From formal claims to controlled execution.
SentinelOps investigates the conditions under which specified properties of autonomous software have meaningful consequences at runtime.
The work connects precise specifications, machine-checkable arguments, mediation of consequential effects, and records suitable for independent review. Each connection introduces assumptions that need to be made explicit.
A proof has a domain of validity.
Empirical testing measures observed behaviour under selected conditions. Formal verification establishes a defined property relative to a model and its assumptions. Both contribute evidence, with distinct failure modes and limits of inference.
An implementation introduces further questions. The executable system must reflect the model, the relevant actions must pass through the specified control, and changes to dependencies or environmental conditions must be evaluated against the original claim. A theorem about an abstract transition system does not establish the safety of every deployed instance.
Verification identifies what follows from explicit assumptions. Runtime evaluation identifies the paths and effects actually governed by the implementation. An independent conclusion depends on both records.
The interface is the research problem.
For tool-using agents, critical behavior occurs at interfaces with external systems. A policy expressed in natural language needs a precise interpretation. A proved obligation needs an implementation that enforces it. Evidence about an execution needs enough provenance to distinguish a checked action from an unobserved path.
Four connected research tasks.
A proposed system must account for each interface before a broad end-to-end claim has adequate support.
- 01
Specify
State the property and its assumptions.
Identify the relevant actors, effects, authorizations, environmental conditions, and failure cases. Ambiguous policy text is a starting point for formalization, not a theorem.
- 02
Verify
Check the stated proposition.
Use a proof assistant, model checker, or appropriately scoped test to evaluate an explicit obligation. Record tool versions, dependencies, and the trusted components of the result.
- 03
Mediate
Connect the specification to execution.
Place a control at the relevant action boundary. Its scope, coverage, bypass assumptions, and response to failed checks define the enforcement claim.
- 04
Inspect
Make the evidence independently examinable.
Retain records needed to reconstruct a decision and distinguish an implemented control from a proposed one. Versioned code and reproducible tests supply the supporting evidence.
Claims proportionate to their evidence.
- I
Safety first
Safety properties need precise statements, justified assumptions, and scrutiny of the execution paths to which a verified property applies.
- II
Rigour
Formal verification and proof-carrying code provide evidence for explicit specifications and proof obligations, within their stated scope.
- III
Open collaboration
Publishing source and evidence enables external examination. Reproducible checks and clear limitations make that examination meaningful.
- IV
Practical innovation
Connect academic formal methods to actual interfaces, deployment constraints, and the engineering evidence needed to assess their use.
Testing and formal verification.
The original site compared empirical test coverage with adoption of formal verification. Its numerical estimates lacked cited measurement methods. This comparison preserves the subject without presenting invented percentages.
Empirical testing
Examines observed outcomes under selected inputs and environments.
Evidence to inspect: sampling, coverage, baselines, instrumentation, and failure records.
Formal verification
Establishes a precisely stated property under a model and explicit assumptions.
Evidence to inspect: formal statement, dependencies, trusted checker, and implementation correspondence.
Programme milestones and ambitions.
The dated programme sequence is retained from the original website as an editorial roadmap. A milestone's inclusion does not establish that a project, pilot, or standard was completed on that date.
- 2024
Foundation
The original site described the establishment of an open-source organisation and a core verification toolchain.
- 2025
Industry adoption
The previous roadmap named finance, healthcare, and autonomous-systems pilots. Confirm individual pilot records before treating these as completed deployments.
- 2026
Standard setting
The original objective concerned interoperable specifications and proof-exchange standards. Contributions must be identified through the relevant artifacts.
- 2027+
Ubiquitous verification
The long-range ambition is wider formal verification in consequential applications, not a forecast or a claim of universal adoption.
Research is collective work.
Implementations, open issues, release records, and research notes provide different kinds of evidence. Review the project repository for the versioned code and tests relevant to a particular claim.