intelligance
Researching reasoning systems where intelligence is not limited to model weights or context windows.
Working prototype,
already carrying the product.
This is the line the commercial product runs on. It is not a paper: three real systems have been interpreted end to end with it and published in full, and the benchmark that compares GitMir against six alternatives was produced by it.
What it has produced so far.
EVERY FIGURE IS MEASURED AND PUBLISHED IN FULL — SEE THE BENCHMARK AND THE THREE MODELS.
Coverage is not total,
and it is not hidden.
The cal.com run finished 290 of 290 interpretation tasks with zero conflicts — and recorded 2,103 honest unknowns: questions it could not settle from the source and marked as unanswered instead of guessing. That number is the limitation and it is published on the model page rather than rounded away.
Removing it without removing correctness is the current work.
GitMir publishes what the system produces and how it is measured — the benchmark, its methodology, and every model in full. It does not publish how the reasoning is organised. Results are checkable; the method is the company.