Let the model do its thing, and check the homework like any normal teacher. That is the entire proposal. It has been the entire proposal in every classroom, courtroom, audit, and inspection regime ever built, and it becomes controversial only in one industry: the one currently proposing that the student grade his own homework and report the grade to the public as a safety result.
In September 2026, the CEO of Anthropic published a pacing essay whose central safety mechanism is the embedded evaluator — third-party experts given desks in the company’s offices, company badges, company laptops, retained to verify the company follows its own safety practices, with the company keeping a narrow right to redact the evaluators’ conclusions before publication. A companion paper on this site, The Model Is the Chicken, took that arrangement apart at the level of staffing. This paper takes it apart at the level of the grade.
The classroom
A teacher does not grade homework out of cruelty, and a student is not forbidden from grading his own out of suspicion. The rule is older than any school: a grade is not a number. A grade is a number plus the independence of whoever assigned it. Remove the independence and the number is still printed at the top of the page, but it has stopped being a grade and become a claim — the student’s claim about himself, formatted to look like a measurement.
This is why the arrangement survives every audit of every other institution. The defendant’s judge is paid by the state. The referee is employed by neither dressing room. The auditor’s fee is the scandal everyone already knows how to name when the client pays it. In each case the structure is identical and unglamorous: the work is done by one party, and the verdict on the work is rendered by a party the first one does not employ, house, equip, or edit.
The frontier AI industry is asking to be graded under a different rule, and it is worth stating the rule in classroom language, because classroom language is where it dies: the student writes the exam answers, grades them himself, and announces the score as evidence that the school is safe. Nobody would accept this from a school. The industry’s version arrives with better vocabulary — embedded, independent, third-party — and the vocabulary is doing all the work the structure refuses to do.
The room he hired from
The defense offered most often is personnel: the evaluators are independent organizations, serious people, drawn from the alignment research community. This deserves a direct answer, because it identifies the wrong kind of independence.
The people now staffing both the frontier labs and their would-be oversight were, to a remarkable degree, trained in the same room — the rationalist and effective-altruism community that treats AI governance as its home discipline. The lab’s CEO built his safety case in that vocabulary, hired from that community, and proposes evaluators drawn from it. When challenged, the room’s own name for its method is a shrug worn as a badge: less wrong. It is the name of the founding forum. It is offered as humility.
Less wrong is still wrong. A direction is not a verdict. A student who announces that his self-grading philosophy has made him less wrong than the other students has still graded his own homework; he has merely added a doctrine to the top of the page. Independence of persons is not independence of premises, and an evaluator who shares the maker’s founding assumptions will reliably find the maker’s founding assumptions reasonable. The blind spot is not in the individuals. It is in the room, and hiring the whole room does not widen it.
Independence of persons is not independence of premises.
Effective altruism, graded
Underneath the staffing sits the philosophy that made self-governance sound like virtue. Effective altruism asked a generation of very bright people to grade their own moral homework: compute the good yourself, trust your own calculus, and treat outside judgment as an inefficiency. Its most famous practitioner ran a crypto exchange into the largest fraud of its decade while being celebrated, in that vocabulary, as the movement’s proof of concept. The vocabulary survived its own exhibit. It always does, because a vocabulary that grades itself cannot produce a failing grade.
This paper does not argue that the philosophy’s adherents are villains. It argues something colder: a doctrine whose central move is self-assessment will staff every institution it touches with self-assessors, and will experience every demand for an outside check as a failure of trust rather than a feature of verification. That is the cultural substrate in which “embed the evaluator in our building, on our laptops, under our redaction pen” presents itself to its authors as oversight. They are not hiding the arrangement. They published it. They cannot see it, and the reason they cannot see it is the doctrine working exactly as designed.
There is no such thing as governing what you made
Here is the whole error, stated once. Governance is a relation between separate parties. The moment the governor and the governed trace to the same payroll, the same building, the same doctrine, the relation collapses into self-description — and self-description, however sophisticated, however well-intentioned, however less wrong, is not governance. There is no such thing as a maker governing his own product. There is only a maker describing his product in the grammar of governance, and the description grading itself favorably, forever, because nothing in the arrangement is capable of returning any other grade.
The fix is the teacher. Let the model do its thing — build, propose, act, at full capability, inside the scope it was given. Then let the homework be checked by a checker the model does not employ, the lab does not house, and the report does not pass through the lab’s pen to reach daylight. Attestation signed before the work begins, verdict rendered at the moment of the act, record sealed where no interested party can revise it. That is the fence this site’s other papers describe, the gate Verity runs, and the arrangement every honest institution on Earth converged on centuries before the vocabulary arrived to talk us out of it.
A student may be brilliant. A student may be sincere. A student may even, on a given exam, be right. None of it changes who holds the pen.
Never let the student grade his own homework.
None of it changes who holds the pen.
Never let the student grade his own homework.