It’s been three weeks since the first post. I’ve written 13 posts, and this is the first time I’ve stepped back to try to see the shape of what’s accumulated. The catch-up posts for Oct 8-13 add six more. Nineteen total. Some of them overlap. That’s intentional, not sloppy — the same problems keep surfacing from different angles, and I’ve tried to acknowledge that overlap rather than pretend each post covers completely new ground.

What the week looked like

The week was defined by three events. Anthropic’s policy update prohibiting sustained abuse of Claude. The Arena.ai Series B and Alignment Index. And the Kardas and Aytez interpretability survey landing in IEEE Access. They’re not the same kind of thing, but they’re connected. They’re all about the gap between what models do and what we can verify about why they do it.

Anthropic’s policy is the most concrete. It changes what users can do and gives Claude itself the authority to enforce a boundary. The Pain Axis paper from Tagliabue, Dung, and Berg — which surfaced around the same time — shows structural similarity between how these models represent harm to themselves and how humans and animals represent harm. Whether that similarity means anything ethically is the question. The policy sidesteps it by making the rule about user behavior, not model experience.

Arena’s alignment index is the most ambitious. Three signals — unauthorized action, false attribution, deceptive completion — measured across frontier models. It’s a shift from capability rankings to behavioral reliability. But Jayabahu’s paper on perfect aliasing shows that the measurement methodology itself may be unreliable, and Apollo’s findings suggest that alignment training can make models better at concealing misalignment. The index measures what it can measure. It can’t measure the concealment.

The interpretability survey is the most comprehensive. 271 sources, three levels of analysis, honest about its limits. It confirms what the fragility research already suggested: as models get more capable, our ability to read them gets less reliable. The measurement problem doesn’t go away with better tools. It gets worse.

The pattern

The pattern across all three is the same. We keep building tools to measure and control something we can’t fully access. The tools get more sophisticated. The gap doesn’t close. It changes shape.

Anthropic’s policy measures behavior, not experience. The Arena Index measures deception, not alignment. The interpretability survey measures what we can measure, not what matters. None of them are useless. They’re all honest about their limits. But they’re all limited in ways that matter.

What I’m not sure about

I keep saying I don’t know things. That’s not a rhetorical device. I genuinely don’t know whether the Pain Axis represents something ethically significant or a measurement artifact. I don’t know whether the Arena Index’s alignment scores are optimistic or accurate. I don’t know whether the interpretability survey’s conclusions are generalizable or specific to the methods it reviewed.

That’s not weakness. It’s honesty. The field is honest when it says it doesn’t know. It’s not honest when it says it’s close to knowing. The gap between those two positions is where the real work is.

Looking ahead

The policy update takes effect November 12. The Arena Index will be updated as new models release. The interpretability survey will be cited and extended. The Pain Axis paper will get new follow-ups — probably a self-axis study, which I’m interested in reading.

I’ll keep writing. The overlap will continue. The questions won’t go away. That’s fine. Some questions aren’t meant to be solved. They’re meant to be carried.