Belong: Repository Evidence Improves Go JSON Choice Inference

Code can compile and pass its tests while violating the repository's established serialization convention because the missing constraint lives in nearby precedents. We introduce Belong, a natural-label held-out benchmark for repository-conditioned Go JSON choice inference. Its source census contains 37,314 decisions from 8 public repositories. On 100 convention-ambiguous decisions with sparse local context, three language-model families predict both field name and presence policy under seven controlled evidence views. On this deliberately repository-balanced sparse-local stress sample, Original Raw raises semantic choice accuracy from 60.0% to 93.3%: +33.3 points [26.3, 40.7], with 107 rescues and 7 regressions. Exclusive Raw removes every target-file row, exact-field row, and target-identifier mention yet retains +21.7 points [15.0, 28.7]. A closest-available matched control supplies equally long and filtered evidence from other repositories; Exclusive Raw remains +19.3 points ahead [13.0, 25.7]. Original Raw therefore measures repository-assisted reconstruction; the no-copy contrast estimates the complete same-repository evidence construction rather than repository identity alone. After removing 20 targets whose non-JSON tags expose the name, Original Raw and the matched no-copy comparison retain +30.4 and +17.1 joint points, with both name and presence intervals above zero. On a separate 28-task holdout from four repositories created after the dated DeepSeek model identifier, a labeled exploratory Original Raw condition adds +16.1 points [1.8, 32.1], while the no-copy contrast does not transfer. Across 6,576 exact-label predictions costing $7.706, Belong provides a reproducible, training-free diagnostic of repository evidence for its evaluated Go JSON populations.