This is independent research, done on my own initiative, into representing data as derivable geometry (dimensional programming, built on z = x · y as an organizing manifold) so a system computes what it needs instead of storing and re-sending it. I hold it to one rule: separate what is established from what is a defensible model from what is a conjecture under test, and label every claim. The bold ideas stay bold; the honesty keeps them credible.
A dependency-free library that represents data as nested points on the manifold and hands a model only the slice it needs. Measured result: a roughly 99.7% token reduction answering a localized question over a large nested structure. The Russian Doll and Car hello-worlds run live, with the full API reference.
Open the API and demos →A whole field of values computed live from one small definition, with no array of values anywhere on the page, plus the honest storage-versus-compute number and where it does not apply.
Open the demonstration →The same discipline reduced to a number: a deterministic verifier took unsupported model output from about 39% to 0% while best-of-N kept the answer rate high. Open source, reproducible.
Source on GitHub →Why fitting AI to geometry may help, the nested point as scale recursion, and the expansion-collapse rhythm of z = x · y. Each section tagged established, defensible model, or conjecture under test.
Read it →z = x · y² as a real surface, the helix behind a wave, and why a formula has a body. Offered as a conjecture and a lens, with the boundary to physics kept bright on purpose.
Read it →An honest note on scope. The measured demonstrations above are real and reproducible. The larger thesis, that representing data as geometry can reduce the drift and hallucination caused by paradigm mismatch, is a conjecture under test, not a settled result, and it is presented that way. The geometry is an organizing principle; it gives no raw compute speedup. The full theory, with claims labeled, lives at dimensionalprogramming.com.