Ternary-Aware Evaluation Protocol (TAEP): filtered molecular-glue benchmark, per-pocket measurements and analysis code

This record holds the data and code behind the paper "Molecular glue generation has no shared pose-quality metric, and the ones in use cannot see the protein". WHAT IS HERE - The Ternary-Aware Evaluation Protocol (TAEP): five ordered pose-quality checks for ternary complexes, each shipped with the control that falsifies it. - The filtered 94-pocket benchmark, as the pocket identifiers and the selection criteria that produced them, together with the ipTM calibration that justifies the cut. - The independent 18-structure PROTAC stratum, likewise as identifiers and criteria. - The split index files, including the decontaminated subset and the grouped split. - Every per-pocket measurement reported in the paper, as machine-readable JSON. - The analysis code. Each script carries positive and negative controls behind a --self-test entry point and must fail on deliberately corrupted input before its output is used. WHAT IS NOT HERE, AND WHY The generator under study is not released, in weights or in training code. The protocol is the contribution and it runs on any generator; the rows of the paper's tables that describe this particular generator therefore cannot be regenerated from scratch by a reader, and we say so rather than implying otherwise. Every protocol check, every split index and every released measurement can be re-run and re-checked independently, and so can the two third-party models, whose code and weights are public. Third-party structures are not redistributed here. The benchmark is defined by identifier lists and selection criteria; the underlying structures come from their original sources: MG-PDB and MGBench (doi:10.1021/acs.jcim.5c01860), MGTbind (doi:10.1093/nar/gkaf1075), CrossDocked2020 (doi:10.1021/acs.jcim.0c00411), and TernaryDB, released with DeepTernary (doi:10.1038/s41467-025-61272-5). Redistributing them would put this record under their licences rather than ours. LICENCES Data are released under CC BY 4.0. Code is released under the MIT License, whose text is in the code directory. The record-level licence field is set to CC BY 4.0 because Zenodo takes one; the MIT licence governs the contents of code/. HOW TO START Read README.md first. It maps every quantitative claim in the paper to the script that produces it and the artifact that holds it.