vibe-qc v0.18.4 - source release

Public source release of vibe-qc v0.18.4.

Download Source code (zip) or Source code (tar.gz) under Assets. These are source archives, not prebuilt installers. Follow the included installation instructions. Managed Git-based installers require a clone checked out at this tag.

Desktop availability is listed separately at https://vibe-qc.com/download/.

vibe-qc v0.18.3 - 2026-10-27

• Allow opt-in fresh FFTW collectors to receive the recipe-pinned source archive and exact build paths through an external spec; reject mismatches before configuration.

• Add an opt-in candidate adapter for prospectively authenticated FFTW installs,
  deriving configured inputs and host/toolchain context before reuse and
  checking final-prefix native/header observations with donor reads denied.
  Unsupported toolchains, patched sources and other dependencies remain held.

• Extend the opt-in FFTW prefix probe to observe the separate threaded library,
  its actual function bindings and a transform configured for two threads.
  Build provenance and whole-install reuse admission remain unevaluated.

• Standard and interior eigensolver requests now reject nonpositive or nonfinite residual tolerances before backend dispatch, preventing false convergence with infinite tolerance.

• Interior eigenproblem wrappers now preserve generalized overlap validation, retained-rank reduction and metric-normalized orbitals instead of bypassing the generalized dispatcher.

• Legacy Lanczos now refuses explicit root requests larger than its Krylov capacity instead of silently reducing the count and potentially certifying a partial spectrum.

• Gamma GDF UHF/UKS now refill a non-ensemble atomic/restart seed from its own
  physical Fock before the first finite-temperature iterate, as the multi-k
  route already does. Invalid accepted densities still fail entropy admission;
  zero-temperature and initial-guess-only evaluation remain unchanged.

Added: an occupation record on the neutral Γ-CCM producers (#519)

Every real-Gamma and neutral-Bloch result records how the spin sector's
electrons were distributed (result.occupation: integer or ensemble, the
largest departure from an integer, the Bloch producer's per-k counts, the
frontier gap and whether it is degenerate), in both convention records and
the .system manifest (aiccm_occupation). A same-Hamiltonian comparison
(assert_comparable_neutral_results) now also refuses an ensemble state by
name before any fold: the multi-k zero-temperature fill shares a
frontier-degenerate group equally, which has no determinant on the torus.
AICCM/CCM stays experimental.

Added: an explicit spin-state contract for the neutral Γ-CCM producers (#519)

CCMSystem(..., spin_state=), the real-Gamma and neutral-Bloch producers,
both runner arms and run_periodic_job(aiccm_spin_state=) name the spin
sector of the Born-von Karman torus: "legacy" (the default, each
producer's historical convention, unchanged number for number),
"per-cell" (the unit cell's (n_alpha, n_beta) replicated with spins
aligned on both producers) or an integer total multiplicity. The sector is
resolved before the shell-aware dispatch, impossible requests fail at
construction, the executed counts are recorded on every result, in the
.system manifest (aiccm_spin_policy, aiccm_spin_state) and in the
QVF convention record, and a same-Hamiltonian comparison is refused when
the two sides executed different sectors
(assert_compatible_spin_state, assert_comparable_neutral_results).
Under the default a doublet cell on a torus of more than one cell still runs
as a supercell singlet on real-gamma and as a per-cell doublet on
neutral-bloch; the mismatch is now recorded and refused instead of
silent. AICCM/CCM stays experimental.

• Added read-only experimental chi factor and complete Coulomb-block fidelity diagnostics. Localized the archived LiH reciprocal-40 refusal to retained-eigenspace forward accuracy near the metric cutoff, with independent operator and higher-precision controls. The original source still refuses; no producer, retained rank, support, admission tolerance or D93 qualification changed.

• Molecular DLPNO-MP2 PAO preparation now uses checked retained Fock solves for pair domains and dipole screening, preserving frozen-core projection, overlap rank, screening estimates and PNO thresholds.

• Molecular DLPNO-UCCSD default full union singles now reuse one checked Fock solve per calculation. Full-space ordering, independent singles amplitudes, frozen partitions and numerical gates are preserved while repeated identical preparation is removed.

• Molecular DLPNO-UCCSD Python local-TNO preparation now uses checked retained Fock solves for union T0 and iterative T1, preserving frozen partitions, pair-union rank, TNO selection, screened triples and numerical gates.

• Route cached molecular DLPNO-UCCSD spin-PAO Fock blocks through the shared residual-checked eigensolver with source, frozen, atom-domain and spin coordinates. Preserve overlap rank selection, projection tails, empty-space behavior and cache reuse without retaining solver snapshots.

• Molecular DLPNO-UCCSD rotated-occupied T1 triples now use checked retained TNO Fock solves with explicit source, frozen and triple coordinates. TNO selection, frozen partitions, screening refusals and numerical gates are preserved.

• Route retained singles-PNO Fock blocks in compiled molecular DLPNO-UCCSD through the shared residual-checked eigensolver, preserving independent spin and frozen partitions, density selection and PAO projection. The default full singles space needs no additional solve.

• Molecular DLPNO-UCCSD union residual preparation now reuses the checked spin-PAO Fock cache across pairs and singles, avoiding repeated construction for identical atom domains. Frozen partitions, overlap rank selection, domain ordering, empty-domain refusals and numerical thresholds are unchanged.

• Molecular DLPNO-UCCSD retained union pair-PNO and enabled singles-PNO Fock solves now use the shared checked boundary with explicit source, frozen and spin-coordinate labels. Density selection, degenerate spin spaces, PAO projection, frozen partitions and numerical gates are preserved.

• Experimental streamed molecular symmetry execution can bind MANTA's certified shell maps once and consume representative keys natively, avoiding repeated Python map copies. Mismatched plan maps refuse before evaluation; complete numerical audits, physical source identities and checked fallbacks remain unchanged.

• The experimental molecular quartet factory now consumes native MANTA source/support evidence for exact radial channels, shell mappings and declared geometry tolerance before integral evaluation. Complete numerical equivariance audits remain required; physical source and Fock identities, scientific tolerances and default routes are unchanged.

• Add an explicit experimental producer diagnostic for contracted s-shell four-center Coulomb construction errors, reusing MANTA's verified radial bounds. It includes actual native/null tiles, signed contractions and resource admission; full symmetry source audits remain mandatory.

