The report
Guidelight assessed frontier AI labs against six control practices using public disclosures and externally visible evidence. The result was uneven implementation rather than a simple safe/unsafe split.
The methodological limit matters: public documentation can reveal whether a lab describes a control, publishes evidence for it or exposes a governance mechanism, but it cannot substitute for direct access to internal systems.
DiggingBeagle should therefore use the assessment as comparative evidence about disclosed control maturity, not as a complete security audit of hidden infrastructure.