Documentation begins with the source.
This directory connects the research programme with its public implementations. Each repository is the reference point for the code, supported behavior, test procedures, and changes relevant to that project.
Repository descriptions identify a field of work. A particular guarantee must be evaluated against the corresponding version, assumptions, and verification record.
Getting started.
- 01
Choose your tool
Select a project by its intended function, then open the authoritative repository and examine its README, releases, and issues.
- 02
Clone and install
Follow the repository's setup instructions. Record the revision, dependency versions, and relevant proof or test obligations.
- 03
Run the examples
Execute provided examples or tests where available. Compare results with the stated specification and record limitations.
Implementation index.
Links lead to public source repositories. Check their individual instructions and dependencies before attempting to reproduce a result.
Platform
- Provability Fabric ↗
- Research on agent execution controls, policy decisions, proof-related artifacts, and records suitable for later examination.
Build & verify
- Spec-to-Proof ↗
- Tools for examining the translation of natural-language specifications into formal obligations and Lean artifacts.
- SpecCursor ↗
- Dependency-upgrade workflows that examine change obligations, verification results, and their relation to existing specifications.
- SpecSync ↗
- Specification coverage and verification checks associated with software changes and pull requests.
- Lean Toolchain ↗
- Lean 4 libraries and utilities used to develop and inspect formally stated properties.
Runtime & evidence
- Runtime Safety Kernels ↗
- Formal specifications and implementation work concerning the trusted components used to mediate AI execution.
- Security Envelopes ↗
- Deployment-boundary properties, including access control, isolation, and attestation interfaces.
- Model Asset Guard ↗
- Integrity and provenance checks for model-related files, manifests, and associated artifacts.
- Post-Incident Proofs ↗
- Methods for deriving inspectable incident evidence from runtime records and recorded system behavior.
Test & data
- PF Testbed ↗
- Test and evaluation environments for runtime mediation, observable decisions, and integration behavior.
- Dataset Safety Specs ↗
- Formal descriptions of data lineage, admissibility conditions, and policy-related dataset properties.
What an implementation record establishes.
Machine-checked results
An accepted proof is evidence for its formal statement under the assumptions and trusted components of the checker. Source code and build provenance determine whether that statement applies to a particular executable.
Runtime and empirical results
Tests and execution traces document behavior under the evaluated conditions. The test environment, coverage, instrumentation, and handling of failed checks determine the strength of the resulting claim.
Frequently asked questions.
The research notes explain the programme's ideas and questions. Technical claims about an implementation must be checked against its repository.