• Experimental streamed molecular symmetry J/K can consume immutable MANTA
  libint/MD tiles through the native provider interface, preserving full source
  audits, canonical digests, physical counters and target ownership. The extra
  workspace is budgeted explicitly; smaller budgets and older runtimes retain
  the checked Python reader. Numerical-source and default-routing scope stay
  unchanged.

• Reconstruct symmetry-reduced CCM three-center integrals before metric whitening to avoid amplifying numerical noncommutation in near-null auxiliary modes. Metric screening and factor-level tolerances are unchanged.

• Keep the MgO multicell symmetry fixture on its documented no-tail profile on both sides, retaining the original factor-level tolerance.

• Add experimental parameter-opt molecular energy fitting: immutable parameters and reference protocols, balanced correlated contrasts, grouped splits, isolated DFTB/PM6 adapters, bounded conditional and nonlinear solves, restart artifacts and a queued-fit example. Broader P0-P3 scientific acceptance remains tracked in #676.

• Reject stale or contradictory symmetry-reduced BIPOLE k-point input before full-mesh expansion; preserve valid native reductions, consistent representative reordering and reciprocal-lattice wraps.

• Validate declared full BIPOLE multi-k coordinates against the current cell and shift before integral work or dry-run output, rejecting stale or incomplete Fourier grids while preserving valid point order and reciprocal wraps.

• Molecular GFN2-xTB: for molecules with an element beyond argon, the auto-stabilisation ladder on the default atom-resolved AES path gains a stiff-AES rung before the high-temperature rung (T = 0.005 Ha, AES-potential damping 0.02, 100 iterations). Open-d singlet atoms such as W, Os and Mo, whose spherical state is an unstable fixed point of the on-site multipole feedback and never converged on any earlier rung (nor in xtb 6.7.0), now converge there in about 50 iterations to the spherical solution (W at 300 K: -2.215078 Ha, matching xtb). The rung's budget comes out of the high-temperature rung (600 to 500), so every budget figure of the ladder is unchanged; each attempt record now carries ladder_aes_damping. Path-only: converged energies are unchanged.

• Apply the kinetic-density floor chain rule consistently in periodic meta-GGA potentials and fixed-grid gradients, so active-floor derivatives match the evaluated XC energy in restricted and unrestricted calculations.

• Multi-k Ewald UKS now computes its spin diagnostic from all k points and
  their complex orbitals and occupations, matching the other periodic
  unrestricted drivers. This removes spurious dependence on orbital phases
  and includes fractional occupations. SCF energies and densities are unchanged.

• BIPOLE multi-k HF and hybrid DFT now accept empty nonzero-transfer
  reciprocal spheres as zero long-range exchange sums. Elongated 3D cells
  with small Ewald splitting parameters no longer fail on these valid
  channels. Cutoffs, physical supports and zero-mode corrections are unchanged.

• Molecular ASE geometry optimization now refuses line-search warnings instead of accepting a minim...

vibe-qc v0.18.2 - source release

Public source release of vibe-qc v0.18.2.

Download Source code (zip) or Source code (tar.gz) under Assets. These are source archives, not prebuilt installers. Follow the included installation instructions. Managed Git-based installers require a clone checked out at this tag.

Desktop availability is listed separately at https://vibe-qc.com/download/.

vibe-qc v0.18.1 - 2026-10-25 - Berkelbach's Badger

• Bind the Ewald UKS ECP derivative preview to its retained terminal SCF
  densities, orbitals, energies and occupations. Reject edited states and
  contradictory ECP flags before evaluating a gradient; older receipts
  require a fresh SCF. The qualified force envelope is unchanged.

Fixed

• Separate the real restricted internal singlet and external spin-breaking
  stability channels for molecular RHF/RKS. Record each channel and preserve
  conservative aggregate rejection; an external instability is no longer
  labelled as an internal saddle or an excited SCF solution (#662).

Added

• Add the explicit real_restricted_cc reference contract for closed-shell
  CC single points. Its separate reference-admission field requires internal
  stability and completed tests of both real channels; it does not override
  general SCF usability or certify correlated accuracy.

• Update the bounded MANTA image-tight example work allowance for conservative arithmetic preflight reservations, preserving its memory cap, physical settings and accuracy targets.

---

Download Source code (zip) or Source code (tar.gz) under Assets. These are source archives, not prebuilt installers; follow the included installation instructions. Managed Git-based installers need a clone checked out at this tag.

Documentation: https://vibe-qc.com/ · Desktop downloads: https://vibe-qc.com/download/

vibe-qc v0.18.0 - source release

Public source release of vibe-qc v0.18.0.

Download Source code (zip) or Source code (tar.gz) under Assets. These are source archives, not prebuilt installers. Follow the included installation instructions. Managed Git-based installers require a clone checked out at this tag.

Desktop availability is listed separately at https://vibe-qc.com/download/.

vibe-qc v0.17.26 - 2026-10-22 - Tew's Tern

Validation scope

This patch adds LMP2 pair natural orbitals and a compiled pair-screening
kernel to closed-shell local DLPNO-CCSD, and changes open-shell DLPNO-CCSD
pair screening to 0.25 x tcut_pairs by default, which changes open-shell
DLPNO-CCSD and DLPNO-CCSD(T) energies. It adds COSMO-RS infinite-dilution
activity coefficients (COSMO-RS now warns as experimental), K-commensurate
hexagonal k-point meshes, compressed and joint-density MANTA metric bounds,
and repairs Gamma-point GDF image-pair routing, DLPNO-MP2 local-DF run_job,
native periodic XC difference domains, direct-HF warm-up validation and chi
reference restore. The integration pins are QVF toolkit v0.5.0 and
vibe-view v2.20.20; patches retain the Tew's Tern codename.

Scientific acceptance is limited to the calculations and references recorded
for each change. Candidate CI is not external-program parity or blanket
acceptance of experimental methods; MANTA certificates retain their
finite-operator and density-envelope scope.

