What Certanum has published, and the Field Record: a dated record of developments in AI numerical evidence. Every entry links to its primary source.
Not company news. Each entry gives the date and primary source, what happened in one neutral sentence, and one sentence of Certanum’s view.
NVIDIA launches the Open Agent Safety Platform, with OpenShell, an open-source runtime that governs what AI agents can access and do. NVIDIA names Citi and JPMorganChase among the financial-services firms collaborating on its agent-safety technologies.
For numerical evidenceRuntime control governs the action; whether a consequential figure was computed from its sources remains a separate question.
Reuters reports that Anthropic has set up a biology wet lab in the San Francisco Bay Area as it expands its life-sciences work. Anthropic says the lab’s main focus is fundamental biology rather than drug discovery; its head of life sciences told Reuters that the final test in biology is still real lab work.
For numerical evidenceThe same holds for numbers: the test of a computed figure is its source data and its re-execution, not the model’s account of it.
Novo Nordisk and Anthropic announce a collaboration on drug-discovery problems identified by Novo’s scientists, including a trial in which Novo tests Claude Science on specific research and development workflows.
For numerical evidenceWorkflow-level trials are where the standard of evidence for an AI-computed figure is first set in practice.
The FCA publishes the final report of the Mills Review on AI in retail financial services. It expects firms to be able to reconstruct and trace how AI-assisted outcomes were reached.
For numerical evidenceFor a consequential figure, reconstructing the basis means its operands, its operation and their sources, recorded when the figure was produced.
Bristol Myers Squibb announces a strategic agreement with Anthropic to deploy Claude Enterprise across research, clinical development, manufacturing, commercial and corporate functions.
For numerical evidenceOnce AI works inside clinical development and manufacturing, the figures it produces enter regulated records, where each must be traceable to its source.
The Federal Reserve, OCC and FDIC issue revised model-risk management guidance (SR 26-2, OCC Bulletin 2026-13, FDIC FIL-15-2026). Generative and agentic AI are placed outside its scope, to be governed under firms’ broader risk management.
For numerical evidenceWith these systems outside model-risk guidance, each firm must decide for itself how a figure they produce is evidenced.
The FDA and EMA publish guiding principles of good AI practice in drug development. Principle 6 asks that data provenance, processing steps and analytical decisions be documented in a traceable and verifiable way, in line with GxP.
For numerical evidenceA figure an AI system derives from study data is one such processing step; its operands, operation and sources are what make it traceable and verifiable.
Proof-Carrying Numbers, by Aivin V. Solatorio of the World Bank’s Development Data Group, proposes verifying the numbers in a model’s output against structured claims at the presentation layer, failing closed. The paper leaves derived values, such as ratios and aggregates, to future work.
For numerical evidenceThe closest published precedent. AP-1 and Certanum address the derived figure: its operands, its operation and their provenance.
We correct errors in this record in public, with the date and what changed. If you find one, tell us: contact@certanum.com.
Certanum is not affiliated with, endorsed by, or connected to any company, authority or author named on this page. Descriptions are paraphrased and are not legal advice.