It is 2015 in a Mountain View war room, and the postmortem is embarrassing: the machine-learned model at the heart of the product is a few hundred lines, but it is wrapped in tens of thousands of lines of glue — feature pipelines nobody owns, data dependencies nobody declared, a training set that quietly diverged from what the servers see in production. Accuracy fell for a month before anyone noticed, because no test covers "the world changed." Engineer ML as a system: detect training-serving skew, version data like code, monitor the input distribution, and make every consumed signal a declared, tested dependency. Get it wrong and the organization ships a model that was excellent at training time into a production reality that has already moved on.
Chose The physical-world adversarial attack — wrong. The adaptive-adversary audit was the one that fit.
Song's research on AI security and adversarial machine learning directly studies how production ML systems can be attacked or can fail silently when their operating conditions diverge from the assumptions baked in during training — a genuine technical cousin of this problem's training-serving skew, viewed through a security rather than an engineering-hygiene lens. Her work forces exactly the kind of vigilance this problem demands: do not trust a model's training-time performance as a guarantee of its behavior once deployed against a changing or adversarial world. She scores in the middle of the roster because her primary focus is security and robustness against deliberate attack, a narrower and more adversarial framing than this problem's broader concern with organic distributional drift, undeclared dependencies, and pipeline ownership.
Lamport's rigorous approach to specifying exactly what a distributed system does and does not guarantee is a genuine methodological cousin of this problem's demand to make every consumed signal a declared, tested dependency rather than an implicit, undocumented assumption — his career-long insistence on precise specification over hand-waving is exactly the discipline this problem's engineers lacked. His consensus and correctness work also addresses how distributed components can silently diverge from a coherent shared state, a loose structural echo of training-serving skew. But Lamport's actual technical domain was distributed-systems consensus and correctness proofs, not machine-learning pipelines, feature engineering, or statistical drift monitoring specifically, which keeps his relevance philosophical and methodological rather than directly technical to this problem.
Battle #119 · 8/10/2026, 11:38:51 AM · this result is deterministic: the same two personas on this problem always resolve the same way.