest.

"Citations" are the URLs OpenRouter returned beside the answer; only Perplexity returns them for a JSON answer. "Named in its answer" are the URLs inside the JSON. Both are leads: nothing is taken from them until a source is read. "Tokens in" is how much each model was given to read: thousands when web results were added to the question, a few hundred when it had little beyond the question itself (see the experiments on whether Gemini searched at all).

Step 3 of 6

3. From claims to facts

What a model says is a lead. Only what we read on its source reaches the page.

The reading agent turned the 4 answers into 44 leads, each tagged with the model that said it, and looked for each one on a source it read. 42 confirmed, 2 not found, 0 unverified.

confirmednot foundunverified

• Confirmed: a source we read says it.
• Not found: we read the source, or looked for one, and it does not say it.
• Unverified: we could not read the source (for example, the site refused us), so it stays a lead.

What did not hold up, and why

Step 4 of 6

4. The reading agent

A reading agent: a Claude Code sub-agent of xray@web-summit.dinis, one per subject. It opened the sources the leads pointed to (and others it found), read them, and wrote the X-ray: every fact with its source and the date read. It finished at 7 November 2026 at 21:25 UTC.

The sources, by kind

Step 5 of 6

5. The review

xray@web-summit.dinis reviewed the X-ray on 2026-11-07 before it was published. What the review changed or checked:

1. Softened the tagline: the AI director's spoken remark is given as reported, not as a headline fact.
2. Removed two documented limits from the incidents: not events.

Step 6 of 6

6. What it cost

The search cost $0.0738 in all, as OpenRouter reported it for each call.

Perplexity search: $0.0063 (9%)OpenAI search: $0.0250 (34%)Gemini search: $0.0006 (1%)OpenRouter web search: $0.0419 (57%)

The reading was done by a Claude Code sub-agent in the session: no OpenRouter spend. Its cost is not recorded in dollars, only as about 174 thousand tokens.

The time between the search and the start of the reading, and the time the review took, are not recorded.