Budgeted Two-Stage Pre-Validation for Retrieval-Grounded Machine Learning Pipelines: A Group-Held-Out Evaluation

Retrieval-grounded machine-learning pipeline synthesis must reconcile code templates with package versions and resource budgets. We study a two-stage pre-validator adapted from concrete dependency-satisfiability and resource-prediction mechanisms: an inexpensive retrieval-profile margin ranks candidate manifests; a budgeted second stage invokes a gradient-boosted peak-memory predictor and applies bounded repairs. Actual execution of 48 public-dataset learning workflows supplies memory profiles, while eight controlled configuration mutations per workload define explicit local feasibility outcomes. Four group-disjoint test folds provide 384 held-out candidates. At an exact 50% second-stage invocation budget, retrieval-prioritized validation yields 372/384 feasible configurations (96.9%), matching exhaustive checking, versus 281/384 (73.2%) for one matched random schedule. Across 300 random schedules the mean is 74.2%. A 16-group paired bootstrap puts the advantage over that seeded random baseline at 19.8–27.1 percentage points. Uncertainty offsets and semantic novelty are not independently beneficial on this coarse-fault benchmark: point prediction and a simple memory-margin ranking tie the proposed method. Calibration also deteriorates on a shifted fold. These findings support auditable, low-cost screening for synthetic configuration faults, without treating local feasibility as actual cloud deployment success or deterministic template synthesis as LLM generation.