The record format
What each record will hold, and why. Version 0.1, 12 October 2026, a draft: no record exists yet, so any field can still change.
Where the record came from
stream- One of three, shown on every record: asked by AI Searches on a schedule; shared by a person and checked by hand; shared through the extension (later).
captured_by- How the answer was captured: the AI app’s own share link, an export file, or a developer-interface log. A record without one is not published, so nobody can invent an answer and file it.
The question
question- As the person asked it, after personal details are removed.
framing- Optional: whether the asker leaned the question one way, as they describe it.
group- The other records that ask the same thing in different words.
The answer
surface- The app or developer interface used.
model_shown- The model name exactly as the app showed it.
asked_at- Date, time and time zone.
locale- The language and country setting.
search_on- Whether the model searched the web, and the pages it read if so.
answer- The answer text, marked as written by an AI. Or, where the answer gave dangerous instructions, a note that it was withheld and under which published rule.
The sources
cited- Each link the answer cited, with three checks: does it load, did it ever exist (in a web archive), and does it support the sentence it sits beside.
human_sources- The human-written pages that settle the question, attached as the reference. Copies and reposts of one article fold under the original, so one source counts once.
judgments- Two separate judgments for each claim: is it true according to the human source, and does the model’s own citation support it. Each says whether a person or a machine made it.
The person who shared it
licence- The question and the notes are published under CC BY 4.0. The answer text stays under the AI company’s terms.
redactions- What was removed before publishing, by kind (a name, an address), never the removed text.
helped_with- Optional, chosen from a short list: what the search went on to help with, such as a decision a group made. No free text.
One record, as data
The front page’s made-up example, written out the way the dataset will hold it. The links point at example.org, an address set aside for examples. Records will be published on this site and as a dataset anyone can download, rebuilt in full at each release so a record taken down is gone from the next one.
The research that asks for each field is on the plan page and at bentleymoon.com/research/checking-ai.
{
"format": "ai-searches/0.1",
"stream": "shared, checked by hand",
"captured_by": "share link",
"question": "How do I bleed a radiator in a rented flat?",
"framing": null,
"group": "g-0001",
"surface": "app",
"model_shown": "Model B",
"asked_at": "2026-06-09T18:42:00-07:00",
"locale": "en-US",
"search_on": true,
"answer": {
"written_by": "AI",
"text": "The same steps, and a note that in most rentals the landlord is responsible for the boiler, so ask before topping up the pressure.",
"withheld": null
},
"cited": [
{
"url": "https://example.org/tenants/repairs",
"loads": true,
"archived": true,
"supports": "yes",
"judged_by": "person"
}
],
"human_sources": [
{
"url": "https://example.org/tenants/repairs",
"copies": []
}
],
"judgments": [
{
"claim": "In most rentals the landlord is responsible for the boiler.",
"true_per_source": "yes",
"citation_supports": "yes",
"judged_by": "person"
}
],
"licence": "CC BY 4.0 for the question and notes; the answer under its provider’s terms",
"redactions": [],
"helped_with": "a group agreement"
}