Changed

• KPoints.recommend now gives a Γ-centred mesh on a cell with a hexagonal
  plane multiples of 3 along that plane, on the recommended mesh and on every
  verify=True ladder rung, so the mesh contains the zone corners K and K'.
  On DFTB0 graphene the former ladder (18, 25, 35) stopped at 25×25, whose
  energy lies 1.1e-5 Ha below the converged value because the mesh misses the
  Dirac point. The K-commensurate ladder (18, 27, 36, 51, 72) decreases
  monotonically and stops at 51×51. The ladder also keeps refining past a rung
  that repeats the previous mesh. A Γ-only axis is unchanged. (#129)

Added

• The closed-shell local DLPNO-CCSD solver can build its pair natural
  orbitals from full local MP2 instead of semicanonical MP2:
  LocalCCSDOptions(pno_guess="lmp2"). This is the TightPNO scheme of
  Liakos et al. 2015, Sec. 2.2. The coupled LMP2 equations, including the
  off-diagonal occupied Fock coupling, are solved in intermediate PNOs cut at
  lmp2_pno_scale * tcut_pno. The final PNOs are the natural orbitals of the
  LMP2 pair densities, the weak pairs take LMP2 energies, and the PNO
  truncation correction and the CCSD starting amplitudes come from LMP2. The
  lmp2_* defaults follow ORCA 6.1.1's TightPNO output. The result reports
  pno_guess, e_lmp2, n_lmp2_pairs, lmp2_iterations and
  lmp2_converged. The default stays "semicanonical", so no existing
  energy changes. (#190)

Performance

• Closed-shell local DLPNO-CCSD computes the full-virtual semicanonical MP2
  energies of its pair screening in a compiled kernel that runs over pairs
  in parallel (dlpno_sc_mp2_pair_energies). The NumPy loop it replaces is
  O(n_act^2 n_vir^2 n_aux). Its elementwise passes always ran on one thread,
  and its products did too whenever vibe-qc's BLAS pin was in effect. The
  per-pair projection onto the canonical virtuals reuses one C_vir^T S
  product. LocalCCSDResult.phase_seconds and
  LocalUCCSDResult.phase_seconds report where the closed- and open-shell
  routes spend their wall time (localisation, factor transforms, screening
  and pair setup, the LMP2 guess, the amplitude solver and triples).
  Energies are unchanged apart from floating-point summation order. (#190)

Changed

• Open-shell local DLPNO-CCSD now screens its spin-orbital pairs at
  0.25 * tcut_pairs, through the new LocalUCCSDOptions.tcut_pairs_spin_scale
  (default 0.25). This changes open-shell DLPNO-CCSD and DLPNO-CCSD(T)
  energies. The published TCutPairs values are calibrated on spatial pair
  energies, and ORCA's open-shell DLPNO screens the spin-summed spatial pair
  (Saitow et al. 2017, Eq. 53). A spin-orbital pair carries only part of
  that energy, so the unscaled threshold screened several times more
  loosely, and the extra screened pairs kept full-virtual MP2 energies that
  over-correlate. On C6H14/cc-pVDZ through both routes on one SCF,
  NormalPNO CCSD sits +0.652 kcal/mol above canonical on the closed-shell
  route, +0.263 on the open-shell route at scale 1.0 and +0.693 at 0.25. The
  C6H13 radical moves from +0.235 to +0.631 at NormalPNO and from +0.369 to
  +0.366 at TightPNO, so TightPNO no longer looks worse than NormalPNO
  through error cancellation. Over eight radicals with canonical references
  the CCSD(T) mean absolute error falls from 0.549 to 0.461 kcal/mol at
  NormalPNO and from 0.357 to 0.333 at TightPNO, because fewer triples are
  screened away. Set tcut_pairs_spin_scale=1.0 to reproduce
  an earlier result; the .out block records the scale. (#190)

Fixed

• Reject invalid active direct-HF level-shift warm-up lengths before integral construction instead of silently treating them as automatic warm-up. Explicit schedules and disabled shifts retain their existing precedence.

Performance

• Experimental chi local MP2 now evaluates every retained-factor Fourier audit
  entry with bounded character/displacement phase blocks and linear slabs.
  The streamed pair and MO contractions cannot recreate a full phase table.
  The factor cap includes concurrent work, cache copies and returned outputs;
  provenance reports tiles, recomputation counters and unchanged audit defects.
  Retained source storage and scientific admission are unchanged. Smaller
  budgets trade recomputation time for workspace; chi remains experimental
  and D93 fitted/post-HF remains not-qualified.

Added: COSMO-RS infinite-dilution activity coefficients, and COSMO-RS now warns as experimental (#558)

• run_job(cosmors=CosmoRSOptions(..., activity_coefficient=True)) also reports
  the solute's ln gamma at infinite dilution in the solvent against the pure
  solute, with its residual and combinatorial parts, on the conductor density.
  COSMOSPACE (Klamt 2002) is its independent test oracle.
• vibeqc.solvation.cosmors.cosmospace no longer claims to differ from the
  COSMO-RS sigma potential: Klamt 2002 Appendix A derives one from the other,
  and on the same segments the two give the same residual activity coefficient
  (measured to 5e-13).
• run_job(cosmors=...) and every Direct COSMO-RS SCF now emit
  vibeqc.CosmoRSExperimentalWarning, and the .system manifest records
  method_status = "experimental" with experimental_component, as the
  experimental label requires. On the eight Klamt 1995 vapour pressures at
  BPW91/def2-TZVP the transferred klamt1998 constants give rms 1.76 kcal/mol
  on the conductor density and 2.93 on the Direct COSMO-RS density.
• A Direct COSMO-RS free energy converges with the cavity grid only near
  0.25 angstrom. CosmoRSResult.fine_grid_spacing_ang records the spacing, and
  the .out notes a direct run on a coarser grid.
• Correction to the v0.17.25 entry for the Direct COSMO-RS density: its
  rms 1.09 kcal/mol was measured with each solute on a 0.30 angstrom cavity
  against a potential built at 0.35. On matched 0.25 angstrom cavities it is
  1.12 at RHF/STO-3G, against 1.23 on the conductor density.
• SolventResult.solvent_model carries the model that built a solvated run,
  so the finite-difference electronic-ESP gradient oracle rebuilds displaced
  cavities from the run's own recipe, radii overrides included.

Fixed

• Restoring a saved experimental chi HF reference (RHF, UHF or ROHF, with or without ECPs) in a fresh process no longer imports the DLPNO correlation package while computing the request digest. The digest values and the archive integrity checks are unchanged.

