I come from engineering in regulated industries, where a model that almost gets it right is worthless. I bring that bias to AI: I do not care whether a detector looks good on a benchmark, I care about knowing exactly what it catches and what it does not.

Working on Groundlens I ran into an uncomfortable limit. Similarity-based hallucination detectors fall to chance exactly when the error matters most, when the false sentence uses the same words as the true one. I called it the register wall. Instead of hiding it, I retracted the numbers that effect had inflated in earlier work and published the blind spot as a figure.

Groundlens is the deterministic, explainable first stage of a pipeline: it filters cheaply what it can settle and escalates to entailment or a person what it cannot. I work alone and fund my own time. I am asking for support to keep measuring these limits rigorously, instead of selling them as solved.