Validation and tolerances

Every algorithm is compared with Crystal-Builder or sci-form golden data. The tolerances below are the initial contract; a test that needs a wider tolerance must record the reason in the algorithm page.

Feature	Comparison	Tolerance	Phase
Element reference data	exact	0 (copied from Crystal-Builder)	1
Cell parameters	absolute	1e-6 Angstrom / 1e-6 degree	1
Cell volume	relative	1e-9	1
Fractional coordinates	absolute	1e-6	1
Atom sets	up to translation and permutation	1e-6 fractional	1
Bond length	absolute	1e-4 Angstrom	1
Bond order	absolute	1e-3	1
Bond set	canonical (i, j, image) after P1 atom mapping	exact	1
Bond length (full corpus)	absolute	0.05 Angstrom (inherits the P1 mapping tolerance)	1
Coordination numbers	exact	0	1
P1 atom mapping (golden)	absolute periodic distance	0.05 Angstrom (symmetry-standardized sites)	1
Ring size/count	exact	0	1
Minimum-image distance	absolute	1e-9 Angstrom	1
CIF cell parameters	absolute	1e-6 Angstrom / degree	1
CIF atom count	exact	0	1
POSCAR/XYZ round trip	absolute	1e-9 Angstrom	1
Neighbor distance	absolute	1e-9 Angstrom	1
Slab atoms per layer	exact	0	1
Ring paths	atom indices and images	exact after atom mapping, translation home and direction	1
Ring budget	too-dense refusal	exact parity with the golden error	1
Functional groups	atom sets	exact after atom mapping	1
Space group number / Hall number	exact	exact	2
Wyckoff letters	exact	exact	2
Standardized cell	absolute	1e-5 Angstrom / 1e-4 degree	2
K-path	points and labels	1e-6 fractional	2
PXRD positions	absolute	0.01 degree 2-theta	5
PXRD intensities	absolute after normalization	1e-3	5
UFF energy	absolute	1e-6 eV per atom	4
UFF forces	absolute	1e-4 eV/Angstrom	4
UFF stress	absolute	1e-3 GPa	4
Phonon q frequencies	absolute	0.5 cm^-1 (phonopy spring golden)	4
Indexing figures of merit	relative to the Python reference	1e-4	7
Indexed lattice (golden)	Q-spectrum agreement	5e-4 relative	7
Peak fit centre	absolute	2e-3 degree (synthetic Gaussian)	7
Pawley cell	absolute	2e-3 Angstrom (golden pattern)	7
Pawley zero shift	absolute	5e-3 degree (golden pattern)	7
Pawley intensity	absolute	3 esds	7
Pawley Rwp / chi^2	absolute	0.02 (golden pattern)	7
Le Bail intensity	four esds	90% of a noisy synthetic pattern	7
Rietveld position	absolute	3e-3 fractional (rutile golden)	7
Rietveld cell	absolute	2e-3 Angstrom (rutile golden)	7
Rietveld zero shift	absolute	0.02 degree (rutile golden)	7
Rietveld Rwp	absolute	0.25 (rutile golden, single-Gaussian model)	7
Space-group screen	exact	rutile P4_2/mnm among the candidates; 001 removes it	7
Space-group screen	exact	diamond Fd-3m among the candidates; 200 removes it	7
Extinction class	ranking	rutile P4_2/mnm is the first golden row	7
EQeq charges	absolute	1e-3 e against the authors' program on MOF-5	4
UFF energy	absolute	1e-4 kcal/mol against sci-form on methane and water	4
Charges	absolute	1e-6 e	4
Porosity	absolute	0.01 Angstrom / 0.1 percent	5
Topology	canonical graph	exact	5
SMILES round trip	graph and charges	exact after atom mapping	6
Embedding bond length	absolute	0.3 Angstrom for ring closure	6
MOF assembly	atom count and finite coordinates	exact/deterministic	6
Powder position	absolute	0.1 degree 2-theta	7
Rietveld scale/background	relative	1e-3	7
Structural refinement	Rwp	non-increasing over accepted steps	7
Instrument doublet	ratio	1e-9	7

Rules

• The comparison helpers live in xtalforge-testkit (Task 0.9.4).
• Golden files are generated by scripts/golden/ and never by CI.
• The independent end-to-end battery (tests/e2e/, scripts/e2e.sh) compares the WASM surface with gemmi, spglib, ASE and NumPy references and with committed 3D screenshots; its scope, findings and known gaps are in the audit page.
• A deliberate deviation from Crystal-Builder is recorded in the algorithm page with the reason and the affected structures.
• Tightening a tolerance is a normal change; widening one requires a note in the changelog and in the algorithm page.