Trustworthy AI governance requires shared knowledge, reproducibility, and community scrutiny. Overdog publishes methodology openly to establish shared foundations and prior art -- not to create proprietary lock-in. We publish the measurement science progressively and openly. The production-grade execution is where we build a business.
This page defines the publication policy. Position pieces, research notes, and white papers are collected on the Content hub.
Formal working papers and governance papers, collected in the White Papers section of Content, starting with "Safety Is Retrospective"
High-level framework descriptions, architecture diagrams, and the regulatory join key specification
The epistemic status of every claim -- labelled as theorem, definition, hypothesis, or open question
As methodology matures and pilots conclude, we will publish:
arXiv preprints of methodological extensions, beginning with the Atom 0 paper on conformal e-prediction for agentic monitoring
Educational resources, explainers, and reference diagrams
Non-sensitive tooling -- trace canonicalisation utilities, replay verifier prototypes, Standard Candle test harnesses -- under open licences once validated
This staged approach mirrors mature safety-critical domains. In aviation, the principles of flight data recording are public and standardised; the certified recorder implementations are controlled. In nuclear safety, the physics of neutron flux monitoring is published; the reactor instrumentation is engineered under strict regulatory oversight. In clinical genomics, the statistical methods for variant classification are published and debated openly; the validated diagnostic pipelines that implement them in clinical laboratories operate under formal accreditation.
The methods are public. The high-stakes implementation is controlled. That is not a contradiction. It is how serious measurement infrastructure works.
We publish core methodology as defensive prior art. The intent is to prevent enclosure of foundational measurement concepts -- not to establish exclusivity. If conformal prediction, e-values, and sequential testing become standard tools for AI governance, that outcome serves the mission regardless of who implements them.
Our competitive position rests on execution quality, deployment experience, and the operational corpus that accumulates through commissioned work -- not on restricting access to the underlying science.
We welcome collaboration with researchers in conformal prediction, sequential testing, distribution-free inference, AI governance, and measurement science. If you are working on related problems -- particularly in the application of calibrated uncertainty quantification to runtime AI monitoring -- we would like to hear from you.
We are also engaging with standards bodies (BSI, CEN-CENELEC JTC 21) and measurement institutions (NPL Centre for AI Measurement) to ensure the methodology develops in dialogue with the communities that will use it.