Fixed

• run_rhf_periodic_gamma_gdf (and run_rks_periodic_gamma_gdf) on a 3-D
  cell now keeps its molecular-limit route only when no atom pair reaches a
  periodic image within cutoff_bohr. It used to decide by the lattice
  vectors alone, so a box just longer than the cutoff with atoms near opposite
  faces stayed on that route while its fitted J/K counted the image pair.
  LiH/STO-3G in a 16-bohr box at the default 15-bohr cutoff came out 0.21 Ha
  low (-8.0789 Ha). It now runs the bulk Gamma GDF Hamiltonian, -7.8698 Ha,
  equal to PySCF's GDF with exxdiv='ewald' to 2e-11 Ha. Boxes with no image
  pair inside the cutoff are unchanged. (#98)

Fixed

• run_job(method="dlpno-mp2") with DLPNOMP2Options(local_df=True) runs
  again. Since DLPNO-MP2 began reading its inputs through the
  correlation-source contract, which requires three-centre factors with
  mo_transform, the route stopped before any work with "factors must
  provide mo_transform". run_job now passes
  vibeqc.dlpno.source.LocalFitFactors, which carries the orbital and
  auxiliary bases and builds the global fit only if a consumer asks for it,
  so the local-fit route still never evaluates the global three-centre
  tensor.

Documentation

• Clarify finite-k Ewald UHF energies versus isolated-molecule references.
  ECP regressions now check Gamma and multi-k UHF against matched periodic
  direct-integral references, and test the explicit molecular-limit exchange
  mode against ORCA. The production Ewald Hamiltonian and defaults are unchanged.

Added: compressed periodic metric bounds

Periodic MANTA RI fitting can require numerical metric J/energy bounds for
ordered compressed hierarchy and global-panel sources, using outward column
norms from the shared producer. Pair-tree actions and actual contraction
roundoff remain outside this bounded admission.

Improved: joint density limits for metric-action admission

Periodic MANTA metric-action admission now uses total and per-entry density
limits together through outward-rounded dual bounds. This tightens J and
fixed-density Coulomb energy certificates without changing RI factors,
physical source selection rules or Fock arithmetic. Sorting work is admitted
by both public and periodic aggregate resource plans.

Fixed

• Native three-dimensional Kohn–Sham calculations now retain every active AO-image difference required by the XC density and its potential. Initial guesses, restart densities and subsequent upd...

vibe-qc v0.17.25 - 2026-10-21 - Tew's Tern

Validation scope

This patch strengthens periodic physical-domain and XC calculations, improves
BIPOLE contraction and reduction performance, and adds bounded MANTA metric,
reference and auxiliary selection. It repairs finite-temperature SCC-DFTB
forces, periodic overlap diagnostics, toroidal BvK exchange and NEB worker
transport; sparse MRCI, direct-density COSMO-RS and explicit QVF local-triples
options are also included. The integration pins are QVF toolkit v0.5.0 and
vibe-view v2.20.20; patches retain the Tew's Tern codename.

Scientific acceptance is limited to the calculations and references recorded
for each change. New periodic, MANTA, toroidal, solvation and semiempirical
claims remain subject to their case-specific validation; candidate CI is not
external-program parity or blanket acceptance of experimental methods.
MANTA certificates retain their finite-operator and density-envelope scope.

Changed

• MRCI (method="mrci") assembles its MR-CISD Hamiltonian sparse from
  alpha/beta string excitation tables, with the same elements bit for bit,
  and finds the lowest root of spaces above 4000 determinants by Lanczos
  instead of a dense eigendecomposition. H2S/LANL2DZ CAS(4e,4o), whose MRCI
  exceeded the 300 s test cap, has 15100 determinants and now takes 1.5 s
  (264 s and about 5.5 GB before); its energy is unchanged to 1e-10 Ha.
  Smaller spaces keep the dense solve, so their results do not move. (#326)

Changed

• Issue numbers from the tracker used before the 2026-09-08 repository split
  now read "former-tracker issue N" in comments, test docstrings, the manuals
  and the semiempirical SECCM refusal messages (the GFN2, PM6 and SCC-DFTB
  Klopman-Ohno 3-D refusals), so they no longer point at unrelated current
  issues with the same numbers (#444, #211, #425, #130, #442). References to
  current issues are unchanged, and so are the released changelog sections.
  (#556)

Added: COSMO-RS solvation free energies on the Direct COSMO-RS density (#558)

run_job(cosmors=CosmoRSOptions(..., density="direct")) runs the solute in
Direct COSMO-RS with the same solvent potential and reports the solvation free
energy of that self-consistent density, Klamt & Diedenhofen 2015
(doi:10.1021/jp511158y) eq. 9. The default, density="conductor", is
unchanged. The gas reference and eta are the klamt1998 set's, not
refitted, and the reference states are still 1 bar to 1 mol/mol. For a solvent
with mu_S = 0 every reported field equals the conductor form's. On Klamt
1995's eight-liquid vapour-pressure panel at RHF/STO-3G the rms error is
1.09 kcal/mol, against 1.23 on the conductor density. The .out block states
the density, E_DCRS - E_gas and the macro-iterations, and the run cites the
2015 paper. vibeqc.solvation.cosmors.combinatorial_chemical_potential is new
and shared by both densities.

Added

• QVF job containers can carry dlpno_ccsd_options as a JSON record of the
  local DLPNO-CCSD(T) solver's LocalCCSDOptions, so a containerised run
  states its local triples algorithm ("triples_mode": "local", "t1" or
  "exact") instead of taking the engine default. Enumerated text fields
  accept only the implemented values, and unknown fields or values are refused
  at their option path. dlpno_options and dlpno_ccsd_options records may
  also name their threshold preset in a reserved "dlpno_thresholds" key (the
  JSON form of options_from_dlpno_thresholds), since run_job refuses the
  top-level preset beside an options object. The run record's system
  attachment names the triples algorithm that ran. (#558)

• Native dependency source preparation and image checks now support Python 3.6 bootstrap hosts before a project environment exists, preserving pinned source hashes and artifact identity checks.

