A FairByDesign reporting series
The Control File
Who can actually stop an AI system, and what stops when they do? Since June we have been following that question through incidents, export controls, protocols and governance claims. This page collects the reporting so far and what comes next.
The question we are following
The reporting so far
Newest first. Mercury Intelligence pieces show a public summary; members can read the full investigation.
- InvestigationSep 8, 2026
The Kill-Switch Test
Follows the AI incidents behind the kill-switch debate and asks whether the tasks, credentials and consequences stop with the chat.
- Mercury BriefJul 27, 2026
Who Holds the Wheel? Open Source and the Fight for AI Decision Intelligence
Whoever controls telemetry, evaluation criteria, dashboards and intervention mechanisms can influence what institutions see and measure.
- Mercury BriefJun 17, 2026
Can You Reconstruct What Your AI Did?
The Reconstruction Principle and a six-question test for whether governance can actually be demonstrated after something goes wrong.
- Mercury BriefJun 16, 2026
The Export-Control Order That Took Anthropic’s Frontier Model Offline
“The off switch was real. It just wasn’t in the building you thought it was.”
- Mercury BriefJun 4, 2026
MCP Is an Operational Surface. Governance Has to Catch Up.
The Model Context Protocol as a point of control and risk, not just an integration breakthrough.
- Mercury BriefJun 4, 2026
The Accountability Control Plane
Most organizations can prove they have an AI policy. Far fewer can prove what actually happened the last time an AI system shaped a real decision.
Testing it, not just reporting it
Evidence Lab 001 — The Agent Control Gate
A controlled, reproducible experiment asking whether runtime controls can stop unauthorized AI-agent tool actions, and whether the evidence left behind can reconstruct what happened. The protocol and project plan are published.
See the studyThe Lab 001 results
The next part of this file will report what the experiment found, including failures and limits, with the method and data needed to check it.
Get the Mercury Brief freeListen & watchHelp us report this
Have you tested what happens when an AI agent is interrupted in a real system? We are looking for evidence of what stopped, what continued, and how the event was reconstructed, including tests that challenge the premise of this file.
General tips: info@fairbydesign.org. Sensitive material: viktoriabakos@proton.me (Proton Mail, end-to-end encrypted when you write from a Proton account). How we handle sources.
Go deeper with Mercury Intelligence