Abstract
Critical heat flux is a safety-critical limit in boiling systems because overprediction assigns thermal margin that is not physically available. Machine-learning studies of critical heat flux commonly validate random record splits, although public databases consist of dense parameter sweeps from individual experimental campaigns. Such splits may measure interpolation within familiar facilities rather than transfer to unseen experiments. This study asks whether reported accuracy survives campaign-disjoint validation, and whether anchoring a learner to a closed-form correlation improves transfer. A database of 24,579 measurements from 60 experimental campaign groups in uniformly heated vertical round tubes was analyzed using standard water and steam thermophysical properties, histogram gradient boosting and extreme gradient boosting regressors, and hybrids that learn the multiplicative logarithmic residual of the Bowring correlation. Campaign groups were withheld entirely from fitting; random-to-grouped degradation was quantified using a leakage amplification factor; and predictors derived from the measured outcome were excluded through a causal audit. The direct learner deteriorated from 4.69 percent random-split error to 11.91 percent under campaign-disjoint validation, a 2.54-fold increase, whereas non-trainable correlations remained near unity. The anchored hybrids achieved grouped error rates of 3.94 and 4.96 percent, with amplifications of 1.36 and 1.06, respectively. Across thirty held-out partitions, the anchored models usually retained lower amplification and grouped error than their direct counterparts and were insensitive to a six percent systematic anchor error. The contribution is a reproducible evidence framework for machine learning in complex boiling heat transfer.