Evidence register

Claims mapped to artifacts.

This page states what the public work demonstrates, how to inspect it, and where the evidence stops.

Publicly reproducible

AI Governance Benchmarks

Bounded claim: A deterministic rule-based scorer can evaluate synthetic agent-boundary cases reproducibly.
Inspect: Repository
Reproduce: pytest and the repository runner documented in its README.
Limitation: Synthetic evaluation does not establish real-world safety or independent validation.

Publicly reproducible

Bounded claim: Schema-backed receipts can distinguish example safe, blocked, and human-review actions.
Inspect: Repository
Reproduce: Validate the included JSON examples and run the repository tests.
Limitation: A template demonstrates a pattern, not deployment inside third-party agent systems.

Publicly reproducible

Bounded claim: A clean-room workflow can record risk classification, approval gates, delegation, and decisions.
Inspect: Repository
Reproduce: Follow the repository's documented example workflow.
Limitation: This is public methodological evidence, not independent certification.

Public methodology

Bounded claim: Engineering activity is tracked at exact versions and kept separate from review, readiness, and authority to act.
Inspect: Public governance summary
Private control surface: The detailed activity report is owner-restricted because it contains unreleased repository identities, review findings, and development status.
Limitation: Passing automation does not establish production readiness, safety, compliance, independent validation, consciousness, or general intelligence.