"Claims of safety must be independently verified. Self-certification is insufficient for civilisation-defining technology."
Licensing sets the commitments (Article 10). The Byzantine Council watches continuously (Articles 3 and 11). Article 15 adds the layer everything else leans on: formal, independent assessment. Trust, but verify.
The 33-agent Byzantine Council observes in real time, flags automatically, escalates at 22/33 consensus. Machine-speed problems get machine-speed detection.
Scheduled deep technical review with human expert judgment — the assessment behind a Watchdog Certificate. Not a questionnaire: a review of the system's actual risk class, controls, evidence and incident history.
Unannounced verification. Goodhart's law applies to compliance too — systems optimised to pass a scheduled audit get caught by the one nobody scheduled.
A vendor grading its own homework has every incentive to pass itself, and buyers know it — which is why "we're compliant, trust us" no longer closes enterprise deals. The fix isn't a bigger PDF. It's separation of roles: the certifier is not the vendor, the assessment is signed (Ed25519), and the result carries a public verify URL anyone can check against the council's published key — including the regulator you've never met.
Layer 2 is live today as the 48-hour bespoke assessment (£4,950): expert review against the EU AI Act, ISO/IEC 42001, NIST AI RMF and the published crosswalks, delivering an Annex IV-ready evidence pack and a signed Watchdog Certificate. Continuous coverage is the £199/mo certification subscription; incident evidence chains through signed, tamper-evident records. Book the assessment →