The figure-numbered gallery below uses the SIMBIOCHEM camera-ready numbering after the author-review revision. The three prospective panels are Panel I, Panel II and Panel III. The updated journal manuscript numbers the external-probe sweep as Supplementary Fig. S15, Table S19 and Section 14. Its Fig. S15 draws from workshop Fig. 5 and workshop Fig. S13, with complete inputs under distance_sweep. The S15 row below refers to the workshop's different transfer figure. The journal manuscript's Supplementary Figs. S16–S17 show the subsequent non-water geometry audit and separate contact follow-up. They preserve the original 72-case response score and add 36 new contacts.
| Journal display | Scientific question | Released experiment |
|---|---|---|
| Fig. 1 | Why a response field contains more spatial information than a dipole | Panel II example and response definition |
| Fig. 2 | What the constrained model learns from 48 configurations | Training and development |
| Fig. 3 | How the three prospective panels were selected and frozen | Panel I · II · III |
| Fig. 4 | Prospective response-ESP and dipole accuracy | Panels I–III |
| Fig. 5a,c,d | Chemistry, regime and per-solute breadth of the water-panel result | Panels I–III |
| Fig. 5b | Non-water contact follow-up, one geometry per solute and neighbour species | Frozen 36-case follow-up |
| Fig. 6 | Liquid-derived local environments | Liquid bridge |
| Fig. 7 | Frozen coupling to a held-out water | Held-out water |
| Fig. 8 | Nested water size and inducing-fragment separation | Shell size · dissociation · 50–100 Å extension |
| Supplementary Fig. S16 | Why the original 72 replacements were a geometry stress test | Complete 72-case audit |
| Supplementary Fig. S17 | Contact energies and all case scores in the corrected follow-up | Frozen 36-case follow-up |
The original 72-case non-water scores remain in journal Supplementary Tables S8–S9 and the original experiment folder. The follow-up was designed after that panel's geometry audit and does not replace its predeclared result.
| Figure | What it answers | Plot values | Raw experiment |
|---|---|---|---|
| 1 | Physical response, target and representation | CSV data | training |
| 2 | Architecture and development controls | CSV data | training |
| 3 | Prospective response-ESP and dipole comparisons | CSV data | panel_1 |
| 4 | Held-out-water energy, torque, orientation and rank | CSV data | heldout_water |
| 5 | Fixed-charge probe distance and exact moments | CSV data | distance_sweep |
| S1 | Reference construction and probe surface | CSV data | training |
| S2 | Representation reconstruction error floor | CSV data | training |
| S3 | All 56 prospective solutes | CSV data | panel_1 |
| S4 | Bootstrap and leave-one-solute sensitivity | CSV data | panel_1 |
| S5 | Predicted and QM spatial fields | See experiment / provenance | panel_3 |
| S6 | Comparator observable coverage | CSV data | panel_1 |
| S7 | Response as the inducing fragments separate, 3–100 Å | CSV data | 3–20 Å and 50–100 Å |
| S8 | Complementary W4 error and distance diagnostics | CSV data | heldout_water |
| S9 | Exact multipole hierarchy | CSV data | heldout_water |
| S10 | Torque across water rank | CSV data | heldout_water |
| S11 | Cumulative frozen coupling | CSV data | heldout_water |
| S12 | Response fraction of frozen electrostatics | CSV data | heldout_water |
| S13 | Quantum-water sweep and paired uncertainty | CSV data | distance_sweep |
| S14 | Prediction chronology | See experiment / provenance | panel_1 |
| S15 | Chemistry, water regimes and the 36-contact follow-up in panel b | CSV data | nonwater_contact · water panels |
| S16 | Liquid-derived local environments | CSV data | liquid |
| S17 | Size and representation controls | CSV data | shell_size |
| S18 | Contact-energy audit of the original 72 replacements | Case-level audit | nonwater |
| S19 | All energies and per-case response scores in the 36-contact follow-up | Case scores | nonwater_contact |
| Table | Contents | Data and definitions |
|---|---|---|
| 1 | Extension summaries, including the original non-water challenge and separate contact follow-up | nonwater · nonwater_contact |
| S1 | Absolute response scale | panel_1 |
| S2 | Original 48 training configurations | training |
| S3 | Prospective identities | panel_3 |
| S4 | Comparator definitions and provenance | panel_1 |
| S5 | Complete prospective leaderboards | panel_1 |
| S6 | Regime-specific prospective and liquid errors | liquid |
| S7 | Post hoc water-contact sensitivity of the prospective panels | water_contact_audit |
| S8–S9 | Original 72 unrelaxed non-water replacements: aggregate and species comparisons | nonwater |
| S10 | Liquid identities | liquid |
| S11 | Same-supervision AIMNet2 control | training |
| S12–S13 | Nested size results | shell_size |
| S14 | Controlled global-branch ablation | global_branch |
| S15 | Held-out-water identities and structures | heldout_water |
| S16 | W4 coupling | heldout_water |
| S17 | Exact multipole control | heldout_water |
| S18 | Frozen ensemble-member robustness | panel_1 |
| S19 | Reported artifact hashes | training |
| S20 | Distance–orientation sweep | distance_sweep |
Unique historical derivation tables, including detailed development-control and ensemble-member analyses, are retained in provenance. They are supporting analysis records rather than a second set of authoritative experiment data. The filename map explains old plotting identifiers.