Frontier AI · Technology
Inside the audit that finally opened the black box
A consortium of 14 governments completed the first independent evaluation of a frontier model. What they found — and what they still can't see.
By Raj Patel · 18 min read
The audit took fourteen months and produced nine hundred pages, most of which describe what the consortium could not verify. That is precisely why it matters: it is the first independent evaluation of a frontier model with full cooperative access, and its central finding is an inventory of blind spots.
The evaluators could measure what the model does; they could not reconstruct why. Interpretability, one auditor noted, remains closer to seismology than to surgery — we read the tremors, not the rock.
The consortium''s recommendations will now test the thing no audit can: whether disclosure obligations survive their first contact with commercial competition.
Read more: The quiet revolution in GLP-2 therapeutics