Understand the answer.
Inspect the evidence.
EvidencePilot helps you research verifiable claims without hiding how its judgments were formed. Here's what happens between your input and your report.
Try the analyzer ↗Identify what can be checked
EvidencePilot asks NVIDIA Nemotron, served through Nebius Token Factory, to extract up to three concrete, independently verifiable factual claims from your text. Opinions and predictions are not treated as facts.
Find relevant context
Tavily retrieves live evidence for each claim. Deep mode adds a Nemotron research plan and an optional targeted follow-up search when initial evidence is weak. A search excerpt is a lead, not independent proof.
Compare the available evidence
Nemotron assesses those excerpts and returns a supported, contradicted, mixed, or insufficient-evidence verdict. Linked URLs are drawn from actual search results, not generated by the model.
Write more carefully
A final pass suggests a cautious rewrite that avoids overstatement. You can inspect source relationships, compare the original with the corrected text, and export Markdown or JSON.
One product. Two useful workflows.
Both modes perform live research and label uncertainty. Choose the lighter workflow for routine claims or the fuller investigation when a claim deserves additional scrutiny.
Check the essentials.
Uses Nemotron to extract and assess claims, with one basic Tavily search per claim. Skips separate planning and follow-up calls to reduce usage and expected latency.
Best for: straightforward factual claims and quick editorial checksExplore beyond the first answer.
Uses Nemotron to plan research, advanced Tavily searches, and a bounded follow-up when initial results are weak. Returns more investigation details when available.
Best for: nuanced claims and research that benefits from a stronger source trailDeep mode uses more API requests and may take longer. More requests do not guarantee a correct verdict. Neither mode independently authenticates retrieved sources.
Traceable sources, not automatic truth.
EvidencePilot validates that displayed links came from retrieved results, but it does not authenticate the websites or verify that an excerpt correctly represents the complete source. Be especially careful with scientific, political, medical, legal, or financial claims.
Questions worth asking.
Does a 'supported' verdict guarantee a claim is true?
No. It means the model found support within the retrieved material. Results may be incomplete, outdated, biased, or misinterpreted. Open the original sources and corroborate important conclusions.
Why do I see 'insufficient evidence'?
The available excerpts might not address the claim directly, or may not provide enough reliable detail to make a defensible judgment. Uncertainty is preferable to inventing a confident verdict.
Can I use this for medical, legal, or financial advice?
You can use it to organize preliminary research, but never as the sole basis for a consequential decision. Consult qualified professionals and primary sources for high-stakes questions.
Will loading another example or clearing the editor erase my report?
No. Your last completed report, source-review notes and editorial draft remain visible below, clearly labelled as belonging to the previously analyzed text. Use Restore analyzed text to return to the exact input without another AI request. Running a new analysis or reopening a different saved JSON report replaces the previous report, so export JSON first if you want to keep it.
Can I reopen a report without another AI request?
Yes. Export your investigation as EvidencePilot JSON, then use Reopen saved report on the analyzer. The file is read locally without contacting Nemotron or Tavily. Old verdicts are not rechecked, and imported files are labelled as saved snapshots.
Can I compare two investigations?
Yes. With a report open, use Compare two evidence reports to load another saved EvidencePilot JSON export in your browser. The comparison reports changes in verdicts, evidence strength and distinct cited pages for near-identical claims, ignoring tracking-only URL parameters. It is not a new fact-check, and text with different wording is not matched automatically.
Can I edit the AI's suggested correction?
Yes. The Human Editorial Desk lets you choose a personal editing decision for each claim and rewrite the proposed text locally. Your edits are not reassessed by Nemotron, and they do not change the original AI verdicts. Export JSON to preserve your decisions and working draft.
Does this page need an account?
No account is required. The public demo uses server-side AI credentials and rate limiting, and analysis may be unavailable if its usage limits are reached.