Superseded. This is an earlier draft of the methodology chapter, retained as a record of what the work believed at the time. It is not current: the composition it describes includes constrained elevation, which the design chapter withdrew after implementation established that no axis of a basis is elevatable. The current chapter is the one published under the methodology category.
The chapter set itself one task at the outset, inferred from the review's finding rather than taken from a methods textbook: to establish that a discipline for refusing unwarranted claims is itself warranted, without relying on the designer's judgement as the evidence for its validity. The inference is short and it is stated in the opening section so that a reader can reject it, in which case what follows loses its organising principle. This closing section takes stock of how far that task was carried, separates what is settled from what is genuinely open, and states the limitations plainly rather than leaving them for a reader to discover.
What is settled
Most of the chapter is firm, and it is firm for a particular reason: it follows from the review rather than from a preference about how the work should be conducted. The paradigm follows from the review's finding that the artefact does not exist, so there is nothing to observe and the thing must be built to be known. The typing of the artefact is read off the composition the review found missing, into categories the paradigm itself supplies. The circularity threat is named before its answer, and the commitment that answers its timing half is the paradigm's own requirement that objectives precede design rather than a scruple this work volunteered. The criteria that test whether the right thing was built are quotations and faithful reports from the surveyed fields' own statements of where they stop, authored by people who never heard of this work. And failure is declared in two faces given equal weight, with over-conservatism named as a failure mode rather than a safe harbour, so that the evaluation cannot be passed by a discipline that simply refuses everything. None of these rests on the designer's judgement at the point where the review said judgement is exactly what must not be relied upon.
What is open
One part is open, and it is marked open rather than dressed as closed. The discrimination question, whether the discipline separates the claims it should refuse from those it should admit, has a route rather than a settled answer. The coverage and risk relationship borrowed from selective prediction shows how such a mechanism can be evaluated without a human labelling any case, and it converts the ground truth from opinion into outcome, which is what removes the designer from the decisive point. But it is a route with work still owed on it: the selective-prediction literature re-read on its evaluation protocol specifically, and a defensible account of what error means when the quantity is an unwarranted action rather than a misclassification. The residual that the scenario distribution remains the designer's is real and is not claimed to be solved. Publishing the chapter with this part marked open is deliberate, and it is the same discipline the chapter argues for, applied to the chapter itself: if the discrimination problem has no good answer, that is worth discovering while the structure is still cheap to change rather than in a chapter written as though the answer were in hand.
The limitations that remain
Three limitations bound the work and are stated here rather than conceded reluctantly under questioning.
The sourcing is open-access only. Every field surveyed in the review was read in its own primary source, but only in sources that are openly available, and paywalled material was excluded and the exclusion recorded rather than hidden. This bites least where the load-bearing argument sits, since the near-neighbour comparison rests on public primary standards, and most on the flanking literature. It remains a limitation, and a reader is entitled to know that the survey was drawn from what could be read in full without a subscription.
The work has a single researcher. There is no second party to label scenarios, adjudicate disputed cases, or supply an independent verdict, and this is not a incidental fact but part of why the evaluation had to be arranged as it was. The reliance on inherited criteria and on outcome-based ground truth is partly a response to the absence of anyone else to vouch, and while that arrangement is defensible on its own terms, the constraint that produced it should be named.
Triton is bounded to illustration. The incident is a strong demonstration that an unsupported diagnosis can be acted upon at a handover with physical consequence, and it is used for that and no more. It is not evidence about probabilistic generation specifically, and the chapter does not lean on it as though it were. Its role is to make the failure concrete, not to carry the argument about machine-generated claims, which rests elsewhere.
What the chapter hands on
What the methodology leaves for the design chapter is a specification it did not have to invent and a set of tests it did not have to author. The parts to be built are typed and their required content named. The threat that the designer both sets and passes the test is met where it can be met, and recorded where it cannot. The criteria for having built the right thing are quoted from other fields, and the criterion for the right thing working is identified even though its instrument still needs sharpening. The design chapter begins, then, not with a blank page but with a commissioned one: build the smallest basis model that could work, discharge the scheduled debt against its load-bearing semantics before fixing it, and take it to the evaluation the discrimination section has already begun to lay out. Whether it can be built is still the question the rest of the work asks. The methodology's contribution is to have made that question answerable without the answer resting on the word of the person asking it.