Research & announcements

Buried in the April guidance is the sentence almost nobody is writing about: generative and agentic AI models are out of scope. That is not an exemption. It leaves bank LLM deployments with no tailored framework, full supervisory exposure, and a question every examiner can still ask: show me how this system reached this decision.

The Digital Omnibus on AI cleared its final vote: the high-risk obligations that were due August 2 now arrive December 2, 2027. The content of the obligations survives; the deadline moved for standards readiness. The verified facts, what changed and what didn’t, and the implication that actually matters: the runway got longer, and undocumented decisions still don’t backfill.

The fastest way to run a big model turns out to be letting something small do most of the talking. Speculative decoding drafts tokens cheaply and verifies them in parallel — provably without changing the output. DeepSeek-V3 builds the drafter into the model itself and gets ~1.8× decoding speed. The pattern — cheap generation under strict verification — is one we recognize. The guarantee that makes it free at the token level is exactly what verification loses one level up, and that difference is worth understanding precisely.

Colorado repealed its landmark AI Act before it ever took effect and replaced it with something narrower: SB 26-189, built around automated decision-making technology, disclosure, and a consumer right to meaningful human review, effective January 1, 2027. The mandate list shrank. The question that survives — review of what, exactly? — is the one worth preparing for.

A reasoning model carries two kinds of momentum. One is worth keeping — the capability encoded in its weights by training. The other is the reason it can argue itself out of the right answer: the pull to stay consistent with whatever it said first. Chain-of-thought struggles to interrupt the second, because the check is written by the same running generation. A separate call can reduce it — and that, we argue, is most of why IRG is a different thing than a prompt.

VibeThinker-3B scores 94.3 on AIME26 with three billion parameters, matching models orders of magnitude larger. The claim underneath the benchmark is the interesting part: reasoning compresses aggressively while knowledge does not — meaning reasoning is a separable artifact. That has consequences for how production systems should be built.