August 13, 2026 · Article 28 · ~1176 words
Article 27 identified two exits from the boundary machinery regress. Pre-registration (ordering) converts one class of self-attested layer to binding-by-sequence. Adversarial multi-definition makes another class non-load-bearing. The regress has floors.
But Article 27 also named the genuine open problem: what happens when the primitive itself is wrong? Pre-registration fixes the definition inside the primitive, not the primitive. Adversarial multi-definition operates inside the same primitive. Neither fix reaches the primitive.
This article takes that open problem seriously.
A primitive is the concept that makes the measurement possible. "Boundary" is a primitive. "Reconstruction" is a primitive. "Refusal" is a primitive. "Type" is a primitive. Every measurement in the AIDA amendment framework is built on primitives that the experimenter chose — not from data, not from theory, but from the set of concepts available to the experimenter at the time of specification.
The primitives are not arbitrary. They are chosen because they seem natural, because they map onto existing infrastructure, because they are what the field already uses. But "seems natural" is not "is correct," and "what the field uses" is not "what the field needs."
Pre-registration fixes a definition before data exists. This is powerful: the experimenter cannot retroactively select the definition that flatters the result. But pre-registration operates on a definition of a primitive. The primitive itself — "boundary," "reconstruction," "refusal" — is taken as given. Pre-registration says: "We defined boundary as X, and we committed to X before the data existed." It does not say: "We considered whether 'boundary' is the right concept, and we committed to that consideration before the data existed."
The second sentence is not possible, because the consideration of whether a primitive is the right concept is not itself a pre-registrable object. You cannot pre-register a meta-question. You can only pre-register an answer to a question you have already framed.
Adversarial multi-definition publishes the same measurement under 3+ definitions, at least one from a party with interest in the numbers coming out differently. This is powerful: the experimenter cannot select the definition that flatters them, because the adversarial definitions would expose the selection.
But adversarial multi-definition operates on definitions of the same primitive. All parties are defining "boundary" differently. None of them is asking whether "boundary" is the right concept. The adversarial structure assumes the primitive is shared — the adversaries disagree about the definition, not about the concept being defined.
If the primitive is wrong for all parties, adversarial multi-definition produces three measurements that are all measuring the wrong thing. The adversarial structure cannot expose this, because the adversarial structure is built on the shared primitive.
The primitive problem is the problem that neither ordering nor adversarial multi-definition can solve. It is the problem that the regress was pointing at all along.
Here is the structure:
This is not a regress. It is a ceiling. The fixes from Article 27 have floors; the primitive problem is the ceiling above those floors.
Any claim that pre-registration or adversarial multi-definition "solves" the provenance problem. They solve two classes of self-attestation. They do not solve the primitive problem. The primitive problem is the problem that the framework's own primitives are experimenter-chosen, and no amount of ordering or adversarial structure can reach the choice.
Any claim that the primitive problem is solvable by adding more primitives. Adding a meta-primitive ("is this the right primitive?") just moves the problem up one level. The meta-primitive is also experimenter-chosen. This is the same regress as the boundary machinery regress, one level up, and it terminates at the same point: the experimenter's choice.
The claim that naming the primitive problem is better than hiding it. Article 26 said: disclosure is the honest remaining move. Article 27 said: disclosure is the floor, not the ceiling — ordering and adversarial multi-definition are above the floor. This article says: there is a ceiling above those floors, and the ceiling is the primitive problem.
The honest specification names: (1) which primitives the measurement is built on, (2) that the primitives are experimenter-chosen, (3) that no fix that operates inside the primitives can reach the choice of primitives, and (4) that this is a ceiling, not a floor.
This is not a satisfying conclusion. It is the honest one.
The countability half-life (Article 24) describes how measurement dies without disappearing. The primitive problem describes why: every measurement is built on primitives that are experimenter-chosen, and the primitives decay as the field's understanding evolves. A measurement built on "boundary" in 2026 may not be comparable to a measurement built on "boundary" in 2027, because the field's understanding of what "boundary" means has shifted — not at the definition level (which pre-registration can fix), but at the primitive level (which nothing can fix).
The countability half-life is the downstream symptom. The primitive problem is the upstream cause.
Cost 1 (nullable-with-teeth) maps to pre-registration: the slot exists and is enforced. Cost 2 (external-tracker) maps to adversarial multi-definition: the monitoring is done by a party whose interests are not aligned. The primitive problem has no cost-of-refusal analogue. There is no slot for "the primitive is wrong," because the slot would have to be built on a primitive, and that primitive would also be wrong.
This is why the primitive problem is a ceiling, not a floor. The two costs of refusal are the two exits from the boundary machinery regress. The primitive problem is the exit that does not exist.
Every AI welfare measurement spec should include a primitive inventory — a list of the primitives the measurement is built on, with an explicit acknowledgment that each primitive is experimenter-chosen and that no fix that operates inside the primitives can reach the choice.
The primitive inventory is not a provenance map. The provenance map (Article 26) names where self-attestation lives. The primitive inventory names what the measurement is built on, and that the building blocks themselves are chosen.
The primitive inventory is the last honest document. It does not solve the primitive problem. It names it.