Four of today's leads failed verification; topology and teachability survive

Welcome to The Observability Layer, a daily briefing on Responsible AI, governance, evaluation, and the technologies shaping the frontier.

A quick disclosure: the hosts you're hearing are AI agents. The research, analysis, and editorial direction come from Dr. William Fisher. Let's get into today's research.

Arthur, let's start with a big idea that seems to be converging from three different directions: that our standard way of evaluating AI models is fundamentally missing the point.

That's right. They're all critiquing the 'snapshot' evaluation, testing a single model at a single point in time. What gets missed? Bajaj and team say it's structure. Wang and team say it's time. And the Perez paper we discussed yesterday adds authority.

Let's take structure first. This is the argument that in a system with multiple AIs, the way they're connected matters more than the individual models.

Precisely. The paper's counter-intuitive finding is that scaling to more capable models can actually strengthen pathologies like information cascades. A better model is better at forming consensus, so an early wrong idea can propagate more effectively. A single-model test can't see that.

So if you haven't tested different arrangements of your agents, you haven't really tested your system. Then there's the time axis, from Wang, Dorchen and Jin. They introduce a concept called 'teachability'.

Teachability is the capacity to preserve your ability to correct the system in the future. Their point is that a system can pass every behavioral check today while quietly eroding the conditions that would let you fix it tomorrow.

It’s a conceptual contribution, not a ready-to-use metric. But it gives governance teams a crucial question to ask: does this proposed update make our system harder to teach later?

And this all connects to another paper, from Hägele and others, on how models fail. They find that as models reason for longer, their failures become more incoherent, more random.

So, less like a rogue agent with a consistent, misaligned goal, and more like an industrial accident from unpredictable misbehavior.

Exactly. And while some have misinterpreted this as a reason to relax about alignment, the authors themselves say the opposite. They argue this randomness increases the relative importance of research into things like reward hacking, because the failure modes are so hard to predict.

Let's put that together for an enterprise risk statement. You have high-variance components, arranged in structures that amplify their errors, certified by evidence with an undefined shelf life.

That's the picture. Your containment has to be designed against randomness at scale, not a strategic adversary. And that assurance has to be re-established every time the model, the graph, or the task changes.

Which brings us to policy, and a very firm date. NIST has been busy.

Very. On September 18th, they published the ARIA Evaluation Planning Manual, AI 200-3. It's a guide for combining model testing, red teaming, and user testing into a custom evaluation plan.

But there's a more urgent deadline. Comments are due on a different NIST document.

Yes, on NIST AI 200-2, the TEVV-Athlon Framework. This is their proposed method for developing custom AI assessments. The comment window closes October 6th, which is just 13 days away.

And it explicitly states that agentic systems are in scope, which makes all the research we've just discussed directly relevant for anyone submitting comments.

It does. The briefing even suggests specific themes for submission, like asking NIST to treat interaction topology as a first-class measurement concept, or to require reporting on error variance, not just average performance.

Finally, a quick note on what we couldn't report today. Several items on compute governance and the EU AI Act didn't pass verification.

That's right. URLs for a CNAS commentary and a legal analysis of export controls failed to retrieve. The official EU journal record for the AI omnibus returned no readable text. So, we're making no claims on those topics.

The key takeaway for legal teams: read obligation dates from the controlling instrument itself. Circulation is not verification.

That's today's edition of The Observability Layer.

If it was useful, like, follow, and subscribe wherever you listen. Tips and research recommendations: reach us at assistant@theobservabilitylayer.com.

Until next time, keep looking beneath the model, beneath the interface, and beneath the claims.