Skip to main content

Responsible AI & governance

omnibias is scientific-computing infrastructure, so "responsible AI" here means something concrete and technical: can you trust, reproduce, and bound the numbers it produces? This section is about scientific-model governance, not generic LLM content safety.

The five pillars

PillarPage
Governance process for scientific modelsModel governance
Where numbers come from, sealedProvenance & certificates
Same inputs → same bits, anywhereReproducibility
What must pass before releaseEvaluation gates
What we do and do not claimScope boundaries

The core principle: honesty about exactness

Every result omnibias produces is labelled by how it is exact — closed-form, autodiff-exact, or numerically approximate. Conflating these is the central governance failure for scientific ML, and omnibias is designed to make it impossible to do so by accident. See Exactness & scope.

Who this section is for

  • Researchers publishing results that depend on omnibias numbers.
  • Reviewers / auditors who need to re-derive or bound a claim.
  • Engineering leads putting a scientific model into a regulated or safety-relevant pipeline.