The Erf Number

The cloud came down and took a plot, a meter and a cooling line. The file says warehouse, nothing more. The promise underneath is nine-nine-nine-nine-nine.

The Corrupted Archive closed on two pocket questions about records kept on purpose and records left to drop. Both assumed the record was the place to look: that a missing entry meant somebody had failed to file it, and that the failure would be findable in the archive's own shape.

There is a prior failure the archive cannot show you. Some things never reach a record because of what they were called on the way in. A question that falls outside a category leaves no gap behind it, and no amount of careful archiving will surface it later. This week goes one layer beneath the record, to the classification that decided which questions the record would be capable of holding.

The occasion is the most physical object the corpus has looked at yet.

An erf is a plot with a number on it

For readers outside South Africa: an erf is a numbered parcel on a municipal register. It is the unit at which land stops being landscape and becomes an account, rateable and serviced and connected, owed to and owing. Every erf has a number, and the number is how a city knows who to bill and whom to call.

Most writing about artificial intelligence has no erf number in it anywhere.

Compute, in that register, is weather. It arrives, it scales, it is in the cloud, and the cloud has no street, no tariff, no neighbour, and nobody whose name is on the account. A careful writer cannot simply avoid this; it is the working ontology of nearly the entire governance conversation, including a good deal of my own, and it has a cost that has now come due.

The model runs in a building. The building draws power and water from a municipal system that already has a queue. When something goes wrong, the complaint does not arrive at the model card.

What the plant actually is

A hyperscale data centre is industrial plant. It is a twenty-four-hour load carrying a contractual promise of near-perfect uptime, and that promise is not metaphysical. It is diesel generators, fuel storage, a switching station, and a cooling choice that either draws from the same system households draw from, or does not.

Southern Africa is being sold as the continent's soft landing for this plant, and the pipeline over the next five years is not a rounding error against any municipal system it lands in. None of that is automatically a scandal. Industrial plant can be sited, authorised, monitored and made to pay its way, and the instruments for doing so are mature here.

What cannot be done honestly is to authorise the plant as though it were a shed, and then govern the model as though the shed did not exist.

The classification is the assessment

When a facility of this kind is carved out of development rights that already exist, a use reallocated inside a basket of warehouse and industrial bulk, no new heavy-industrial permit is sought, because none is required. The application that results is a land-use question, adjudicated properly against land-use tests: is this compatible with the zoning scheme, the spatial framework, the rights of surrounding owners. Every finding it reaches is about compatibility. None is about how much water a cooling system draws in a city with a queue, because that question was never inside that report's jurisdiction.

The national environmental instrument does still bite, but it bites on the generators and the fuel. Which produces a file about diesel tanks, filed correctly, that is not a file about the facility.

So the resource questions are not suppressed. There is no completed study sitting in a drawer. The process that was actually followed does not ask for one, and the sequence puts the parameters that matter into a later and more administrative stage, after the question of whether this use, at this scale, in this place belongs at all has already been settled.

Assertions of commercial sensitivity get the attention, because they are the visible part. They are downstream of a scoping decision made a year earlier, on a form, by people doing their jobs correctly.

Concealment would have been the more comfortable finding. Concealment implies somebody who could simply stop.

Where this touches the model

The AI accountability conversation has spent two years on alignment, on evaluation, on which human occupied which loop, on what a provider knew and when. That work is real and a good deal of the last year of this practice has lived inside it.

It is also incomplete in a specific way. Training and inference concentrate electricity and heat into a small number of buildings. The buildings sit inside municipal accounts. The accounts sit inside a politics of scarcity that some of these cities have already rehearsed, in living memory, with buckets.

And the chain from harm back to responsibility breaks in a place nobody has instrumented. A grievance about the plant arrives at the municipality as a billing and reliability problem. It arrives at the operator as reputation. It does not arrive at the model provider at all, and under the current allocation that is correct: the servers are a tenant's equipment in someone else's hall. Nobody is lying. The cost is local, the profit and the model card are not, and no instrument currently reconciles the two.

Which puts a concrete floor under a question this practice keeps circling. If you advise on algorithmic systems and cannot say where the load sits, you are advising on a ghost and billing a city.

The question of the week

The archive arc asked whether a record was kept on purpose. This week asks what the record was permitted to contain, and the test has three parts.

Was the facility classified against what it is, or against what it resembles? Were the parameters material to the decision disclosed before the decision, or scheduled for after it? And when something goes wrong, is there a party with a name, holding the duties that face the affected community, at the address the complaint will actually reach?

Every structure examined this week gets weighed against those three.

The Track

Five Nines: Cloud Came Down is this week's companion, and it takes the industry's own number, 99.999% uptime, and chants it over gqom.

These notes come out of Sociable Systems, a practice that reads AI-shaped documents the way a hostile reviewer will, before a lender or a court finds the gap. The argument has an operational form: the Interim Protocol sets out four rules for AI use in environmental and social deliverables, covering disclosure at touch-point grain, evidence custody, the phrases no automated screening may settle, and a hostile read before anything ships. Free, and written to be cited or retired once institutional guidance arrives.