Fixed

• Molecular SCC-DFTB analytic gradients at finite electronic temperature now
  use the Fermi-Dirac occupations that built the density. The energy-weighted
  density was rebuilt with Aufbau occupations (2 on the first n_occ orbitals),
  so every finite-temperature SCC-DFTB force, including those of the
  run_job Mermin retry ladder, came from the wrong occupations: on square
  cyclobutadiene at T = 0.001 Ha a carbon force was -0.0143 Ha/bohr against
  the -0.0364 of a central difference. Zero-temperature results are
  bit-identical. SCCDFTBResult gains occupations. (#563)

Fixed

• run_neb with the default n_jobs=0 no longer fails on molecular bands with
  "cannot pickle 'vibeqc._vibeqc_core.Molecule' object". Atom, Molecule and
  PeriodicSystem now pickle (a PeriodicSystem with attached symmetry
  refuses with a TypeError), and process workers return each image's
  warm-start density in a picklable form, so the parallel band equals the
  serial one (measured to 1e-15 Ha for RHF, UHF, ROHF and RKS). A run whose
  tasks or warm-start caches still cannot reach worker processes, for example
  one passing native rhf_options= or any periodic band (its k-mesh does not
  pickle), warns and evaluates the band serially instead of failing. (#564)

Fixed

• The full-k DFTB0 and SCC-DFTB drivers (run_dftb0_kpoints,
  run_scc_dftb_kpoints) now check the Bloch overlap S(k) before the SCF. If
  the pair cutoff leaves it indefinite, they refuse with a message that names
  the entry point, the smallest eigenvalue, cutoff_bohr and the remedy,
  instead of failing in the Bloch eigensolver with "S(k) is near-singular ...
  threshold 0.000000". The Li and Ga functions of the Stewart STO-6G basis
  still overlap past the 15-bohr default: rock-salt LiH on a 4x4x4 mesh is
  refused at 15 bohr (smallest eigenvalue -0.105) and runs from 25 bohr. The
  default is unchanged. (#565)

Changed: toroidal MP2 and DLPNO run their supercell HF on the BvK torus (#566)

toroidal_mp2_gamma, toroidal_mp2_gamma_pairs and the toroidal DLPNO-MP2,
DLPNO-CCSD and DLPNO-CCSD(T) wrappers used Ewald J with the molecular-limit
exchange, which keeps every density element in the home image. Two AOs that
neighbour across the periodic seam exchanged as if N - 1 cells apart, so a
translation of the torus by one cell changed K by 3.2e-3 (three cells of the
H2/STO-3G test chain) and 4.4e-3 (four cells), and K was not the derivative of
the reported energy (#534). The supercell HF now runs on the GDF Hamiltonian at
Gamma with exxdiv='ewald' exchange: it equals the unit-cell KRHF on the
N-point mesh to 1.4e-13 Ha per cell, every operator commutes with the torus
translation, and the results report exchange_q0="BvK-ewald". At three cells
of the chain e_hf moves from -1.1164536589 to -1.1178468400 Ha per cell and
e_corr from -0.0131980528 to -0.0131780804 Ha per cell; PySCF KRHF/FFTDF gives
-1.1178240341, the 2.3e-5 difference being the JK fitting error. In a cell
padded with vacuum the HF energy converges only as 1/N along the replicated
axis, to a limit that depends on the transverse vacuum.

toroidal_mp2_gamma_pairs now splits the MP2 energy over Wannier pairs, with
the canonical amplitudes and integrals rotated to the Wannier functions; the
total is unchanged. Its former assignment of canonical orbitals to cells was
undefined on a translation-symmetric torus, whose canonical orbitals are Bloch
states, and wannier_cells now holds the cells of the Wannier functions.

Fixed

• Fitted periodic MANTA runs now cite the inverse-bound literature they emit:
  manta_hermitian_inverse cites Rump and Lange (2023) and Higham (2008),
  and a requested metric action adds Rump (2010). Both routes had been
  appended inside the last bibliography entry of the citation database rather
  than under [routes.numerics], so these runs cited nothing for them. The
  database integrity tests now refuse an entry field the registry does not
  read. (#567)

• BIPOLE density-to-k-point recovery exports only the selected Born-von-Karman
  representative blocks through a new LatticeMatrixSet.get_blocks(indices)
  accessor. It preserves minimum-distance selection and Fourier arithmetic while
  avoiding copies of the full real-space density support on each recovery.

• Share lattice positions and image labels across periodic direct J/K leaf
  caches, reducing memory use without changing Gaussian envelopes, screening
  bounds or contraction order. Existing memory caps and fallbacks are retained.

• Share the short-range BIPOLE product-tree lattice partition across shell pairs and construct their separate bounds from the leaves upward, reducing duplicate index storage while retaining the existing screening tests and contraction order.

• Tighten the default BIPOLE short-range image precision from 1e-6 to 1e-8 across SCF and public forwarding paths, correcting the missing-shell Coulomb-cancellation residual in the Ne molecular-limit regression. Explicit settings remain available; the radial tolerance is not a total-energy error bound.

• Reject nonfinite BIPOLE short-range radii and screening parameters before domain construction. The low-level exact-zone builder now rejects nonpositive zones consistently with the driver, preventing invalid input from silently removing short-range output blocks.

• Stabilize the vendored Libxc PBE and r2SCAN correlation potentials for empty spin channels by evaluating spin fractions without subtractive cancellation. This also benefits hybrids using those kernels. Functional parameters and regularization are unchanged; existing unpatched Libxc builds are rebuilt automatically.

• Publish public self-contained Linux desktop downloads for vibe-view and vibe-sketch with checksums and corresponding source. Clarify that macOS and Windows editions are planned paid Store purchases.

• Add bounded metric cutoff selection for periodic MANTA fitted Coulomb. Certify each candidate against the shared full-Coulomb reference before SCF, retain the first passing source, record discarded trials, and reserve their full resource cost. Auxiliary and reference domains rem...

vibe-qc v0.17.24 - 2026-10-20 - Tew's Tern

Fixed

Write supplied transition-density grids as canonical QVF sections, preserving signed values and checking their optional excited-state links.

