groundlens

Check whether an answer was actually drawn from the material it was given.

This checks grounding, not truth. It tells you whether the answer engaged its source. It does not tell you whether the facts are right, and a plausible wrong fact stated in the right frame will pass. That case has to go to a second stage, so the demo runs one for you: groundlens (geometry, milliseconds, deterministic, free) and Vectara HHEM (entailment, slower, and the one that holds up when the error stays in register). Watch where they part company.

GitHub · Docs · PyPI · SGI paper · Taxonomy · Mechanistic paper

If you have no source document, this falls back to DGI, which compares your answer against a reference direction learned from 212 answers written in a single style by a single model. Text written any other way scores low however faithful it is, so a low DGI means "not comparable to that corpus", not "wrong". Paste a source whenever you have one — the check becomes SGI and means much more. To use DGI on your own material, fit it to your own data with groundlens.calibrate().

1. What did you ask the AI?

2. Did you give the AI any source material? (optional)

If you gave the AI a document, a webpage, an Excel file, or any reference material to base its answer on, paste the text below. If you just asked a question with no source, skip this step.

3. What did the AI answer?

Try an example

The third example is the blind spot, on purpose: every word is copied from the source except one number. Geometry passes it. Entailment catches it.

The last two have no source, so they run DGI. Both will read low — that is the calibration limit above, not a verdict on the first one's accuracy.

groundlens is open source (Apache 2.0). Built by Javier Marin. This demo runs the same library available via pip install groundlens.
groundlens is triage, not a truth oracle. It tells you which answers came from the source and which need a closer look. It does not check facts.