MANTA controls for the bounded 0.18 release

The approved 0.18 scope qualifies existing MANTA controls; it does not require completion of the full MANTA design. MANTA remains optional and experimental. The examples here use neutral, all-electron three-dimensional cells, fixed occupations, unreduced regular meshes, RI-J and exact full-Coulomb K with the explicit BvK Ewald convention. Physical supports remain fixed during SCF. The public entry point is vibeqc.run_periodic_job with the existing typed MANTA options. No default, symmetry permission or Fock convention changes.

Frozen examples

The source-tree qualification driver is benchmarks/manta_release_controls.py. It reuses the accepted fixtures rather than introducing another physical producer. Run it from a matching development environment with NumPy, SciPy and pytest installed, through the calculation queue for scientific acceptance:

PYTHONDONTWRITEBYTECODE=1 OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 \
VECLIB_MAXIMUM_THREADS=1 MKL_NUM_THREADS=1 \
.venv/bin/python benchmarks/manta_release_controls.py --output /tmp/manta-controls

Use --profile to select only a case whose acceptance is missing. The driver runs each enabled case and its fixed-source control, checks convergence and QVF validity, and verifies strict-target refusal before a computed archive is written. It prints source/native identity, full basis and physical supports, certificate targets/results, modeled memory and measured CPU/wall cost. Managed calculation output and QVF preserve the evaluated state. Keep this record outside the checkout and associate it with the queue job and runtime.

Profile	Method and mesh	Controls and requested targets
image-tight	Shifted two-point Be RHF	S, T and density-family kinetic-energy image targets 1e-6; physical pair radius 10 bohr and auxiliary radius 12.1 bohr
ordinary-rhf	Shifted two-point Be RHF	Joint image, full-Coulomb exchange and nuclear admission; loose physical targets exercise the control path
global-pbe0	Shifted two-point Be PBE0 RKS	The same joint controls with the method’s exact-exchange fraction and a fixed XC quadrature
compact-tight	Gamma He RHF, compact custom basis	Image, exchange and nuclear component targets 1e-8; pair radius 1 bohr, nuclear radius 18 bohr, reciprocal energy cutoff 512 Eh

The ordinary RHF/PBE0 cases request image targets 1e6, exchange targets 1e12 and nuclear targets 1e6. These intentionally loose targets are not production accuracy recommendations. Their RI coupling J/energy targets are 1e-6 over the automatically admitted canonical density family. The compact case uses exact finite coupling; it does not request a multipole truncation target. The explicit resource caps are 4 GiB memory, 1e12 shell calls and 1e12 primitive products. Scalar/kernel allowances are 7e18 for image-tight and 2e18 for the other profiles, including current arithmetic reservations. These are worst-case admission allowances, not measured operations or runtime estimates. A resource refusal remains a failure to admit the requested case, not a scientific PASS.

The driver prints the precise contracted basis coefficients, geometry, mesh shift and domains. These custom small fixtures demonstrate the controls; they are not converged material predictions for a paper. A paper’s actual materials still need their own basis, k mesh, XC grid and physical-domain convergence evidence.

What the bounds mean

Control	Included comparison	Exclusions relevant to release claims
one_electron_accuracy	Omitted AO-image overlap and kinetic matrix contributions, plus kinetic energy over the admitted canonical occupied-space family	Arithmetic, overlap-rank stability, other energy components and density relaxation
exchange_accuracy	Omitted full-Coulomb exchange domains, with the method scale and fixed Madelung correction accounted for	Madelung numerics, arithmetic, screened/RI exchange, finite fitting error, relaxation and k-mesh error
nuclear_accuracy	Omitted nuclear potential and repulsion contributions, plus attraction energy over the same density family	Arithmetic, metric solve, other energy components and density relaxation
RI coupling accuracy	Approximate coupling action relative to its declared finite source, propagated to fixed-family J/Coulomb energy	Finite fitting-basis error, omitted physical source domains, arithmetic and relaxed SCF energy

These certificates are separate ledgers with different reference objects. Do not sum them and call the result a total relaxed SCF energy tolerance. The requested budgets for the frozen profiles are:

