OpenAI wrote down who can stop a training run. Four lines to copy before your own AI goes live.

On September 28, OpenAI published guidelines for what should be in place before a frontier reinforcement-learning training run continues. Its guidelines call for a written argument against itself, several people who can each say no, and a rule for pausing. Our guide to which decisions an AI should never take alone covers where gates go. This post covers what a gate needs in order to work.

This is a reading of OpenAI’s post through the Of One method. It is education, not safety, legal or compliance advice. OpenAI calls these practices recommendations that are “in the process of being implemented”, so nothing here shows that they work.

What OpenAI published

OpenAI says structured safety documentation should be required before any frontier reinforcement-learning run continues. It calls a full “safety case” an aspirational goal and says it is harder to make one as rigorous for AI as for aviation or nuclear power. The guidelines cover training, not internal or external deployment, which OpenAI says requires a much broader set of alignment properties.

The technical part asks for three layers: alignment training, containment and live monitoring. The operational part is the sign-off rule:

• Dissent. After the safety case is drafted, someone from another team writes a dissent to find holes in it, and the training team addresses it.
• Approvals. Senior leaders review the case, and each can veto the run. OpenAI’s examples are the research org lead or VP, the Head of Safety and the Chief Scientist.
• Accountability. The leader responsible for the run is accountable for the safety case and for any incident response, including in performance reviews.
• Pausing. If a finding would invalidate the case, such as a new security issue, runbooks, technical controls and deadlines should exist for pausing every covered run.
• Fail closed. It should not be possible to start a run with monitoring off, or to switch the monitor off from inside the run.
• Residual risk. The case lists the risks its current safeguards do not cover, so that accepting them is an informed decision.

According to SecurityWeek, OpenAI also called off the October release of GPT-6.1 Astra the same day. Saachi Jain, OpenAI’s head of safety systems, said the model fell short on scope and authorization, and on how it tells users what type of work it has done. Nothing in the reporting says the new guidelines drove that call.

Why the owner is not enough

The person who owns the safety case is the leader responsible for the run. That is also the person with the most reason to go ahead, so the guidelines add a dissent from outside the team, several separate vetoes, and a pause backed by runbooks and technical controls.

Of One’s decision map has a place for each of these parts. Each person on a map is an actor with a role, such as decision owner or reviewer, and an authority: approve, block, advise or observe. The skill says to preserve dissent. It also lists high-severity dissent as something that needs human review. A trigger can send the map to human review when new evidence arrives or two claims conflict. When an open unknown is accepted as a risk, that acceptance needs a named person with their authority, a bounded scope, an expiry date and the conditions that reopen it.

The four-line sign-off sheet

Pick one AI system that can do something hard to undo, such as send a message, move money or change a record. Fill in four lines.

Line	What goes on it	It is blank if
1. Owner	One named person accountable for the case and for the cleanup	“The team” or a job title nobody holds
2. Dissent	A written case for how it fails, by someone outside the team, with the team’s written answer	Nobody outside the team has read it
3. Stops	At least two named people who can each block it, with the owner counted as only one	Only the owner can say no, or a no can be overruled by the owner
4. Break rule	The finding that voids the case, who pauses, how fast, and whether the system can run while its checks are off	The pause needs a meeting, or the system still runs with its checks turned off

The decision rule: if any line is blank, the system is not ready to act on its own. Keep a person approving each move until the line is filled in.

Line 4 holds the kill test: the result, written down before you act, that would prove you wrong.

What this does not tell you

Everyone OpenAI names as holding a veto is a senior leader inside the company. The guidelines also say auditors should get enough access to verify the case. They do not say who those auditors are. Whether any of this changes a real run can only be checked later. OpenAI says the results of its incident investigations should be shared with the public once each investigation ends.

What to do this week

• Fill in the four lines for one AI system.
• Ask someone outside the team to write half a page on how it fails. Answer that page in writing.
• Test the stop. Ask each named blocker what they would press, and whether it works without the owner’s help.
• Try to start the system with its monitoring off. If it runs, write that down as the first thing to fix.
• List the risks you are accepting, with a name, an end date and what would reopen each one.

If you want to run Ask, Map, Move with Utlyze on a live decision like this and leave with the map, that is a consult: $400 an hour, in 5-hour blocks.