README.md

Choose an experiment

Start from the scientific question, then open the corresponding guide. Raw array filenames are stable identifiers resolved by their registries. The data dictionary explains units and schemas.

Experiment	Configurations	Purpose
Training	48	The only response labels used to fit the published checkpoint.
Panel I	48	First prospective solute test and retained joint-gate failure.
Panel II	96	A separately selected 24-solute test.
Panel III	80	Twenty FreeSolv identities, without hydration values.
Water-contact audit	224 screened, 9 flagged	Eight short solute–water contacts and one separate water–water contact; original panel scores remain primary.
Non-water geometry challenge	72	Frozen NH₃, CH₃OH and CH₃CN replacements, with a post hoc QM contact audit.
Non-water contact follow-up	36	Separately frozen, fixed-solute MMFF94s contact geometries with new QM references.
Liquid	24	Four explicit-water configurations per solute.
Shell size	36	Nine parents evaluated with 1, 3, 6 and 12 waters.
Held-out water	10 bases, 90 outer waters	Frozen response coupling and directional observables.
Pilot	6 bases	Original downstream protocol-development test.
Distance sweep	10 bases, many placements	Spatial range of frozen-field reuse.
Global branch	224 evaluation cases	A matched post hoc training intervention.
Dissociation	224	Move the inducing fragment itself to 3–20 Å.
Extended dissociation	56	Follow-up at 50 and 100 Å.
Dissociation QM check	20	Direct complex-minus-fragments QM references for one solute across the same 3–100 Å trajectories.

Prospective solute transfer

Panels I–III contain 56 different solutes and 224 three-water geometries. Their original GLIDER predictions preceded their response QM. They all evaluate the same frozen checkpoint. Panel I was subsequently reused for a different model family, which is clearly separated in data lineage. The original checkpoint was not refit.

All reported results remain, including the unsuccessful joint gate and the later model that failed to improve.