Skip to main content

Scope boundaries

This is the most important page in this section. omnibias is powerful within a precise scope, and overstating that scope is the failure mode we most want to prevent.

What omnibias does claim

  • Closed-form σⁿ(z) for the smooth (Riccati) activation family, at any order, with one σ evaluation — bit-identical across backends.
  • Exact derivatives of one-layer fields and exact propagation of Taylor jets through compositions (Faà di Bruno), to float64 round-off.
  • Rigorous enclosures (interval / affine / Taylor-model) that provably contain the true value, outward-rounded.
  • Tamper-evident certificates, and a theorem_prover_verified flag earned only by a genuine Lean kernel pass.

What omnibias does not claim

Read this before quoting any result
  • It is not a PDE solver and does not prove existence, uniqueness, or regularity of true PDE solutions. It bounds the residual of a given trained surrogate over a given box.
  • It does not close the infinite analytic obligations of hard problems. The Yang–Mills mass-gap scaffold in formal/ keeps those obligations honestly sorry. Do not represent it otherwise.
  • "Exact" always carries a category. Numerical / grid-based pieces (fractional calculus, spectral methods) are not closed form and are labelled so.
  • A theorem_prover_verified = false flag means not proven by the kernel — it must not be reported as if it were proven.

The three exactness categories

Full detail: Exactness & scope.

A one-line honesty test

Before publishing a number, can you name its exactness category and point to the command (and, if rigorous, the certificate) that produced it?

If not, it is not ready to publish.

See also