Validation

Every Claim Has a Number Behind It

Cross-engine agreement, published reference values, exact-answer tests and a list of what is still unverified.

Two scientists at a meeting table comparing printed result charts and a laptop showing two nearly identical curves

Cross-Engine Agreement

Test-only referees, run in a separate environment on every engine release.

Counts from the 2026-10-18 engine handover record. The GPU against CPU row is from the CUDA test on engine 0.6.0.
Module	Referee	Agree	Differ	Notes
Quantum Chemistry	PySCF	419	0	HF energy 1e-8 Eh, gradient 1e-6 Eh per bohr
Quantum Chemistry	Psi4	93	0	Same tolerances
Cheminformatics	RDKit	906	1	Sulfonic acid logP, item SI-03
Cheminformatics	Open Babel	597	3	Documented differences
GPU Against CPU	Our CPU path	54	0	2.0e-12 Eh, also PySCF and Psi4 to 2.4e-11 Eh

Exact Answers and Published Values

Water HF/STO-3G optimised at -74.9659011923 Eh with frequencies 2170.0, 4140.0 and 4391.1 cm-1, matching published 2170, 4140 and 4391
Methane -39.7268636795 Eh; helium -2.8077839575 Eh within 4.25e-8 of published
Analytic gradients against finite differences: 1.3e-9 and 9.6e-10
Point groups: 27 of 27 NIST CCCBDB reference molecules
Constants match CODATA 2018 to the printed digits
MMFF94 energy terms: 761 molecules in the published suite, none outside the rounding bound
Harmonic, Morse, hydrogen and Eckart tests at 1e-10 to 1e-14

What 493 Self-Tests Mean

435

Passed

Pending source check

Failed

Engine 0.7.11 on the CPU server. The GPU server runs extra CUDA checks: 444 passed, 58 pending, 0 failed.

Pending, Unverified and Different

Rows we do not count as verified, stated up front.

Data Awaiting Papers

Parts of the Crippen and TPSA tables and some Model Systems defaults still need checking against the original publications. STO-3G was verified in engine 0.7.11, the D2 table by an independent three-way check, and element weights and masses (CIAAW 2024, AME2020, hydrogen to xenon) in engine 0.7.13, rolling out. Results flag them as unverified.

Known Differences

Sulfonic acid logP against RDKit; 3D coordinates up to 0.026 Å across compilers; Crippen types for O7, N14 and metals not implemented.

Two Registers

Science issues (SI items) and a Capability Gap Register of 51 gaps, each with a severity, a priority, a plan phase and dated sources.

Validation Questions

What does PEND mean in a self-test?

PEND means the value agrees within tolerance but the data table behind it has not yet been checked against the original paper. We count these separately and never as passes: 58 checks are pending today.

Are the cross-engine tests part of the product?

No. PySCF, Psi4, RDKit and Open Babel run in a separate test environment. Nothing from them is bundled or called by the engine.

How big is the tolerance?

Hartree-Fock energies are compared at 1e-8 Eh, nuclear repulsion at 1e-10 Eh and gradients at 1e-6 Eh per bohr.

Where are the open problems listed?

In two registers: the science issues register (SI items for unverified data and known differences) and the Capability Gap Register inside Chemistry Lab, with 51 gaps across theory, data, software, validation, scale, integration, hardware and usability.

Run Your Next Molecule Through AiMolSim

See your own molecule optimised, its frequencies computed and the checks printed beside the result. A demo takes about 30 minutes.

Get Started Request a Demo