Imagine you're growing a bonsai tree using a rule that says "stop pruning when the shape stops improving." Theory tells you bigger pots (more soil) should grow bigger trees. You run the experiment five times with pots of increasing size — and the trees come out random sizes, with no pattern. The pot-size rule is real (bigger pots never shrink the tree below some minimum), but it tells you almost nothing about what you'll actually get. That's this paper. The committed claim: when you combine Q-FLAIR's adaptive circuit-growing mechanism with Caro et al.'s generalization bound (which says fewer trainable gates need less data), you get a valid but useless guarantee. The bound holds in 14 of 15 runs — it's a real upper limit — but the correlation between the bound's value and the actual generalization gap is r = 0.12. That's noise. The number of active gates K does not explain the variation you observe. The experimental setup is narrow but honest. They reimplement Q-FLAIR faithfully on full-resolution 784-pixel MNIST 3-vs-5 classification, run it at five training-set sizes (N = 2,000 to 10,000), and fine-tune each resulting circuit to measure Caro et al.'s notion of active gates. The result is a negative finding: circuit size and test accuracy both vary non-monotonically with N, and seed-to-seed variance is nearly as large as any trend across training-set sizes. The adaptive growth mechanism does not converge to predictable circuit depths as data scales. This matters because the field's implicit hope is that quantum machine learning will eventually have the kind of predictable scaling laws that classical deep learning enjoys — bigger model + more data = better performance on a knowable curve. This paper shows that for at least one well-studied adaptive architecture, that scaling law does not emerge. The bound from Caro et al. is mathematically correct but operationally vacuous: it guarantees you won't exceed a ceiling, but the ceiling is so far above the floor that knowing it doesn't help you plan. The integrity picture is mixed. The authors faithfully reimplement an existing method rather than proposing their own, which is good. They use a real dataset (MNIST) rather than a toy problem. But 15 runs across 5 training-set sizes is a small sample, and the seed-to-seed variance they report actually undermines their ability to detect any trend that might exist. No code release is mentioned. The validation is purely simulation — no quantum hardware in the loop. The honest read on what's missing: they test only one adaptive growth strategy (Q-FLAIR) on one dataset (MNIST 3-vs-5) with one encoding scheme. The obvious next experiment is to repeat this with alternative adaptive architectures — ADAPT-VQE, evolutionary circuit search, or reinforcement-learning-based growth — to see whether the non-monotonicity is a Q-FLAIR artifact or a deeper property of adaptive quantum circuits. The authors likely ran out of compute budget rather than hiding failed experiments; the paper reads as a careful negative result from a small team. The paper is worth reading not for what it proves but for what it honestly fails to prove. Joint optimization of circuit depth and training data in quantum ML remains an open problem, and this work narrows the path by showing that one plausible combination of tools does not yield the scaling law the field needs. The question it raises — why valid guarantees can coexist with zero predictive power — is genuinely important for quantum ML theory.
