Bounded PBC 0.18 qualification configurations

This inventory records the demonstrated PBC configurations selected for the bounded release acceptance in issue #557. It does not expand that release scope or close the CRYSTAL23 parity programme (#593). Method availability, input matching and successful SCF termination each establish less than a matched numerical reference comparison.

The machine-readable source is pbc_018_supported_configurations.json. It contains seven converged calculation rows, a separate refusal boundary, and explicit exclusions. This is an evidence inventory, not an executable run_suite.py matrix or a blanket supported-method declaration. Each row’s claim identifies exactly what passed. The runner’s #557 evidence note maps evidence_key to immutable calculation archives, source/native identities, queued jobs and original judgements. Operational receipts remain outside the product repository.

Included rows and measured results

Energies are Hartree per cell. All seven rows are neutral, three-dimensional, zero-temperature singlets, with unreduced meshes. The unrestricted ECP rows have seven electrons in each spin channel. No row here establishes magnetic, low-dimensional, symmetry-reduced or SKALA acceptance.

Row ID	Method and basis	Passed comparison	Measured cost
lih-rsgdf-gamma	RHF, STO-3G, SR/LR GDF, explicit metric cutoff 1e-13	E = -8.335327447080662; matched PySCF 2.6.2 GDF error -1.31490e-9, below 1e-7	1317.001 s including setup; 7 iterations
lih-rsgdf-222	Same LiH configuration, 2 x 2 x 2 mesh	E = -7.922007391408032; matched fitted reference error -4.22104e-9, below 1e-7	11063.026 s including setup; 8 iterations
diamond-gamma-pbe0-xc	Restricted PBE0, def2-SVP, Gamma GDF	E = -75.08648342894216; 9.91167e-8 from the converged XC-domain reference; T2 splitting 3.33738e-6; both below 1e-5	2268.98 s for diamond plus diffuse-cell check; separate SCF timer unavailable
gpw-rhf-h2-9	RHF, H2/STO-3G, 9-bohr cube, 24 cubed grid	Same-model energy/Fock derivative error 4.90266e-11, below 2e-9	0.6743 s for SCF plus directional derivative
gpw-rhf-h2-16	Same model, 16-bohr cube, 48 cubed grid	Derivative error 4.50180e-11, below 2e-9	0.4604 s for SCF plus directional derivative
cl2-uks-ecp-gamma	Unrestricted PBE, Cl2/LANL2DZ, 20-bohr cube	Bond-gradient error 6.09538e-9 Ha/bohr, below 2e-6; molecular energy offset 6.02557e-6, below 1e-5	Shared two-row command: 378.71 s, including displaced SCFs, gradients and molecular controls
cl2-uks-ecp-311	Same ECP configuration, 3 x 1 x 1 mesh	Bond-gradient error 6.10973e-9 Ha/bohr; same energy-offset gate passes	Same shared command; do not count twice

LiH used four CPU slots. Diamond used eight; its whole-command CPU time was 17421.98 s and peak RSS 457.25 MiB. The ECP command used four slots, 669.70 s CPU time and 5165.26 MiB peak RSS. The original ECP pytest receipt separately reports 432.697 s. These timers have different scopes and are not interchangeable. GPW used four numerical threads. These are individual observations, not a same-hardware speed comparison with CRYSTAL23. LiH’s original combined job completed Gamma before timing out during 222 setup; the separately resumed 222 calculation passed on the same immutable runtime. The timeout remains part of the evidence.

Exact inputs and support conventions

The JSON gives Cartesian atoms, lattice matrices in bohr, actual fractional reciprocal k coordinates and their normalized weights. Its convention is k_cart = 2*pi*inverse(A).T @ k_fractional, with lattice vectors in columns. Gamma is (0,0,0) with weight one. The 222 mesh is every triple from {0, 1/2} with weight 1/8; 311 is (0,0,0), (1/3,0,0), (2/3,0,0) with weight 1/3. These coordinates are resolved from the archived callers and pinned mesh implementation. They are not an attestation of a CRYSTAL output mesh; matching SHRINK dimensions alone remains insufficient (#645).