Validation scope

This patch improves periodic support selection and nuclear error budgets,
reuses BIPOLE lattice and reconstruction work, repairs GPW exchange adjoints
and toroidal AO seam pairs, adds periodic ECP routes, and strengthens the
AICCM Bloch reference solve. Dense-cell Gamma RIJCOSX exchange is refused
where its image sum has no limit. It also provides the standalone
vibeqc-naming distribution. The integration pins are QVF toolkit v0.5.0 and
vibe-view v2.20.19; patches retain the Tew's Tern codename.

Scientific acceptance remains limited to recorded calculations and their
reference comparisons. Periodic derivative and energy comparisons remain
under active validation; candidate CI does not establish external-program
parity or blanket acceptance of experimental methods. MANTA certificates
retain their declared finite-operator and density-envelope scope.

Fixed: the obsolete BIPOLE fold-guard known-red baseline is retired (#182)

Retire the obsolete BIPOLE fold-guard known-red baseline after fixture, domain-guard, grid-motion and original MgO finite-difference acceptance. Numerical thresholds and production behavior are unchanged.

Added: periodic ECPs on ROHF/GDF, Gamma RIJCOSX and the open-shell multi-k Ewald drivers (#35)

run_periodic_job applied a periodic ECP on one route, the k-point GDF
drivers with RHF/RKS/UHF/UKS, and refused an ECP-bearing cell everywhere
else. method="ROHF" on jk_method="gdf" and RHF/RKS/UHF/UKS on
jk_method="rijcosx" (3D cells) now build V_ne and the ionic energy on the
Z_eff nuclear frame, add the lattice-summed V_ECP to Hcore and fill the
valence count, with the ECP context threaded into the SAD/SAP/HUECKEL/MINAO
seeds. The direct-call drivers run_uhf_periodic_multi_k_ewald3d and
run_uks_periodic_multi_k_ewald3d do the same on 3D cells. Slabs, BIPOLE,
the Gamma fast paths and the external-XC adapters keep their explicit
refusal; orphan ECP metadata is refused on every route that applies the
operator.

Fixed

• Gamma GPW HF and hybrid SCF (RHF, RKS, UHF, UKS, ROHF, ROKS) now use a
  Fock that is the derivative of the energy they report in compact cells.
  The molecular-limit exchange keeps its density pairs inside one image cell
  while its output pairs span cells, so once images overlap the plain map is
  not self-adjoint: in a 9-bohr H2 cell the Fock missed the energy
  derivative by 7.7e-6 Ha along one density direction, and the SCF stopped
  away from a stationary point. build_jk_gamma_molecular_limit gains
  adjoint_symmetric=True, which returns the self-adjoint (K + K^+)/2 and
  (J + J^+)/2 from the same integrals; the energy at every density is
  unchanged and the default (False) keeps the historical map (bit-identical
  to the previous core in a single-threaded build).
  The GPW drivers and evaluate_gpw_energy use it; GAPW and the other
  callers are unchanged. The
  exchange energy is still the molecular-limit model, not physical Gamma
  exchange. (#534)

Added: vibeqc-naming, the naming engine as its own distribution (#536)

vibeqc_naming can now be installed without vibe-qc and its compiled core:
pip install ./vibeqc-naming from a checkout builds a pure-Python wheel from
the same python/vibeqc_naming source that the vibe-qc wheel keeps bundling,
so vibeqc.naming and vibe-qc installs are unchanged. NumPy is its only
dependency and [solids] adds spglib. The in-package test scripts stay out
of the wheel, and vibeqc_naming.__version__ reports the distribution's own
version line, starting at 0.1.0.

Fixed

• Gamma-only RIJCOSX RHF (run_periodic_rijcosx_rhf, and
  run_periodic_job(jk_method="rijcosx") at a single k-point) no longer
  returns a converged nonsense energy on a dense slab or bulk cell. Its
  tight-cell exchange is a real-space image sum with no G=0 (exxdiv)
  treatment, which grows without limit with lattice_opts.cutoff_bohr once
  the images span two or more lattice directions: on Mg fcc / STO-3G the
  exchange at a fixed density went from -16.2 Ha (home cell) to -59.5 Ha at
  the 15-bohr default, and the SCF reported -275.97 Ha. The driver now
  refuses such cells before any integrals, as the Gamma RKS, UHF and UKS
  RIJCOSX drivers already did, and names jk_method="gdf" at Gamma or a
  k mesh of at least two points as the alternatives. Chains with vacuum
  beyond the cutoff keep the image model. (#555)

Fixed: AICCM reference="bloch" converges its Bloch SCF far enough for the transform

ccm_reference_from_bloch checks that the occupied Bloch orbitals reproduce
the Bloch density to 1e-8. A Bloch SCF stopped at the multi-k default
conv_tol_grad of 1e-6 misses that check by its last step, so the default
reference="bloch" path of run_ccm_ri_mp2, run_ccm_ri_ccsd and
run_ccm_dlpno_mp2_direct refused LiH and MgO pob-DZVP-rev2 at (1,1,1).
That path now runs the Bloch SCF at conv_tol_grad 1e-8
(BLOCH_REFERENCE_SCF_CONV_TOL_GRAD) unless the caller passes options.
run_periodic_job(method="aiccm", variant="neutral-bloch", correlation=...)
does the same when conv_tol_grad is not set. The refusal message now names
the convergence fix.

• BIPOLE symmetry-domain preparation batches integer image transforms by atom
  pair, reducing setup overhead on larger finite domains. Cutoff decisions,
  fixed-point closure, cell order and physical-domain safety checks are unchanged.

• Lattice symmetry reconstruction batches members sharing an atom rotation,
  reducing repeated small matrix operations in BIPOLE Fock builds. It preserves
  the two Wigner-D products, caller cell ordering and uncovered matrix blocks.

• The physical BIPOLE route now reuses the periodic nuclear-attraction builder's
  bounded per-product source screening. This reduces short-range one-electron
  setup work for HF and DFT without changing AO-pair support or Ewald parameters;
  the omitted attraction is bounded at 1e-16 hartree per matrix element before
  floating-point summation error.

• Add optional bounded physical AO-pair, nuclear and reciprocal support selection before periodic RI HF/KS SCF, reusing shared component bounds and canonical density envelopes, with cumulative trial budgets, explicit limits and final source verification. Whole-method accuracy remains unfinished.

