"Trust me" is not an answer when the stakes are real. So these tools do one thing, cheaply and the same way every time: they tell you whether a machine's answer actually came from the source it was supposed to use.

They read the geometry of the answer, not a second model's opinion, so the clear cases pass in milliseconds and only the doubtful ones cost you a person or a heavier check.

• Want to check whether an answer actually came from the document you gave the model? That is groundlens. Deterministic scores in milliseconds, no second model in the loop, the same result every time. It is the first stage: it decides what your expensive check has to look at.

• Want that check running while you work, inside Claude Desktop, Cursor or Windsurf? groundlens-mcp prints a reading under each answer as it arrives. It is a filter, not a judge, and every reading says so.

• Have to show a supervisor that an answer followed a policy, with the evidence attached? groundlens-rules is hand-authored checklists that carry the text that triggered them. No model, no dependencies, and it never claims to be a measurement.

• Want to know what these detectors actually measure, and where they stop working? grounding-benchmark writes its false answers by hand rather than prompting a model for them, which is the case similarity-based detection finds hardest. Read the datasheet before quoting a number from it.

• Want to see the components working together on a real pipeline? The Cookbook has executable notebooks: what each check settles, and what it hands on.

• Want to know the research where these methods como from? Research

If you are checking generated output in a pipeline that matters, and paying for it in tokens or in people, the problem is interesting to us. Contact: javier@groundlens.dev.

Contributions are welcome across every repository. If you think a number here is wrong, open an issue with the reproduction — corrections get fixed and credited in the commit. Read CONTRIBUTING.md first.

Everything here is Apache 2.0. See LICENSE.

Groundlens is an independent open-source practice for trustworthy modeling, working where applied geometry meets machine learning. Maintained by Javier Marin · javier@groundlens.dev · groundlens.dev