Matched Small Clinical Encoders Do Not Support Hybrid MLM+JEPA Superiority

Compact biomedical and clinical encoders are useful when privacy, auditability, and local deployment constrain model scale, but pretraining-objective comparisons at this scale are easy to confound. Token budgets, auxiliary losses, probe settings, and restricted-data governance can change the apparent winner. We compare six pretraining objectives – span masked language modeling (span-MLM), autoregressive next-token prediction (AR), Hybrid MLM+JEPA, pure JEPA, contrastive learning, and ELECTRA – in one controlled approximately 42M-parameter Transformer family. All objectives use the same corpus/split, tokenizer, architecture, seed IDs, checkpoint schedule, frozen-readout protocol, and common FLOP ledger for auxiliary objective costs; real-token exposure varies by objective under iso-FLOP accounting and is analyzed separately with iso-token sensitivity curves. The locked final-test analysis does not satisfy the pre-specified superiority criterion for Hybrid MLM+JEPA over span-MLM: the PMC-Patients Patient-to-Patient Retrieval (PPR) task instead favors span-MLM after Holm correction (diff = -0.0092, 95 percent unadjusted crossed-bootstrap CI [-0.0185, -0.0029], two-sided exact sign-flip Holm p = 0.0078), while PubMedQA, BC5CDR, and MedNLI remain statistically unresolved. Secondary analyses, including development-only sensitivity curves and PMC-Patients Patient-to-Article Retrieval (PAR), are consistent with readout-specific objective sensitivity, especially for token-level BC5CDR, but the confirmatory result is the failure of the Hybrid MLM+JEPA superiority criterion.