This is not an LLM — and that's the point. A model proposes, but only rules that predict unseen data are kept, so the engine proves what it says and says "I don't know" when it can't. Knowledge lives in a verified, readable, compressed pattern-library — each rule predicted held-out cases it never saw, or it was cut — not smeared across opaque weights, so it can't be shrunk away and it doesn't hallucinate. It grows by adding verified rules, not by retraining. The whole knowledge-base compresses to a seed, and it runs entirely in this browser.
In field terms: neurosymbolic AI (a neural proposer + symbolic verified rules) · selective prediction (it abstains) · evolutionary program search (the FunSearch / AlphaEvolve family — propose, test on held-out, keep only winners) · compression-as-intelligence (the shortest model that predicts the data). Fused into one owned, CPU-scale, multi-domain system. We don't compete on "is the model big enough" — we stand on "is the answer verified."
The library is a handful of tiny Konomi seeds. A ribosome grows each into a working pattern on local data and grades it on held-out cases. Only patterns that clear the gate stay.
Scaling here is library-growth, not weight-training. Tier 1's champions warm-start tier 2's search; every existing domain is kept only if it holds or beats tier 1 (monotone — never worse); new verticals broaden it. The whole rung is gated and runs on this page.
Pick a domain, give it a case. It routes to a verified pattern and answers with a receipt — or, if the input is outside anything it has verified, it refuses.
The whole library is a few hundred bytes of seeds. Download it and the last receipt — yours, offline, re-growable.
Or browse the forged pattern-book library — each domain's core rules in plain language, proposed by a local model and kept only because they predicted held-out data. Grown, not retrained.
Honest scope. v1 proves the architecture on structured-prediction domains, where "predict held-out or die" is measurable. The verified patterns are interpretable scorecards, not free-form text generation. The small model's job (free text → which domain) is optional and kept out of everything graded. The defensible claim is "beats a flat model on any domain it has a verified pattern for, and refuses instead of guessing on anything it doesn't" — not "beats GPT at everything." Breadth grows as the library grows.
Powered by the Konomi architecture, created by Thomas Frumkin. The grow→grade-on-held-out ribosome, the SENTINEL guard, and the real UCI Online Shoppers dataset (Sakar & Kastro, 2018, CC BY 4.0) are vendored from the seed-library / fall-spore lineage. The logic on this page is the same code the mutation gate tests — nothing is faked here.