Epistemological angle: This is the entry that holds AI to the standard the rest of this collection demands of everyone else, and it is worth being explicit that it meets that standard where most of the sham-tested procedures in theme one did not. Randomised, population-based, pre-registered, primary endpoint pre-specified and reported on a hard outcome rather than a surrogate. Note what was actually tested: not the machine replacing the radiologist but the two together against the standard of care. The comparator matters as much as the result — Swedish double reading is a demanding benchmark, and the intervention beat it on sensitivity while halving the human reading burden. What licenses the result is also worth naming: the model's training distribution and its deployment population were the same, which is precisely the condition entry 42 shows to be absent when these systems fail.

Epistemological angle: An orthopaedic classification in use since 1957 was shown to be lossy by a system with no stake in it. That is worth separating from the usual framing. The finding is not that the algorithm is cleverer; it is that the radiographic grade discards information the knee contains, and that what it discards is not randomly distributed. Read against entry 47, this is the same lesson from the opposite direction: an ordinal category imposed on a continuous structure loses signal, and the loss is invisible from inside the category because the category is what you are looking with. It is also the collection's clearest case of a machine functioning as a genuinely third kind of cognizer — not faster at our task, but sensitive to something our task was never set up to see.

Epistemological angle: The reference class problem of entry 17 in its purest mechanical form, and the reason that entry belongs upstream of this whole theme. A trained model's reference class is its training distribution, and nothing in the model selects that class — the data do, silently, through whatever happens to covary with the label. Here the class turned out to be something like "pelvic radiographs from this scanner, at this hospital, ordered with this priority flag." The 0.52 is the number to carry: strip the confounders and the fracture detector detects nothing. Hájek's point that the choice cannot be eliminated but only made visible applies exactly, and it is why the deployment question for these systems is not how accurate they are but on which population that accuracy was estimated, and whether the patient in front of you is in it.

Epistemological angle: The counterpart to entry 40, and the reason a favourable trial does not settle the question. MASAI shows what a well-matched system adds to a reader; this shows what a wrong one subtracts, and the subtraction is larger than the addition. The epistemological content is about testimony. A machine's output enters the reader's judgment not as one datum among several but as an anchor, and the very experienced group — down thirty-seven points — demonstrates that expertise is protection rather than immunity. Read with entry 19: the failure is in the same family as the anchoring effects that moved surgeons' answers by 22 points, except that here the anchor arrives with institutional authority attached.

Epistemological angle: This is the theme's most important entry for anyone thinking about how to integrate these systems, and it is not an argument against them. Entry 40 establishes that the combination outperforms the human alone. This establishes that the human inside the combination may not be the human who entered it. Both can be true, and taken together they reframe the question from whether to adopt to what to measure afterwards — the unaided performance of the assisted clinician is a quantity nobody was tracking, and it turns out to move. The design cannot establish causation and the authors do not claim it does; it is a before-and-after comparison with all the vulnerabilities that implies.

Epistemological angle: Taught as a pair, in the manner of a genuine disagreement between competent people rather than a controversy with a right answer. The dispute is the mechanism-versus-outcome argument this collection has already run twice, transposed. Entry 23 shows a drug that suppressed the mechanism and killed patients; entry 39 shows repairs that fail structurally in patients who do well. In both, the causal story we could tell was disconnected from the result we could verify. London's argument is that this is medicine's normal condition and that demanding explanation from machines while tolerating its absence in ourselves is inconsistent. Ghassemi's is that the inconsistency does not license accepting an explanation we know to be confabulated. Both are correct about something, and the collection cannot resolve it.

Epistemological angle: The right frame is testimony, not competence. The question is not whether the machine is clever but under what conditions its output may be taken on trust, and the answer here is precise: it degrades exactly where clinical judgment lives. A source that is reliable on recall and unreliable on synthesis is a genuinely difficult kind of witness, because its failures arrive in the same confident register as its successes. The specific numbers will date within months. The framing will not.