Groundlens

Geometric methods for trustworthy models — verification over capability.

We turn "trust me" into "check me."

I built Groundlens because "trust me" is not an answer when the stakes are real. So it does one thing, cheaply and the same way every time: it tells you whether a machine's answer actually came from the source it was supposed to use.

It reads 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.

And it names, out loud, the cases it cannot see. Saying what it misses is not a footnote. It is the point.

A world where trusting a machine's answer stops being an act of faith.

Where under every answer sits a deterministic check you own — one you can run again two years later, show a supervisor, and get the same result. Not a black box grading another black box.

A verification floor that runs on everything you ship, so "did this come from the source?" becomes something you prove, not something you hope.

The deterministic first stage for RAG and agent loops. It ranks responses by how faithfully they reflect their sources — deterministic scores, sub-second, no second LLM in the loop — so the answers that earned trust pass and the rest go to review. It decides what your expensive check has to look at.

The same first-stage check inside Claude Desktop, Cursor, Windsurf, and any MCP client. It prints a reading under each answer — did this come from its source? — in milliseconds, with no model in the scoring path. It is a filter, not a judge, and every check says so.

Almost every hallucination benchmark writes its false answers by prompting an LLM. This one does not — a person writes them from memory, producing confabulations that stay inside the register of a correct answer. That is exactly where embedding-similarity detectors fall to chance. The evidence base for what these methods actually measure.

High-frequency monitoring of how a network's internal representations evolve during training, flagging discrete phase transitions. The same DNA as Groundlens: read the geometry of representations to see what a model is actually doing. Three lines to instrument any PyTorch model.

Groundlens is built on peer-reviewed research.

Contributions are welcome across every Groundlens repository. Read CONTRIBUTING.md before opening an issue or a pull request.

All Groundlens open-source projects are released under 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