01 · simple is the goal
small enough to fit in your head.
if the whole system can't be explained on a single page, it is hiding something. we build the smallest tool that does the job, and we show the wiring.
reasoning is the model talking itself through a problem. intuition is the model knowing the answer before it has finished talking. we measure that moment, and we put it to work.
intelligence emerges from coherence — not scale.
three claims. first, you can make a model arrive at the right answer faster by giving it the right scaffolds — not by making it bigger. second, every model already knows when it knows; you just have to read the number sitting in its own output. third, the way to build trustworthy AI is to publish receipts — every claim a measurement someone can re-run on their own machine.
each claim falsifiable. each load-bearing.
if the whole system can't be explained on a single page, it is hiding something. we build the smallest tool that does the job, and we show the wiring.
the model runs on your machine. no cloud round-trip, no data leaving the room. one workstation, one box, one explicit cost in watts and seconds.
every result on this site comes with the prompt, the model, the wall-clock seconds, and the file the test wrote. if you can't reproduce it, we didn't ship it.
negative results are how research moves. we tag notes measured, mixed, or projected. mixed is honest. so is rolling back something that didn't beat its own baseline.
we treat information the way physicists treat matter.
the same math that explains how waves order a crystal lattice — coupling, resonance, hidden phases — turns out to apply to information too. we use it to find structure inside the model's own outputs that nobody else was reading. every claim below is a paper with a number, not a hand-wave.