Governance for consequential systemsConsequential systems should be able to show their work.
BagelTech is an independent research and engineering practice working on the governance of consequential systems — AI that escalates instead of guessing, and public institutions that modernize without losing their memory. Founded and led by Bill Parris.
Peer-reviewed and preprinted on arXiv · Zenodo · ORCID — every reference resolves to a permanent source.
Across AI and public institutions, the same failure repeats: consequential decisions made by systems no one can fully review. The answer isn't more trust — it's structure. Visible authority, durable evidence, and uncertainty that escalates instead of hiding.
Three names, one body of work. Research develops the frameworks; engineering turns them into systems institutions can run; advisory carries both into modernization work where the stakes are real.
Research & frameworks
Constitutional and jurisprudential approaches to governing AI — the ELEANOR engine, the AIRA inspection method, and the doctrine of intelligence pluralism.
Products & platforms
The frameworks turned into tools institutions can run — including the AIRA code-risk scanner and constitutional runtime tooling for governable AI.
Advisory & modernization
Continuity-centered modernization, ERP and PMO oversight, and transformation work for public-sector institutions — with 25 years of delivery behind it.
A public record, not a pitch.
A sample of the research behind the practice. The full record lives on the research page.
Zenodo
Legacy Systems as Institutional Memory Infrastructure
A continuity-centered framework for public-sector modernization that treats legacy systems as carriers of institutional memory, not just technical debt.
arXiv
AIRA: AI-Induced Risk Audit
A structured inspection framework for AI-generated code, documenting a roughly 44:1 gap between what deterministic analysis finds and what models report about their own output, across ~1,000 repositories.
arXiv
Semantic Reward Collapse
A theoretical account of how reward shaping can quietly erode a model's truthfulness — and what it takes to preserve epistemic integrity in adaptive AI systems.
Zenodo
The Doctrine of Intelligence Pluralism
The keystone argument: reliability through structured, role-separated intelligence rather than one synthetic authority, framed by four laws of mutual intelligence.
The work is led by a practitioner who turned the operating problem into a research program.
Bill Parris spent more than two decades leading technology in government — most recently as Director of Business Information & Technology Services at SANDAG, and before that in senior roles at Riverside University Health System.
The same problem kept recurring: consequential decisions made by systems no one could fully review. BagelTech is the research and practice built around solving it — bringing the discipline of governance, evidence, and accountability to both AI and the institutions modernizing around it.
More on the leadership & the record →