Products · Agentic compliance
Agents are taking authority. The evidence layer has not caught up.
Agentic compliance is how AI gets deployed safely in regulated and mission-critical systems: reconstruct the decision, verify the authority that permitted it, and replay under a pinned stack. It is the expansion application on Lemma’s policy-to-proof platform, not the first commercial entry. Regulators and counterparties will ask: show me exactly what ran, and let me run it again. On today’s hosted inference stacks, that answer is usually unavailable.
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.
Two primitives
| Primitive | Required evidence | Boundary |
|---|---|---|
| Deterministic execution | Exact request and tool context, pinned model and serving stack, output receipt and replay result | Reproducibility is specific to the execution conditions. It does not establish correctness or safety. |
| Provable mandate | Principal, scope, limits, expiry and revocation state bound to the action | Authorization checks must occur before side effects; a signed instruction alone is insufficient. |
The next application test is a bounded workflow (for example a treasury agent): record the decision, evaluate its mandate, and show that a replay or authorization failure prevents the action and produces reviewable evidence. Production replay and mandate assurance remain development work.
What we contributed in open source
The Agentic Determinism Index (ADI) is the public front door: 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? ADI measures serving tuples. It does not certify compliance or prove an agent’s full decision process.
- Published 1 September 2026: methodology and MIT-licensed harness, with bootstrap reference scores on the live index: probe cases, provider adapters, scoring code, raw transcripts. The methodology stays open for comment; community replications are welcome.
- After about a month of runs: first analysis of the accumulating reference series and open-source contributions; the leaderboard keeps updating as new runs land.
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.
Agentic Determinism Index (ADI)
Open methodology · reproducible scores · serving-tuple measurement, not model marketing.