Provable safety for autonomous agents.

SentinelOps develops open-source research projects in formal verification, runtime safety, and evidence for autonomous systems. Explore the work, its implementations, and the limits of the claims they support.

Technical information

Formal verification establishes specified properties under explicit assumptions. The evidence for a deployed system must also address the correspondence between its specification, its executable interfaces, and the conditions covered by the original result.

The programme includes work on proof tooling, execution controls, provenance, and reproducible evaluation. Each repository documents a different part of that research.

01 / Method

From policy text to an inspectable control.

Specification, proof, and enforcement form a proposed chain. Each transition needs its own implementation and supporting evidence.

Fig. 1
01

Specify

Interpret policy text as an explicit condition with declared actors and permissions.

Illustrative policy: Explicit consent authorizes agent access to user data.

Fig. 2
02

Prove

State and check a formal obligation with recorded assumptions and toolchain versions.

Schematic proposition: authorized(action) implies allowed(action).

Fig. 3
03

Enforce

Mediate the supported effects and preserve evidence of actual runtime decisions.

Design condition: admit an action only following the applicable check.

Proof over promises.

  1. I

    Specification-driven

    Behavioral obligations need precise specifications and a record of how policy statements are interpreted.

  2. II

    Runtime mediation

    Checks have a defined execution boundary. The governed effects, coverage, and bypass assumptions determine what the control establishes.

  3. III

    Audit by default

    Evidence records connect decisions to policy and implementation identity, making selected effects independently examinable.

Correspondence

Receive new writing.

Occasional dispatches on formal methods, runtime verification, and the infrastructure of trustworthy AI.