What the experiments show

When resampling the 64 solute groups, the conditional 95% interval for water-force error change is −17.7% to −8.5%. For solute forces it is −2.0% to +1.7%, so improvement is not demonstrated. Relative-energy change has interval −20.8% to −3.6%. These intervals describe this fixed comparison, not variation across training seeds or all biological chemistry.

A difficult phosphorus-containing training configuration improved from about 54,000 to 6.56 eV/Å force error. It was used for training and remains inaccurate; this is not evidence of generalisation. The original model also failed on it, and the source labels were retained.

Water labels refer to the source topology of finite quantum clusters. They do not establish bulk salt concentration, pH or physical reaction times. New evaluation charges are −1, 0 and +1; existing dianion checks remain separate. Original model pretraining exposure is possible.

The old molecular energy regressions mostly concern shifts between molecular groups, while within-group variation barely changed. Next, check energy calibration using training data only, with network weights and forces fixed. No correction has been applied and no candidate has been promoted.

This is one seed and one fixed training budget, not a universal architecture ranking. Initial record identities, targets and schedule match; small float32 differences in recomputed predictions are quantified in the download. Tiny solution differences should not be treated as conclusive model superiority.

The three native quantum references retain separate energies and force targets. The control’s auxiliary losses were checked at every update to ensure they could not alter shared parameters. No new quantum calculations were generated.

Keep the existing joint model as the experimental incumbent. Next, measure actual coupled GPU simulation throughput and extend fixed-model solution validation before a larger training release. Reaction rates, unseen biological chemistry and production deployment remain unqualified.

The earlier PME probes retained cutoff energy jumps. Charge-pair crossings explained the saved directional-derivative discrepancies. This comparison uses the authors’ smooth reaction-field helper, unchanged Amber/TIP3P parameters and explicitly recorded Lennard–Jones switching. The learned weights and starting state were unchanged.

This is a distinct classical Hamiltonian from the earlier PME tests and uses a different water model from the full published TIP4P-FB setup. The earlier failed convergence observations remain available. Only one starting solvent environment and 50 femtoseconds per run have been tested.

Both models passed the authors’ force-interface fixture: quantum and solvent energy derivatives, translation, periodic image changes, cubic-axis rotation, cached neighbour lists and scripted execution. This small fixture does not establish stable dynamics.

The full solvated-system CPU derivative probe failed. A classical-only check on the same saved state isolated finite-difference precision: the double-precision Reference backend passed the unchanged tolerance. The failed probe remains recorded; no reaction-rate or production-model qualification follows from these checks.

Deployment uses the authors’ reduced polarization and periodic electrostatic cutoff; native quantum-reference comparisons use the original reference convention. These settings are kept distinct. In the earlier PME setup, both models completed matched 50-femtosecond runs. The smaller timestep increased maximum energy deviation by about 49–52%, so those runs did not establish timestep convergence. The separately reported smooth reaction-field comparison addresses this mechanics issue; longer-time stability remains unqualified.

The overall solution QM-force change is small: −0.49%, with a conditional 95% interval of −1.33% to +0.22% when resampling 116 linked validation groups. It is not a conclusive improvement. The solvent-response change is −1.25%, with interval −1.95% to −0.69%.

Reaction relative-energy error fell 7.29%, although reaction-force errors rose slightly. Small-molecule energy error was nearly unchanged (+0.03%). Tripeptide energy and force errors decreased. These tests do not measure free-energy barriers or physical reaction rates.

Native quantum methods, geometry and force targets remain separate. Source-linked validation groups are protected; 1,342 records are explicitly omitted from this experimental training view and retained in the source bank. The three known numerical failures and their surrounding components are preserved.

Missing native C6 supervision is explicitly disabled in this experiment. Released-model pretraining exposure remains; these are not globally unseen-chemistry tests. The earlier standalone calibration is retained separately, and its learned outputs were not used to initialize this joint run.

These are full-quantum PubChem molecular examples used to integrate public reference sources. They add no new solution reactions. The wider explicit-environment BMS25 data and its quantum/solvent-force components remain separate.

Both sources use their native geometries and labels. Numerical energy offsets depend on element counts and charge and were fitted only on training records; they are not isolated-atom or formation energies. Forces are energy derivatives and errors are per Cartesian component.

The experiment uses a frozen released representation, two method-specific outputs, three fixed epochs and no repeated tuning. Released-model pretraining exposure is not erased by the project split. Missing SPICE spin metadata is preserved. Reactive dynamics, salt response and biological reaction rates remain unqualified.

All three tripeptide configurations contain very short H–H contacts. Their model predictions have nonfinite molecular forces. They remain in the source data; they have not been deleted or relabelled as bad quantum references.

TIP3P water with one sodium/chloride pair in a 3 nm periodic box; 298.15 K, approximately 0.0615 mol/L salt. Each trajectory has 100 ps equilibration and 2 ns production with a 2 fs integration step.

The checks cover sodium self-diffusion statistics at 5, 10, 20, 40 and 80 ps. Uncertainty resamples whole independent trajectories. The reference uses a classical potential and Langevin thermostat.

This is not a reaction-rate measurement or experimental diffusion validation. Protein motion, ion association, membranes and organelle-scale transfer remain untested.

Peptide relative-energy error improved, but force errors worsened for both peptides and miniproteins, and miniprotein relative-energy error increased. Separately, both models fail on an exposed high-energy tripeptide configuration. Finite results on these 899 examples do not establish numerical stability everywhere or validate reactive dynamics.