LiH has conventional fcc parameter 7.7158 bohr, Li at the origin and H at (a/2,a/2,a/2). Both rows explicitly use def2-svp-jk, SR fitting, splitting parameter 0.4 bohr inverse, raw reciprocal precision 1e-10, metric cutoff 1e-13 and all 69 fitting modes. The one-electron domain is pair-complete; its printed radius is 40.0925 bohr. Printed fitting AO/auxiliary radii are 81.42/64.52 bohr and the LR cutoff is 22.66 Ha. Those rounded diagnostics must not replace the source-pinned automatic support policy. SCF energy and commutator tolerances are 1e-11 and 1e-8. J and K are fitted, with neutral 3D Ewald electrostatics and the BvK Ewald exchange correction. This comparison does not remove auxiliary fitting error or qualify a universal metric policy.

Diamond uses fcc parameter 6.740653086540123 bohr and atoms at the origin and (a/4,a/4,a/4). The automatic XC density radius is 18.501284106877893 bohr. The periodic Becke/Gamma-torus grid uses 75 radial points and a 17 x 36 product angular rule, with partition image radius 10 bohr. The comparison is with the same-method converged 25-bohr domain energy, not a CRYSTAL result. A support-planning replay on the original source and native binary recovers the missing non-XC domains:

Diamond domain	Resolved value
One-electron AO pairs, pair-complete support	23.532024462586605 bohr
One-electron planning enclosure, before overlap screening	34.86711490084021 bohr
Fitting AO images	43.9975328815882 bohr
Fitting auxiliary images	57.98878984898395 bohr
Fitting LR reciprocal cutoff	21.031901088329146 Ha
Nuclear cutoff / V_ne reciprocal cutoff	25 bohr / 200 Ha

The enclosure is not the executed one-electron cutoff. The replay also reproduces the accepted XC radius exactly. These values complete the recorded configuration; they do not establish arbitrary-cell convergence.

The later continuous-grid check (qc465-continuous-grid, source f1ad7714) uses the same frozen diamond configuration and the original 25-bohr reference of -75.08648352805882 Ha. Its converged energy is -75.08648292670634 Ha, absolute error 6.01352e-7 Ha and T2 splitting 3.21507e-6 Ha. Both original 1e-5 gates pass and the XC radius is unchanged. This supplement establishes affected-path source linkage; it does not replace the original result, its diffuse-cell check or its measured cost, or qualify the final release binary. The JSON records it separately under followup_checks.

The GPW rows retain the adjoint-symmetric legacy Gamma molecular-limit exchange model, AO-image cutoff 25 bohr and Ewald nuclear convention with the FFT Poisson zero mode omitted. Their accepted carrier calls run_periodic_rhf_gpw without a functional: these are RHF, not PBE results. The finite-difference direction and step are recorded in the matrix. Fixing the derivative of this energy model does not establish physical bulk GPW exchange (#534).

The ECP rows place Cl2 at the centre of a 20-bohr cube with bond length 1.99 angstrom. LANL2DZ removes ten core electrons per atom; nuclear and Hartree electrostatics use the effective charges +7,+7. PBE has no exact exchange contribution. The archived fixture used pair_complete_1e=False and automatic truncation of the legacy cell-ball support. A support-planning replay on the original source and native binary recovers a final AO radius of 15 bohr and nuclear radius of 25 bohr. Both meshes and every bond displacement in the two-step finite difference accept those starting radii in one optimizer evaluation, with minimum overlap eigenvalue above 0.05124. The Schwarz thresholds remain 1e-12 for integrals and 1e-14 for derivatives. This recovers metadata for the captured source-matched preview; it does not qualify a later physical-support fixture with different domains. The molecular control shares this code’s XC implementation; it is not independent ORCA/CRYSTAL force validation. Earlier #35 single-point route/refusal and metadata checks remain accepted under their original scope, without being promoted to general ECP gradients (#579).

The ECP derivative consumer now also requires a receipt binding the terminal spin densities, orbitals, energies and occupations to the SCF source. Edited states and contradictory ECP flags are refused. Clearing both the receipt and the flags still refuses a basis with an applicable ECP requirement or sidecar potential; light all-electron atoms in the same basis family remain allowed. Rerun the same configuration when using older results without this terminal-state receipt; the historical evidence above retains its original source and numerical scope.

