Make an Evidence Atlas with AI
Explore six claims and four fictional sources. Remove a source, reveal a contradiction and download a review that keeps the evidence visible.
Experiment 02 / Follow the thread
A fluent paragraph can hide a fragile claim. The Evidence Atlas opens the paragraph up: six claims on one side, their source passages on the other. Everything concerns a fictional Night Library event. The relationships were written for this exercise; this is a working evidence interface, not an automatic fact checker.
Start with the claim that sounds plausible
Select the claim that the room can hold forty guests. An early email supports that proposal, but an approved plan explicitly replaces it with a twenty-four-seat layout. With both documents included, the interface shows a conflict. The current plan gives a specific reason to reject the old capacity, rather than asking you to trust whichever document has the newest date.
Now remove the approved plan. The status changes because the available evidence changed. The room did not become larger. This distinction is central to reading AI-assisted research: a claim can look supported inside a limited collection while being wrong in the wider world. A system should make that limitation visible instead of treating the absence of a challenge as confirmation.
Separate support, challenge and silence
Other claims teach different distinctions. Mara’s promise to bring lamps is definite. Jules’s offer to help depends on finishing another shift. Calling Jules confirmed removes a condition from the source. No document records a booking payment at all. That claim remains unestablished even though the poster looks like something ready to publish.
Each source can be included or excluded. The review labels describe only the included material. The explanatory note describes the authored example as a whole, while the source cards show exactly which relationships are currently active. Read the passage as well as the label. In a real research tool, you would also need provenance, document versions and a way to correct a mistaken relationship.
Build a source viewer before a verdict machine
To make your own version, begin with a small collection you are allowed to use. Assign stable identifiers to sources and claims. Keep the exact passage beside the relationship. Ask an assistant to build a viewer in which selecting a claim reveals its evidence, rather than producing a confidence number without an explanation. You should be able to inspect every connection in the first version.
If you later ask a model to propose relationships, preserve the distinction between a suggestion and a reviewed link. Give a person a way to accept, reject or revise each one. Test paraphrases that overstate certainty, missing conditions, older plans and documents that are silent on the question. A high average score would not excuse a single consequential claim whose support cannot be found.
A request you can adapt
Build a local claim-and-source viewer for these fictional records. Show supporting passages, contradicting passages and missing evidence separately. Keep relationships explicitly authored in this first version. Let me exclude a source and export the resulting review. Do not infer that a supported claim is independently true.
Keep the evidence attached to the summary
The download includes the selected source set, every claim’s current status and the included passages. It deliberately identifies the material as fictional and the relationships as authored. This makes the file useful outside the interface without pretending it is a finding about a real event. Try excluding all four sources and inspect how the report changes.
A useful next project might trace claims in a school assignment, compare versions of a public plan or organize a few approved research notes. Start with material whose sensitivity and permissions you understand. Keep interpretation next to evidence, and write down unresolved questions. The creative opportunity is an interface that helps people notice what a polished paragraph tends to hide.
Take this into your next task
Evidence is a relationship you should be able to inspect. Removing a source changes the review, not the underlying truth.
Questions people ask
Is this a live AI demo?
No. Astra in Codex helped create these original browser programs. Their controls run locally with authored rules; they do not call a model.
Can I keep what I make?
Yes. Download the current output, or explicitly save a world to your local creations shelf with its settings and camera view. The notebook lets you reopen worlds and export or import a library backup. Unsaved experiments last only while the page is open.
Sources and editorial notes
Reviewed 2026-09-10. Product capabilities depend on the app, plan, region and workspace settings. Workflows and prompts are authored teaching material, not recorded model results or measured time-saving claims.
Keep going
- Build an Agent Workbench You Can Inspect
- Turn Research into a Searchable Website with AI
- Build a Little World with AI
Related Power of AI pages
Keep reading with Start here, everyday uses, the tool guide, the writing workshop, and sources and standards.