The week in AI. Read through the kernel.

01 · Evidence

A record made by AI is not, by itself, a record of anything.

A police officer is under investigation for allegedly using AI to create evidence, reportedly the first case of its kind in the UK. We take no view on the case, and the structure is the lesson either way: once an artifact can be generated on demand, its existence proves nothing about who made it, or when, or how.

Detection tries to guess that after the fact, and guesses worse each model cycle. The answer is provenance carried by the artifact itself: a record made at the moment of creation of what produced it, a human or a machine, that anyone can check without trusting whoever holds it. The first is an argument. The second is evidence.

Architecture relevance. The development bears on a mark made at creation and a signed record of what an AI produced, verifiable by a recipient. See HumanMark™ and Regulayer.

02 · Containment

Where a safety control sits relative to the model is an architectural choice with consequences.

A frontier model shipped this week with its safety built into the model itself, and then had its access restricted by a government order. We take no view on either. A control implemented inside the model shares the model's availability and its access conditions; a control implemented outside it does not. The two arrangements answer different questions.

The layer that completes a model's own alignment sits outside it. It holds no handle the model can reach. It fails toward stop, not toward open. And it keeps a signed record of every decision, so the safety itself can be checked.

Architecture relevance. The development bears on containment that sits outside the model, with a record of what it did. See Regulayer running on the operator's own hardware.

03 · Agents

If agents are the new software, the off switch is the new firewall.

The loudest voice in the industry called agentic AI the next computing platform this week. The agents are coming, and they will not stay in the chat window. They will take actions at machine speed: payments, code, records, handoffs.

Every action is a new way to go wrong, and most agents ship with no external way to stop one mid-step. The layer that completes the platform is a brake on the action path, outside the agent, that defaults to stop and records what it did. Build the agents. Also build what can stop them.

Architecture relevance. The development bears on a control on the action path, with a record made at the moment of action. See Regulayer.

Deadline watch

Archived issue. Deadlines were updated in Vol. 04.

Each date below is verified against its primary source.

Jul 15, 2026 · China, Interim Measures for Anthropomorphic AI Interaction Services, in force
Aug 2, 2026 · EU AI Act Article 50, transparency obligations apply
Dec 2, 2026 · EU AI Act, AI-content marking obligations begin
Jan 1, 2027 · Colorado AI Act (revised), effective
Dec 2027 · EU AI Act high-risk obligations, current target

Procurement runs 6 to 12 months ahead of every date on this list. The buying window for each is open before the law is.

The desk

The Drift Desk names what happened, names the surface that answers it, and moves on. No blame. No catastrophe.