Answers are grounded in this paper and cite their sources.
Literature-mined release curves can support model development while overstating applicability to an unseen publication. We test the consequences of changing the uncertainty unit from a release observation to a formulation and then an entire publication block. A unit-audited public PLGA dataset provides 1-, 3-, and 7day endpoints for 257 formulations from 99 DOI groups. Fixed single-threaded baselines are evaluated with five prespecified 60/20/20 publication train/calibration/test partitions. The primary outcome is simultaneous coverage of every eligible formulation endpoint within a held-out publication. Extra-trees MAE was 0.172 release fraction under publication splitting, versus 0.136 under formulation-random splitting; mean publication-split R2 was 0.002. Pooled-point calibration yielded 68% empirical publication coverage, compared with 88% for publication-max calibration. Mean widths were 0.583 and 0.773, respectively, on a physical 0-1 scale. Additional capacity, feature-block and learning-curve ablations use the same publication holdouts. Only 1.6% of harmonized endpoints coincide with a recorded source time, motivating a resolution audit. Publicationlevel calibration improves empirical simultaneous coverage at the cost of wide intervals; exchangeability and prospective performance remain unestablished.
No takes yet. Share an insight, caveat, or question.
Shereen Anand (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: