The stage list is where the authority actually sits. Attribution runs first and asks whether the failure is even ours. Compression asks how many distinct problems the failures reduce to. Generation is one stage. Then the oracle, then a human gate [17]. Filtering, grouping, classification and verification are ordinary code, and the generator is invoked only on instances that deterministic code could not dispose of, and only for the one judgement that requires reading unstructured evidence [6].
That ordering follows from the failure mode, not from taste. A generated function that is wrong usually throws, fails a test, or fails to compile [12]. A wrong extraction rule returns a number, and a wrong mapping returns a record, so the system stays green while the data quietly rots [11]. Which is why the check cannot be a second opinion: a reimplemented oracle drifts from production silently, and its drift is indistinguishable from a model failure [9].
Volume is what makes the economics work. The problem shape is a machine-readable artifact that silently stops producing correct output because the world it points at changed, where the cost is triage and count rather than difficulty [18]. In the reported deployment the generator saw roughly one instance in three thousand [7]. The write-up does not name the denominator. If it is the 1.6 million monitored items, one in three thousand is about 533 model calls [1]. That figure is my arithmetic, not theirs, and it is worth pinning down before anyone quotes a cost per repair.
For the precision result to transfer, all six of the stated applicability conditions would have to hold in your setup [16], and you would need an artifact of that same shape rather than code with a compiler behind it [18]. The gain is also reported as a multiplier with no absolute value attached [3], which is a benchmark table with the axis labels sanded off. Precision from 0.3 to 0.6 and precision from 0.45 to 0.9 are both a doubling, and only one of them is a system you would let write to production.
The detail I find most persuasive is the ordering the author reports: the oracle existed before the model did, and could be run against historical evidence, which is what made the deployment tractable [14]. The five invariants are presented as things learned by violating them [21]. That is the tone of a postmortem rather than a pattern being sold, and it is the part I would copy first.
