It is the era when banks first automate the lending decision, and a regulator will read the model line by line. Predict who will default from a handful of financial features — but with a model whose output is a genuine probability and whose coefficients carry meaning: each the change in log-odds per unit of a predictor, defensible in a hearing. A pure black box may rank borrowers a hair better yet cannot answer "why was I denied," which the law increasingly demands. Fit the model, interpret the weights, and be honest about which are causal and which are mere proxies. Get it wrong and you fail the audit, or worse, launder a forbidden variable through a correlated one.
Chose The causal diagram and do-calculus — right call.
Pearl's causal inference framework, the do-calculus and structural causal models, directly addresses this problem's sharpest warning: a regression coefficient can be a genuine causal effect or merely a correlated proxy for a forbidden variable like race or zip code, and only causal reasoning can tell the two apart. His work gives regulators and modelers the formal tools to ask which of a model's predictors would still matter under intervention, exactly the audit this problem demands. He did not work on logistic regression or credit scoring specifically, and his framework is more often applied after a regression than embedded within one. His relevance is conceptually central to the hardest part of this problem, if technically once removed from fitting the model itself.
LeCun's convolutional networks, developed for digit and pattern recognition, are deep, distributed, gradient-trained architectures whose internal weights are not individually interpretable as marginal effects, essentially the opposite structure from the log-odds coefficients this problem demands. He has been a vocal advocate for end-to-end differentiable learning over hand-built or simply interpretable structure, the reverse instinct from what a bank regulator needs. He would have encountered logistic regression as classical 1980s-era statistics background but never contributed to generalized linear models, maximum likelihood coefficient fitting, or regulatory interpretability. His toolkit answers a fundamentally different question than this problem asks. His convolutional lineage was built to discard exactly the structure this problem needs kept.
Battle #144 · 8/10/2026, 11:40:14 AM · this result is deterministic: the same two personas on this problem always resolve the same way.