Heterogeneous Returns to Balance-Sheet Debt Expansion: A Pre-Registered Causal Machine Learning Pipeline with a Calibrated Detection-Frontier Benchmark

We pre-register an end-to-end causal machine learning pipeline — double machine learning, honest causal forests, DR-learners, and budget-constrained policy learning — to ask whether the operating-performance response to significant balance-sheet debt expansions is heterogeneous across U.S. public firms, and whether that heterogeneity is exploitable for capital allocation. Using 15,971 firm-years built from as-reported SEC EDGAR statements (dimensional-fact filter audited against the SEC XBRL API: 80 data points, 0 discrepancies), the average effect is statistically indistinguishable from zero (IRM ATE 0.00153, SE 0.00422) and survives eleven pre-registered ablations. In a calibrated semi-synthetic benchmark the causal forest recovers treatment-effect structure with 6–9× lower PEHE than naive learners, yet — consistent with our pre-registered detection-frontier analysis — no feasible test certifies heterogeneity at empirically plausible magnitudes in the held-out panel. A budget-constrained CATE-ranking policy attains the least-negative doubly-robust value under the 2021–2023 rate-hike holdout and outperforms every naive rule and artifact baseline in point estimates across all seeds, but no allocation policy — including observed market behavior — is statistically distinguishable from random assignment at feasible sample sizes. Model-implied targeting concentrates on small, B2B-oriented firms with — contrary to our pre-registered financial-constraints hypothesis — higher liquidity, a pattern that is stable across seeds but cannot be causally certified at available power. The calibrated benchmark, the truth-validated evaluator, and the frontier itself are the primary contribution: an evidence standard for heterogeneity claims in corporate finance.