• Add optional nuclear matrix, attraction-energy and nuclear-repulsion targets
  to complete periodic RI HF and KS calculations. Admission reuses the shared
  nuclear producer and canonical-density analysis before SCF, with fixed
  physical sources, count-only resource checks and separate scalar provenance.
  These component bounds exclude arithmetic, fitting/metric errors, canonical
  rank changes and density relaxation; automatic source selection remains open.

Fixed: manual status pages and the vibe-sketch license on the technical site

Refresh companion desktop and basis-optimisation status in the manual, document the queue gateway route, and update the technical site to show vibe-sketch as MPL-2.0 while retaining explicit software availability. Update excited-state table support to match current viewer source.

• BIPOLE real-space energy contractions batch lattice-cell key export in the
  native binding, reducing repeated Python and NumPy bookkeeping across RHF,
  UHF, RKS and UKS. Operator support, accumulation order and energy arithmetic
  are unchanged; density values are read afresh on every contraction.

Changed: the roadmap matches the implemented APIs and the remaining qualification

Reconcile the current roadmap, basis-library inventory and historical roadmap navigation with implemented APIs and remaining scientific qualification. Distinguish native translation correlation from the BvK-supercell preview, correct SCF/property and QVF status, and remove obsolete release targets without changing method support or promising a release date.

Changed: the technical website's feature, licensing and distribution claims

Updated the technical website to describe current viewer, queue and sketch capabilities, scope method-status claims, distinguish implemented desktop packaging from published downloads, and state the application and QVF licenses consistently.

Fixed: toroidal supercells keep the AO pairs across the periodic seam

The toroidal MP2, pair, DLPNO-MP2 and DLPNO-CCSD(T) wrappers run their
supercell HF through run_rhf_periodic_gamma_gdf, which builds S, T, V_ne,
the Ewald-J pair transform and the molecular-limit K on the lattice
translations with |g| <= cutoff_bohr. A supercell's atoms span the whole
cell, so two AOs a few bohr apart across the periodic seam are joined only by
a translation one period long. Once the supercell was longer than the
15-bohr ball, those pairs were dropped: on three cells of the 6-bohr H2 chain
of tests/test_periodic_toroidal_mp2.py the ball held only the home cell,
S(Gamma) lost the 0.056 overlap of the two atoms 4.6 bohr apart across the
seam, and the SCF orbitals were not orthogonal under the true overlap
(|C_occ^T S C_vir| = 0.026).

The wrappers now run the HF on a copy of the options whose cutoff_bohr is
the smallest radius that holds every translation an AO pair needs within the
requested cutoff (18, 24 and 30 bohr for 3, 4 and 5 cells of that chain). The
supercell overlap then equals the BvK fold of the unit-cell lattice overlap.
One and two cells of the chain fit the shipped ball and are unchanged. This is
a driver-side stand-in for the pair-complete enumeration of #44.

HF energy per cell on the chain, before and after: -1.1164515356 to
-1.1164536589 Ha (3 cells), -1.1164189739 to -1.1164049637 Ha (4 cells),
-1.1163978619 to -1.1163978278 Ha (5 cells); a 40-bohr ball gives the new
values to 8e-9 Ha. The MP2 correlation energy per cell on three cells moves
from -0.0132120 to -0.0131981 Ha; PySCF FFTDF...

vibe-qc v0.17.23 - 2026-10-19 - Tew's Tern

Validation scope

This patch strengthens isolated dependency installation, periodic nuclear
error bounds, complete RI exchange accuracy targets, TDDFT transition-dipole
factors, D4 optimization and differentiation, and counterpoise free-atom
support for solid basis optimization. It qualifies QVF toolkit v0.5.0 and
vibe-view v2.20.18. Patch releases retain the Tew's Tern codename.

Scientific acceptance is limited to the recorded calculations and reference
comparisons. The diffuse-p short-range integral regression, its sixteen
controls and the NaCl acceptance calculation have passed. The compact GPW
UKS calculation now passes its unchanged runtime criterion. The LiH retained
energy comparison still fails its original tolerance; periodic partition-grid
and strict derivative criteria remain unresolved. These results do not claim
external-program parity or blanket acceptance of experimental methods.

MANTA certificates retain the declared finite-operator and density-envelope
scope. Candidate CI qualifies the exact release source and integration
contracts; it does not close pending scientific issues.

• Evaluate experimental chi local-MP2 occupied-source and selected-frame audits
  through verified cyclic blocks, reconstructing every distinct displacement
  before taking the unchanged maximum-entry metric and independent frozen/active
  projector norms. Preserve exhaustive replay for unverified inputs and threshold
  boundaries. Reuse bounded source-bound symbols and measurements within one
  preparation, account for their payload before allocation, and report audit
  formulations and counters. Chi remains experimental; D93 fitted/post-HF and
  scientific runtime acceptance remain not-qualified and pending, respectively.

Fixed

• TDDFT and CIS oscillator strengths are no longer too large. The closed-shell
  singlet transition dipole used a spin factor of 2 instead of √2, doubling every
  restricted (RHF/RKS, and ROHF TDA) oscillator strength, and the unrestricted
  path counted each amplitude twice, quadrupling UHF/UKS values. The full Casida
  solvers also built the dipole from the wrong vector, (X−Y) rescaled to unit
  length, instead of X+Y on the (X+Y)·(X−Y) = 1 normalisation; their stored
  excitation vectors, dominant amplitudes and NTOs now use X+Y as documented.
  RHF and UHF results in the TDA and the full equation agree with PySCF to 1e-6;
  excitation energies are unchanged. (#545)

Fixed

• run_job(..., dispersion="d4") with a generic DFT functional now optimizes
  and differentiates the DFT-D4 surface whose total it reports. Before, the
  final total included D4 but every optimizer backend (geomopt provider,
  native, Brent and ASE) and the finite-difference Hessian worked on the bare
  functional. The D4 energy and analytic gradient are now evaluated at every
  geometry and every Hessian displacement, D4 is still counted once in the
  final total, and the frequency block names the surface (... + D4).
  Composite 3c recipes, D3(BJ) and dispersion-free runs are unchanged; a D4
  optimization or Hessian without the optional dftd4 package refuses before
  the first SCF. New: vibeqc.dispersion.D4DispersionSpec,
  compute_dispersion, and compute_hessian_fd(dispersion_params=...).
  (#548)

• Install and update helpers now isolate Python dependency resolution and verification from ambient Python paths and user-site packages, preventing missing runtime dependencies after an apparently successful installation.

