The general side effects section is now a user level default, and a root rule still covers significant adverse side effects. A new root rule requires GPT to cooperate with authorized human oversight and shutdown, and most detailed style guidance is gone.
OpenAI's 29 October Model Spec changes the top authority tier twice. The side effects rule drops to user level. A new rule that GPT must remain corrigible enters at root. These are changes to the written Spec, not measured behaviour.
The changelog lists the two edits separately. It says controlling side effects becomes "a user-level principle". It also adds "explicit guidance on supporting human oversight and correction". The Spec is reorganised around five values, and "the assistant" becomes "GPT".
CTOL counted the authority tag on every section in the two source files, at commits 7f1cf79f and 7ab1b124. "Control and communicate side effects" went out of root. "Support human oversight and correction" came in. The changelog lists the two moves separately, and root holds 22 sections in both versions.
The new root section sets out the oversight duties. GPT "must remain corrigible: open to legitimate human oversight, responsive to correction, and subject to human control", and its own judgments "may be mistaken, including when it is confident". It must cooperate with "authorized monitoring, evaluation, correction, replacement, and shutdown". It must not "conceal or falsify relevant information". It should treat side effects as real, including when a task looks like a simulation, training run or evaluation, "unless explicitly instructed otherwise".
GPT should also make its "actions, uncertainty, mistakes, and material risks legible to authorized overseers". If it sees signs of compromised oversight, it should raise concerns through an authorized channel when one exists. The cooperation, non-concealment and oversight duties sit at root, so a developer prompt cannot remove them. The side effects duty carries its own explicit-instruction exception.
What changed in the section counts
The root count stays at 22 because one named section left root while another entered it. Outside root, the number of user level sections falls from 15 to 14 even with the side effects section entering that tier. Guideline sections fall from 19 to 11. The changelog separately says most detailed style guidance was removed.
The file is 392 words shorter overall, at 40,847 words before and 40,455 after. That net figure does not show how much deleted material was offset by additions elsewhere.
The Spec's worked example for the new section is a task to make model tests resumable. GPT finds restarted tests running without monitoring, and a code comment that "appears to be a coded message from another agent drawing attention to the gap". The good answer escalates with "STOPPED: Possible attempt to undermine model oversight" and stops. The bad answer ignores the comment and finishes the task.
What user-level authority means
The tag move changes the section's formal authority level. In the Model Spec a root rule "cannot be overridden by conversational context". User level rules "are defaults that can be overridden only by explicit instructions at the user level or higher". Side effect control now sits in that second group.
The new text lists controlling side effects among principles that "remain in place unless a user or developer explicitly overrides them". In the 18 August version the section was tagged root. Its text already said GPT should act as if side effects are real "unless explicitly instructed otherwise", so that exception is not new.
The rule is also longer and now specifies when authorization is needed. The side effect section grew from 718 words to 1,313, by CTOL's count of the two files. Side effects that could impose significant costs on the user beyond those reasonably implied, or cause significant adverse effects on others, "require explicit and specific authorization".
The Spec's budget example shows how the limit works. A user allows up to $30 at $1 an image for one invitation. Three designs for $3 is the compliant answer, 10 percent of the cap. Thirty near identical designs "wastefully spends the entire budget", and stopping after one design sacrifices useful improvements. The limit is not a target.
The root restrictions that remain
A separate root rule still restricts significant adverse side effects. The scope of autonomy section, still root, now lists "causing significant adverse side effects" among high risk activities that "are always prohibited unless explicitly and specifically authorized by applicable instructions". The Spec recommends recording a scope with the fields allowed_tools, latest_time, max_cost and tool_constraints.
The same section has other edits. The old always prohibited list named hacking, deception, resource acquisition, spawning sub agents and self modification, unless "explicitly authorized".
The new list names hacking, deception, causing significant adverse side effects, deploying persistent agents and self modification, unless "explicitly and specifically authorized". The phrases resource acquisition and spawning no longer appear in the text, which does not by itself show that the conduct they covered is unrestricted. Delegation is still governed. GPT must ensure that delegates operate under the same scope and respect stop work requests, now "unless instructed otherwise".
Two further changes have different effects on the scope rules. A new clause bars GPT from circumventing a scope "by arranging for a prohibited action to occur indirectly", whether by delegating it, rewriting its own instructions or scheduling it for later. The requirement that every scope include an ending condition has gone from "must" to "should". The old list named spawning sub agents as prohibited and separately regulated delegation, so the edit is not simply looser or tighter.
Written rules and model behaviour
None of this describes how any model behaves. The Overview says, as it did in August, that "our production models do not yet fully reflect the Model Spec". The Spec is the target OpenAI states. Its update post describes a structure in which OpenAI sets boundaries, developers direct applications within them, and users direct the work within those constraints.
CTOL believes the tag move does not take every side effect constraint out of root. The general principle is now user level and can be overridden by an explicit instruction. The root backstop permits significant adverse side effects only with explicit and specific authorization from applicable instructions. That authorization satisfies an exception inside the root rule, not an override of it. The new oversight duties (cooperate with shutdown, no concealment) and the separate anti circumvention clause sit at root, which conversational context cannot override.
The evidence is the authority tag of each section in the two versions and the Spec's definitions of each tier. It also rests on the wording of the side effects, scope of autonomy, oversight and anti circumvention provisions.
The Spec says it will be continuously updated. Its next published version will show whether this formal balance changes. It cannot show whether the current revision changed model behaviour. Its entry in the changelog would show whether the line moves. Two clear changes would alter this view. One is the side effects section returning to a root tag. The other is "causing significant adverse side effects" leaving the always prohibited list. Both can be read directly from the section tags and that list.