The proof, one receipt at a time.

The same model, 76% faster, same chip

No new chip, no new model. Just a leaner 4-bit mode the chip runs in hardware, and the same model typed 76% faster on a fraction of the memory. Rerun it yourself.

One number you can actually check

For math, the right shape of answer is a number you can verify, not prose. Kepler returns the number, a checker compares it to the known answer, and on a locked test it passed 38 of 44. Rerun it yourself.

A $1 test that guards $1,679

Before booking a big cloud training run, we test the design on the desk first. The cheap test catches the bad designs, and one wrong booking it stops pays for the whole box. Rerun it yourself.

An AI lab that ran overnight for 2 cents

Fifty experiments, 73 minutes, about two cents of power, and nothing sent out. The same loop on rented cloud chips runs to dollars plus a fee for every word. Rerun it yourself.

Reasoning you can read step by step

A right answer with a hidden path is hard to trust on the next question. This model shows each step, and its score climbs in a clear, checkable way as you hand it better source. Rerun it yourself.

A 4B model on your desk out-trusts frontier cloud

Same 21 hard questions, three models. The small one you can own beat the big ones you rent, at zero cost and full privacy. Rerun it yourself.

Answers that name their exact source, or refuse

A good answer you cannot trace is still a guess. This receipt foregrounds one gate of the test: name the exact source, or refuse. Same run, same 21 questions.

A model that refuses when its notes do not hold the answer

Honest scope: this is the refuse gate of the test, the model checking what it can answer from what it was given. It is not the model grading how good its own memory is. That gate does not exist here, so we do not claim it.