Representation, not scale.

The question is not whether these systems are capable. It is what conditions let capability convert into applied work.

That means the machine handling what the record supports, humans handling what it does not, and the boundary between them established rather than guessed. Established is the load-bearing word. A guessed boundary moves under pressure, and the pressure arrives exactly when the stakes do.

Every field commits early to a representation, and every representation discards something. The discard is correct when made — it buys tractability, and what it drops is not yet load-bearing.

Then it goes invisible. It stops being a choice and becomes the definition of the field. Practitioners inherit it as the shape of the problem rather than as one answer to it.

When the constraint eventually binds, the field scales rather than revisits, because the discard is not in view as something reconsiderable. More compute, more data, more capacity. Recovering what was discarded requires inverting the primitive, not adding capacity.

The constraint on application was never resources. It was the ability to represent a claim in a form verifiable by someone who was not present.

Each enlarged the radius of application — how far from the principal a commitment could be made and still hold. None created wealth. Each converted stock into flow.

The decisive case is Spain and England. Spain had the silver, in unprecedented inflows, and defaulted in 1557, 1560, 1575, 1596 and 1607. England had far less and borrowed longer and cheaper, because post-1694 arrangements made the Crown’s commitment credible to lenders who were not in the room. North and Weingast give the standard account; the mechanism is debated, the outcome is not. The polity with less resource and better representation out-applied the one with more.

Double-entry belongs in the sequence for one narrow property: it made claims verifiable by an absent party. Sombart claimed a great deal more than that, and Yamey’s critique is the reason to take only the narrow property here.

Every representation in that sequence encodes what is — held, owed, transacted, promised. None encodes the boundary: what is not there.

Completeness is one of the standard audit assertions, and the only one never directly testable. An auditor can verify that a recorded transaction occurred. Nobody can enumerate what is absent. So completeness is established through controls, cutoff testing, analytical review and sampling — and then inferred. Every audit opinion in history rests on an inference at that one point.

A signed non-membership proof makes completeness directly verifiable. Not sampled. Not inferred. Checked, offline, by someone who was not there.

AI discarded the ability to represent not-knowing. Hallucination, ungrounded compliance, agents acting on evidence they cannot establish: one discard, optimised around at enormous cost.

The precise version matters. A model can learn to express uncertainty. It cannot verify absence, because verification is a property of the corpus, not of the reasoner. That is why the layer sits beneath the model rather than inside it.

This is contested, and honestly so. Plenty of serious people hold that scaling continues to work, and the argument here does not depend on winning that. LeCun’s position is the nearest in the literature; the departure is that his claim is architectural and this one is about the substrate a model reasons over. Crystara beats vanilla JEPA on all three graphs tested, which extends that line rather than borrowing its authority.

Authority is being transferred while the discard is still in place. The UAE has mandated conversion of 50% of federal operations to agentic systems within two years — the first large public test of the conversion rate. What an automated system was permitted to conclude stops being an academic question at that point.

Capability is priced as the constraint. Conversion is the constraint.

Compute is capital-intensive, commoditising and crowded. The conversion layer is capital-light, per-organisation, per-jurisdiction, and nearly empty.

NIV’s own distinction supplies the test. Regenerative capital formation compounds; friction accumulates. AI capex is currently booked as the former, on the assumption that deployment converts. The falsifiable prediction: if grounding remains unsolved, AI capex registers in the NIV decomposition as friction before it appears in productivity statistics.

There is no standard for whether a system may be granted authority. Benchmarks measure capability. Evals measure output quality. Safety frameworks measure harm categories. None asks whether the system can establish that the record supports the action, and refuse when it cannot.

Proposed in its own terms:

A system may be granted authority over a domain if, for any action it takes, it can produce a proof — checkable offline by a third party who was not present — that the record supports that action; and if, where the record does not support it, the system refuses with a named reason that survives to the actuator.

Binary. Third-party checkable offline. Architecture-independent, so it does not age with model generations.

I have built one implementation that satisfies this criterion. It is defined so that anything could. A criterion authored by someone whose product satisfies it has to say so on its face, which is what this paragraph is for.