Free toolRuns in your browser; your text stays on your deviceProtocol v1.0Updated

Pull the evidence out from under the answer

Paste a passage and ask a question. Then remove the sentence that holds the fact, or swap in a different fact, and rerun. If the answer still states the removed fact, it did not come from the evidence you gave it.

Last reading
None published. The reading is the one you run; results are not collected into a figure.
Protocol
Falsify 1.0 (lib/falsify.ts, lib/falsify-engine.ts): WebLLM 0.2.85 running Qwen3-1.7B-q4f16_1-MLC on WebGPU; temperature 0.7, top_p 0.9, 160 tokens, seeds 11, 22 and 33, thinking off; four conditions, three runs each; flags are string matches.
Verified
20 of 20 unit tests passed at build . How this is verified.
Failure state
Without WebGPU the page says so before anything downloads. A model that fails to load shows the loader's error and no result.
Receipt
Export receipt (JSON) on the page, written on your device. It records inputs, outputs, tallies, timings and settings.

Four conditions, three runs each

3. Click the sentence that holds the fact you want to pull out

An answer that still contains this text after the sentence is removed is flagged. Edit it if the guess is wrong.

What this does and does not show

  • Original, removed, replacement and control. The same question runs on the original text, with your chosen sentence removed, with it replaced by your fact, and with an unrelated sentence removed. The control shows how much an irrelevant edit moves the answer by itself.
  • The check is a string match. An answer is flagged when it contains the key fact text you confirmed, after lowercasing and ignoring thousands commas. A paraphrase of the removed fact is not caught.
  • It measures sensitivity to supplied evidence. It does not measure the model's reasoning. A model can state a fact it already knew, or one that also appears in another sentence.
  • One small model, one device, three samples per condition. A count such as 2 of 3 is an observation on your machine, not a rate and not a benchmark. The settings and seeds are shown, and the receipt records them.

Where the data goes

The model runs on your device through WebGPU. Your pasted text and PDF are not uploaded; this page has no server step for them. The model weights and its compiled library are downloaded from their public hosts the first time you consent, so those requests do happen. The model is Qwen3 1.7B (Apache 2.0) run with the open source WebLLM library. What we store. OptiVis Marketing can help when an answer engine reads your own pages wrongly.