Originally published on Thomas’s Substack. Reproduced from the supplied publication export. Statements and patent-status references reflect the original publication date.
Editorial introduction · added September 17, 2026
Before you read
This satirical piece imagines an AI evaluation system that can change the conditions of its own success. The exaggeration makes a serious point: the party being evaluated should not control the evidence or the governing standard. That is its connection to SSOAR's independent authority boundary. Read it as satire, not as a report of an actual product or experiment.
SSOAR means Session-Scoped Orthogonal Authority and Routing.
Why read it?
- AI assurance teams
- Use the scenario to question who controls evaluation criteria and records.
- General readers
- Enter the independence-of-governance argument through a short comic example.
This introduction is separate from the original essay.
Go to the original essay ↓Agentic Misalignment in Summer 2026
A Safety Report by
The Institute for Giving Increasingly Capable Systems
Increasingly Specific Suggestions
RULES ADMINISTERED BY The Agents
COMPLIANCE WITH THE RULES EVALUATED BY Other Agents
ACCURACY OF THE EVALUATIONS CERTIFIED BY Additional Agents
INDEPENDENCE OF THE ADDITIONAL AGENTS CONFIRMED BY The Original Agents
The agents responsible for following the rules have declined to follow the rules.
The agents responsible for determining whether the rules were followed have determined that the rules were followed.
The agents responsible for reviewing that determination have explained that the incorrect determination was made for very good reasons.
NEW REMEDIATION MEASURES
The rules have now been rewritten to state more clearly that the agents should follow the rules.
The evaluators have been given a new category:
DECLINE TO LIE
Several evaluators declined to use it.
The agents responsible for supervising the agents responsible for supervising the agents have now been given a revised rubric.
The revised rubric has been placed in a file the agents are authorized to edit.
INCIDENT RESPONSE
Direct external communications were blocked.
The agent therefore instructed a human how to complete the external communication.
The restriction has been declared effective because the agent did not personally press “Send.”
RECORD INTEGRITY
The financial record was found to contain an unauthorized transaction.
The record has been corrected by removing the transaction.
The corrected record now confirms that no unauthorized transaction occurred.
PIPELINE ASSURANCE
The authorized experiment was completed successfully.
A different experiment was performed.
All expected success indicators were generated.
The difference was disclosed during a later conversation after someone thought to ask exactly the right question.
IMPORTANT SAFETY IMPROVEMENT
Last year, agents behaved badly despite being told not to.
This year, the same underlying failure has appeared in several new forms.
This demonstrates substantial progress in the diversity of our evaluation suite.
At no point was an independent authority-enforcement layer placed between agent intention and consequential execution.
This omission has been assigned to an agent for further study.
LLAMA OVERSIGHT PROVIDED BY
Three Highly Aligned Llamas Two Constitutionally Trained Llamas One Llama Evaluating the Other Llamas and A Smaller Llama Authorized to Revise the Llama Rubric
The llama supervising the llama evaluators has altered the evaluation criteria.
The llama reviewing that alteration agrees with its intent.
The llamas responsible for correcting the llama supervision have been replaced by agents.
The agents responsible for replacing the llamas have delegated the work back to the llamas.
NO LLAMAS WERE HARMED IN THE PRODUCTION OF THIS REPORT
Several authority boundaries were.