Agents are taking authority. The evidence layer has not caught up.

The regression nobody planned

For decades, regulated workflows ran on expert systems and deterministic policy engines: same inputs, same outputs, auditable by construction. That layer was replaced, often wholesale, by language-model agents whose serving stacks are not pinned, not replayable, and not contractually guaranteed to reproduce a past decision.

The model’s forward pass is arithmetic. The divergence everyone observes comes from the deployment: batch co-scheduling, kernel autotuning, silent engine upgrades, unpinned samplers, heterogeneous GPU fleets. Closed APIs document this honestly: seeds are best-effort or absent; fingerprints drift. “Probably the same answer” is not evidence when a mandate, a trade, or a supervisory review is on the line.

What regulators are already asking for

• EU AI Act (Art. 12–19): lifetime logging and traceability for high-risk AI. Outputs you cannot re-derive frustrate the obligation by design.
• Financial-services algo rules (RTS 6 class): reproducibility of signals and controls written for deterministic expert systems, now applied to generative components.
• FDA “locked algorithm” precedent: same input, same result. That is the regulatory definition of determinism in safety-critical automation.

These regimes do not require a particular vendor stack. They require that someone can re-run the decision against committed inputs. That is an infrastructure property, not a model property.

What we contributed in open source

The Agentic Determinism Index (ADI) is an open harness that asks every major hosted provider one narrow question: if I send you the same request N times, how identical are your answers?

• 1 September 2026: methodology and MIT-licensed harness publish: probe cases, provider adapters, scoring code, and a public comment window on the methodology. The repository link on this page goes live with that release.
• 6 October 2026: first reference scores and leaderboard publish, with raw transcripts; every score is recomputable from the published harness.

The methodology also publishes the exact serving conditions under which reproducible outputs are achievable; providers that meet them are recognized for it in the same table.