Profile	S / T / kinetic-energy image targets	Exchange matrix / energy targets	Nuclear matrix / attraction-energy / repulsion-energy targets	Finite-source coupling J / Coulomb-energy targets
image-tight	1e-6 / 1e-6 / 1e-6	Not requested	Not requested	1e-6 / 1e-6
ordinary-rhf	1e6 / 1e6 / 1e6	1e12 / 1e12	1e6 / 1e6 / 1e6	1e-6 / 1e-6
global-pbe0	1e6 / 1e6 / 1e6	1e12 / 1e12	1e6 / 1e6 / 1e6	1e-6 / 1e-6
compact-tight	1e-8 / 1e-8 / 1e-8	1e-8 / 1e-8	1e-8 / 1e-8 / 1e-8	Not requested; exact finite coupling

S is dimensionless; the other matrix and energy bounds are in atomic units of energy. Energy bounds apply to the declared canonical density family and component, not to the relaxed total energy. A nuclear certificate’s total_energy_error_bound combines nuclear attraction and repulsion only. Zero coupling-truncation error for an exact finite contraction does not bound fitting-basis incompleteness, omitted physical domains or arithmetic. An unrequested ledger is not a zero-error certificate. The ordinary profiles’ large physical budgets establish admission behavior, not material accuracy.

Inspect every requested certificate’s qualified flag and measured bounds, not just a successful process exit. Do not report an iteration-capped state as converged. The existing full-Coulomb fitting and stored-metric controls retain their separately documented scope in the MANTA manual.

The frozen release examples exclude RI-K exchange certificates, screened exchange certificates, source-specific symmetry reduction, low-dimensional boundaries, gradients/stress, correlated production claims and speedup claims. Pair-tree metric-action bounds are a separate MANTA extension outside these frozen examples. Actual integral/Fock contraction roundoff remains later MANTA work; it does not hold this approved bounded release.

Composition of the certificates

The provenance of a periodic RI HF job carries a component_ledger that composes the separate certificates without double counting. Each requested certificate is mapped to the disjoint energy component it bounds at the canonical density and to its own reference object: one_electron_accuracy to the kinetic energy (omitted overlap and kinetic images), nuclear_accuracy to nuclear attraction plus repulsion (omitted real and reciprocal images), exchange_accuracy to exact exchange (omitted full-Coulomb domains with the fixed Madelung correction and method scale) and, when a coupling target was requested, fitting_accuracy to the Coulomb energy through the RI coupling action against its declared finite source. Because the components are disjoint, fixed_density_energy_error_bound is their outward-rounded sum. Each positive binary64 addition advances toward positive infinity; zero terms and a single stored bound remain exact. The receipt records composition_arithmetic as outward-binary64-sum-of-stored-nonnegative-bounds. This covers only summation of the supplied component certificates, not arithmetic in their construction. Requested energy bounds and supplied matrix bounds must be finite and nonnegative. Missing energy bounds, invalid bounds and sum overflow refuse qualification, return no composed bound and suppress the relaxation delimiter. Invalid energy values are represented by null in the failed ledger so it remains serializable with strict JSON. No scientific target is relaxed. The ledger refuses (qualified false with reasons) when a requested certificate is unqualified, when two certificates claim one component, or when the one-electron and nuclear certificates were evaluated over different canonical density envelopes. Its excludes list is the union of the certificates’ exclusions, with a component another certificate claims removed, plus density relaxation, k-mesh error, basis incompleteness, integral roundoff, overlap-rank stability, Madelung numerics, XC and the omitted physical Hartree domains for these frozen profiles, whose fitting certificate covers only the coupling action relative to its declared finite source. The separately supported full-Coulomb RI-BUPO certificate can replace that coupling contribution when its requested targets, selected Gamma source and canonical density identity qualify; it is never added to an overlapping coupling bound. That extension does not change the frozen profile inputs. The sum is a fixed-density component bound over the claimed components, not a relaxed total-energy tolerance.

