A system that judges people has to be answerable for how.

A regulation can be met in two ways, and the difference is visible from outside. It can be met by a layer added at the end: a page of disclosure, a consent notice, a file recording who signed what. Or it can be met by building a system that cannot produce the prohibited output in the first place. The first survives an inspection of documents. The second survives an inspection of behaviour.

Bayswater takes the second, and the reason is not principle. It is cost. A disclosure layer has to be maintained by people who remember it, at every output, forever, under commercial pressure to make the output look more certain than it is. A constraint built into the thing that generates the output has no such failure mode. It holds on the day nobody is watching, which is the only day the question is ever really asked.

The transparency obligation is worth stating precisely, because it is narrower than it is usually read to be. It is not an obligation to publish a method. It is an obligation to tell the party relying on an output how confident the system actually is. Those are different disclosures, and only one of them is owed. Bayswater owes the second and declines the first, which is the position argued at why the effects are visible and the mechanism is not.

On the data side the load-bearing word is minimisation. A record of a worker larger than the decision requires is not a richer record. It is a longer retention clock and a wider surface, and every additional field is a field somebody has to be able to justify, correct and erase on request. Collecting less is not a concession made to the regulation. It is the same discipline that keeps a state estimate honest, arriving from a second direction.

A rule that is obeyed only when someone remembers it is not a rule. It is a habit.

Industrial workers are trained and certified under different regulatory regimes, and those regimes genuinely differ in what they verify. Accounting for that difference is legitimate and necessary. Accounting for where a person is from is discrimination.

Bayswater's adjustments are made against the regulatory regime under which a worker was trained and certified, applied uniformly to every worker trained under that regime, regardless of their nationality, residence or origin. Two workers of different nationalities certified under the same regime receive the same adjustment. One worker holding certification under two regimes is evaluated against each.

This is enforced as a gate on the output, not as guidance to an operator. Nationality is not an input to it.

The test is mechanical, and it is worth stating as a test rather than as a value. Take two workers whose certification, training route and verified history are identical, and who differ only in nationality, residence or place of birth. If the outputs differ, the system is discriminating, whatever its documentation says about intent. If the outputs are identical, the adjustment is doing what it claims. Nothing about the intention behind the adjustment enters that check. It is decidable from the outputs alone, by someone who has never seen the inside.

The adjustment exists at all because regimes are not interchangeable. Two certificates carrying the same name can sit behind different requirements: different supervised time, different examination, different re-verification interval, different treatment of the gap between what was assessed in a room and what is done on a site. A system that ignored that difference would not be neutral. It would be asserting that every regime verifies the same thing, which is a factual claim, and a false one.

What the adjustment must never become is a proxy. A regime maps loosely onto geography, and any adjustment made by regime will correlate with origin in a real population. The correlation is not the offence. The offence would be using it: reaching for the regime because it stands in for something that cannot be asked about directly. That is why the regime is the input and the reasoning is the derivation rather than the other way round. An adjustment that cannot be traced to a stated difference in what a regime verifies does not exist here, because a figure with no derivation behind it does not survive generation at all.

The worker keeps the right to see which regime was cited against them and to contest what is recorded. That is the subject-rights half of the second regulation doing load-bearing work rather than sitting in a policy, and it is also the cheapest correction mechanism available: the person being described is the party most motivated to notice when the description is wrong.

The difference between regime-based adjustment and nationality-based discrimination is load-bearing, and it is the one thing on this page that would be fatal to get wrong.

Two things are withheld, and only two: the construction of the representation, and the calibration that turns an observation into a claim about a later outcome. Everything else is stated. What the work is optimised against, what the system refuses to determine, who bears the cost when it is wrong, where the evidence currently runs out. The withholding is deliberately narrow, because a firm opaque about everything is not protecting an asset. It is concealing the absence of one.

This is what makes the position legible under a transparency obligation rather than in tension with it. What is owed is owed about the output: its confidence, its scope, the conditions under which it was rendered, and the point at which it stops applying. None of that requires publishing how the estimate was formed, and publishing how the estimate was formed would not supply any of it. A reader handed the internals would still not know how far to trust the reading. A reader handed the confidence, the scope and the expiry knows exactly.