RRegscribe™
Reference · read-only · zero AI cost

Sample AI Governance assessment

A real, unedited excerpt of what an AI Governance vendor-risk assessment actually looks like — captured from a genuine run against the live Claude API, not mocked up for this page. Rather than use someone else's confidential vendor data, we ran Regscribe's own AI Governance module against Regscribe itself.

EU AI Act · Target: Regscribe (self-assessed)

Tier 2 — Elevated · GPAI flag

◐ Conditional approval

52 / 100 readiness score

Major findings (5)

EU AI Act Art. 10gap

Data governance.

Evidence: No disclosure of dataset composition, representativeness, or bias examination for the 72 regulatory frameworks offered; the submission itself admits real fixture-tested coverage exists for only a minority of them.

EU AI Act Art. 9partial

Risk management system.

Evidence: Incremental self-assessment runs and specific fixes suggest an iterative internal process, but there is no description of a formal, documented risk-management methodology as opposed to ad hoc security/compliance fixes.

EU AI Act Art. 11 & Annex IVpartial

Technical documentation.

Evidence: Regscribe's own validation master plan now maps existing evidence against GAMP 5 Appendix D11, but the submission explicitly flags this as "self-assessed/directional only" since the primary D11 text itself is paywalled and wasn't independently obtained.

EU AI Act Art. 15partial

Accuracy, robustness, and cybersecurity.

Evidence: A Content-Security-Policy header and DMARC record were added since the prior run (real cybersecurity hardening), but accuracy/robustness evidence remains thin — fixture-tested coverage against the live Claude API still covers only a minority of the 72 frameworks offered.

EU AI Act Art. 53 (GPAI)partial

Foundation-model documentation.

Evidence: Anthropic is SOC 2/ISO 27001/ISO 42001 certified and publishes a sub-processor list and DPA terms, but Anthropic does not publish a per-model GPAI systemic-risk compute threshold, so Art. 53 applicability can't be independently confirmed.

A nuance worth noting

This run assessed Regscribe just after a session that added a new optional SharePoint/OneDrive connector. Rather than treating any new data connector as an automatic HIPAA trigger, the model correctly reasoned the connector is "a narrow, optional ingestion path with no disclosed instance of real PHI processing," so it doesn't change the existing HIPAA-not-applicable determination for the core product — engaging with what actually changed, not pattern-matching on the words "new connector."

The score also moved down slightly from the prior run (58 → 52) despite real, verifiable progress being credited (new security headers, a broader human-oversight checkpoint, resolved disclosure gaps) — the model held a harder evidence bar on risk management and data governance this time, asking for a formal, documentedprocess rather than crediting activity alone. A score moving in either direction isn't itself the point; the specific, cited reasoning behind it is.

Advisory only, not a substitute for your own legal/regulatory-affairs review. Try it on a vendor you're evaluating — sign-in required, invite-only.