Refusal and exclusions

Issue #652 is excluded specifically for dense translated NaCl/LANL2DZ on the GDF RHF/UHF routes and for general projector-tail completeness. Its historical 25-bohr home-centered projector sphere missed partners when the AO enclosure grew to 63.0539 bohr. Main contains the physical midpoint-domain repair, but a recovered RHF example does not close the remaining shared UHF acceptance or the density-grid artifact failure. The frozen matrix has no NaCl row. Its only ECP rows are the Cl2 Ewald UKS preview with explicitly retained legacy support, AO radius 15 bohr and projector radius 25 bohr. Their evidence cannot qualify the NaCl/GDF route. This is an inventory exclusion, not a claim that every such input is refused at runtime.

The two #179 LiH rows retain their original source 63031b08, matched PySCF 2.6.2 references, explicit metric cutoff 1e-13, 69 fitting modes and 1e-7 Ha tolerance. Their scoped PASS does not close the broader #179 issue or establish default-metric, dense-cell or unfitted accuracy.

No #531 BIPOLE cancellation row is included. In particular, the 100-bohr Ne residual of -9.208e-8 Ha fails the unchanged abs(delta_E) * L / N_e**2 < 1e-8 gate. The 50-bohr observation alone does not qualify cancellation. The remaining programme must bound the combined omitted EE, EN and NN contribution, including charge amplitudes and image multiplicities, and retain both molecular-limit rungs and a non-vacuum origin/atom-representative check. This inventory exclusion does not alter a kernel, weaken the gate or require that programme to close for the included configurations.

The dense-mg-rijcosx row is a qualified refusal (#555): Mg primitive fcc, conventional parameter 4.51 angstrom, STO-3G, Gamma RHF, dense exchange-image rank three. Direct and public entries refuse before integral construction. The retained three-check boundary suite passed in 4.23 s. There is no converged energy, fitting/XC result or corrected dense-cell COSX claim.

No #531 BIPOLE cancellation configuration is included. The 100-bohr Ne residual of -9.208e-8 Ha still fails the unchanged coefficient gate abs(delta_E)*L/N_e^2 < 1e-8. The narrow 50-bohr observation does not qualify the larger-domain claim. The existing formulation analysis follows Dovesi et al., Physical Review B 28, 5781 (1983), section II, doi:10.1103/PhysRevB.28.5781: control the combined electron-electron, electron-nuclear and nuclear-nuclear Coulomb error with physical, relabelling-invariant supports. A monopole estimate or an enlarged radius alone is not that certificate. This remains separate research; the umbrella need not close to retain the other rows.

The #573 restricted starter matrix’s 28-bohr overlap-support check is not accepted total-energy parity. #648’s input/ECP matching tools do not replace actual reference operators, k coordinates/weights, grids, domains, zero modes and numerical results. Those obligations and broader HF, DFT, hybrid, SKALA, spin, dimensionality and performance coverage remain in #593.

MANTA and remaining runner work

Reuse the independently owned MANTA frozen configurations and measurements and control guide. The four profiles are image-tight, ordinary-rhf, global-pbe0 and compact-tight, each with its fixed-source control. Their exact geometry, custom bases, mesh shifts, physical supports, fitting and XC settings are serialized by benchmarks/manta_release_controls.py:configuration. That producer remains the single source for those rows. Component certificates cannot be added into a relaxed total-energy certificate, and no symmetry or bulk parity claim is introduced here.

The last reviewed candidate omitted the already accepted image-tight allowance commit ebbf5957bf93bf1ea7af3bbaef954e3e0fa5f729. The existing candidate inclusion check records the required bridge. This inventory does not replace the candidate owner’s inclusion check or authorize a release.

The runner can reuse the seven scoped PASS rows and dense refusal. The two support-radius metadata gaps are resolved by separate queued planning replays, with original native digests and relevant Python module hashes checked before execution. The matrix links these supplements by support_metadata_evidence_key; the #557 evidence note retains the immutable recovery receipts. These are deterministic support reconstructions, not newly observed SCF results. Original energies, gradient checks, tolerances, costs and source identities remain unchanged. The continuous-grid diamond supplement and the ECP terminal-state checks are separate from that metadata recovery. Release qualification and candidate identity remain with the loop. The broader #593 programme remains separate.