relaxation_delimiter delimits, and does not bound, the scale of density relaxation for closed-shell results: from the certificates’ maximum-entry Fock error bounds and the converged canonical spectrum it reports the uncoupled closed-shell second-order estimate 2 ||dF_mo||_F^2 / gap with ||dF_mo||_F <= n_basis * (sum of entry bounds) / s_min, s_min the smallest retained overlap eigenvalue and gap the smallest occupied-virtual gap over the mesh. It is marked is_certificate: false and applies only with a positive gap at every k point and ||dF_mo||_F < gap/2; overlap perturbations are excluded because they change the metric, and the coupled orbital response is not included (it needs the stability eigenvalue, #625). On the compact-tight profile the delimiter is far below the component sum; on the ordinary profiles with loose targets it does not apply. Neither number is a relaxed total-energy tolerance.

Acceptance reuse and candidate handoff

The accepted component and complete-method evidence is recorded in the image, exchange and nuclear benchmarks. Their independent Gaussian/Boys/Fourier/libint and NumPy/SciPy references remain valid for unchanged components. No matched ORCA/CRYSTAL reference is available for these custom finite domains; KS also shares libxc. Fixed-source comparisons alone do not establish independent bulk parity.

For a containing candidate, retain the accepted source/native identities and verify the relevant implementation and fixture files. Reuse valid evidence; run only changed behavior or missing combinations. The compact simultaneous image/exchange/nuclear configuration is the added integration check here. Keep failed comparisons and unsupported combinations explicit. Exact candidate CI, publication checks and the immutable release tag remain separate release requirements. This page qualifies neither the complete 0.18 release nor all MANTA capabilities by itself.

Recorded qualification

The bounded control benchmark records eight converged complete calculations, ten strict-target refusals and bitwise reuse of saved accepted states on source 7917b560ffd3d48fbbd1bf4f0ecb66da898e7b4d. The compact simultaneous-control example passes all selected 1e-8 component targets. This closes the missing integration check for this frozen MANTA subset. The later shared continuous-grid integration adds two PBE0 calculations, three strict refusals and an independent NumPy/SciPy comparison; the unaffected HF evidence is reused. Its updated source/native identities and measured errors are in the same benchmark. MANTA admission includes the shared partition switching band. A containing 0.18 candidate must retain this scope and evidence and pass the release gates; the overall release is not declared ready here.

The corrected-candidate linkage also qualifies these physical profiles on source 2d0fdb8ca06b5e37c1f6721667800e9954defdfe: candidate 95c18c7e plus the converter-header repair a92b4880 and the explicit image-tight work allowance update after RI preflight reservations changed. Eight complete calculations and ten strict refusals pass with the documented caps; the old work-cap refusal is retained, and saved physical states and component certificates are bitwise unchanged. The benchmark retains the actual local native identity, cost and independent-reference reuse. This establishes the bounded MANTA contribution for that corrected source. The release candidate must include the repair and allowance update and reconcile its final identity. Its other gates and the paper’s material-specific convergence evidence remain separate.

The candidate inclusion check records that candidate 2b2ed883 includes the converter repair but still lacks the accepted work allowance, available on main as ebbf5957. Its numerical sources match the qualified integration. Include that allowance before treating the candidate’s image-tight example as qualified, then reconcile the final source and native receipt while reusing the accepted calculations for unchanged sources and inputs.

The prepared-successor verdict closes the source linkage for 7151963: it contains the accepted allowance and matches all 2015 source/input hashes. The bounded scientific contribution is accepted on the recorded local native build. Final candidate binary identity and publication remain with the release owner; no new calculation is required solely because the prepared source has a different commit ID.

The subsequent selected-candidate linkage confirms both repairs in 5dfc97f0. Its three changed source pins concern the ECP UKS derivative state, outside these all-electron RHF/PBE0 energy profiles. Their accepted scientific evidence remains applicable. The benchmark distinguishes the actual local calculation core, the completed predecessor’s Linux core and the selected candidate’s pending binary receipt; none is treated as a deployed runtime. The bounded MANTA source contribution is complete, while final candidate qualification stays with the release owner.