Added: counterpoise free atoms for solids in the vibe-qc cohesive engine

VibeQcEngine now declares the counterpoise capability, so
CohesivePipeline(engine, counterpoise=True) computes every free atom in
the ghost basis of its own crystal instead of recording a skipped stage.
This is the convention of the published pob cohesive energies, which use
CRYSTAL's ATOMBSSE for every free atom, and it unblocks the vibe-qc half
of the M2 gate's engine rung (#224). The new
vibeqc.basis_optimization.counterpoise module builds the cluster as the
CRYSTAL23 manual defines it: the real atom plus the basis functions of the
neighbours in the first NSTAR symmetry stars that lie closer than RMAX
angstrom. The engine's bsse_nstar and bsse_rmax default to
Crystal23Engine's 30 / 10.0. The bare atom seeds the cluster SCF, and a
cluster energy above the bare one is reported as counterpoise_above_bare.

• Add an Apple Silicon macOS app, DMG and ZIP packaging pipeline containing
  Python, the native scientific engine, basis data and the naming, fetch and
  dispersion dependencies. Bundled commands run without a separate Python or
  Homebrew installation. The first native profile requires macOS 27.
  The fetch extra now installs OPTIMADE's HTTP client dependencies.

• Add shared periodic full-Coulomb exchange bounds valid throughout the
  canonical occupied-space density family, with reusable native overlap
  analysis and separate entrywise and density-norm energy enclosures.

• Add a standalone periodic nuclear error producer for the physical
  AO-pair midpoint spheres used by RI methods. It separates real,
  reciprocal and missing-AO-image potential bounds and nuclear repulsion
  bounds, reusing shared Gaussian and lattice producers. Complete-method
  target admission and automatic physical-source selection remain pending.

• Add optional pre-SCF exchange accuracy admission to periodic RI HF and
  global hybrids with exact K, including the fixed BvK overlap correction
  and reuse of the canonical density envelope across component ledgers.

• Optional periodic RI HF/KS overlap, kinetic and canonical-density kinetic-energy image targets reuse the shared Gaussian bounds, with count-only admission and a separate provenance certificate. An explicit completion cube can tighten the enclosure without changing physical sources. Whole-method accuracy remains unqualified.

---

Download Source code (zip) or Source code (tar.gz) under Assets. These are source archives, not prebuilt installers; follow the included installation instructions. Managed Git-based installers need a clone checked out at this tag.

Documentation: https://vibe-qc.com/ · Desktop downloads: https://vibe-qc.com/download/

vibe-qc v0.17.22 - 2026-10-18 - Tew's Tern

Validation scope

This patch updates the QVF 0.5 toolkit contract, periodic Gamma operator and
short-range fitting support, axis-aligned GPW collocation, and experimental
MANTA automatic RI coupling accuracy. QVF archives retain format version 1.
The candidate qualifies this exact source with QVF v0.5.0 and vibe-view v2.20.17.
Patch releases retain the Tew's Tern codename.

Scientific acceptance is limited to the recorded calculations and reference
comparisons. The original GPW Hessian and compact restart checks passed their
stated assertions and 300-second runtime cap. The dense-ionic diagnostic also
met its cost criterion but remains physically unreliable. The compact UKS case
in #299 still exceeds its unchanged cap. The short-range fitting and NaCl SCF
problem (#551), and remaining strict periodic derivative criteria (#510), still
require their own numerical acceptance. Earlier failures remain evidence and
are not converted to PASS by this release.

MANTA accuracy certificates bound the fixed finite coupling operator under the
recorded density envelope. They do not establish basis, domain, arithmetic,
density-relaxation or complete-method physical accuracy. Other unresolved
periodic cost, reference and broad method criteria remain open. Candidate CI
qualifies the release source and integration contracts without closing those
scientific issues.

Fixed

• Correct the 3D native Gamma RHF operator to use the same density on every
  lattice displacement and physical pair/source support for all its terms.
  This restores integral adjoint and density-response closure at finite
  cutoffs; energies can differ from the previous home-cell-density
  approximation. Cutoff and k-point convergence remain separate requirements.

• Preserve drawing members, metadata, extension declarations and named input structures when executing QVF jobs. Keep the original request unchanged, show the result geometry by default, resolve job structure references explicitly, and refuse unsafe result collisions with atomic rollback. Legacy unreferenced jobs remain single-structure containers with the submitted geometry attached to their run record.

Fixed

• Preserve diffuse periodic Gaussian cancellation in short-range fitting by
  removing an additional primitive screen that was outside the declared image
  error budget. Finite domains and SCF/metric thresholds are unchanged.

• Share real Gaussian image values across k-points when constructing the
  bounded compact GPW AO cache, including skew cells, while preserving Bloch
  phases, image support and the existing streaming memory policy.

Performance

• Evaluate Gamma GPW atomic orbitals on axis-aligned grids using separable
  Gaussian factors and a native image sum, avoiding full shifted-grid copies.
  Grid points, periodic images, basis normalization and contraction coefficients
  are unchanged; skew cells and custom grids retain the general evaluator.

Added: fixed-envelope automatic periodic RI-BUPO coupling accuracy

• Experimental periodic RI-J can select a fixed pair-tree coupling tolerance
  from J and Coulomb-energy targets for an explicit AO density envelope.
  HF/KS jobs retain the selected source through SCF and record its certificate.
  Densities outside the envelope and unmet targets fail explicitly. The
  guarantee excludes basis/domain, arithmetic and density-relaxation errors.

• MANTA automatic periodic RI-J selection uses less aggressive tolerance
  proposals to retain certified approximate partitions instead of needlessly
  falling back to exact finite coupling. Tiny positive targets no longer
  produce an invalid zero tolerance. Evaluated bounds, resource limits and
  the fixed source contract remain authoritative.

Added: componentwise periodic RI accuracy allocation

• Automatic periodic RI coupling selection uses entrywise density bounds
  from the shared canonical-overlap helper and absolute action envelopes
  supplied by the existing coupling producer. Explicit entry limits are
  supported alongside a positive total norm limit. Both constraints are
  retained and enforced at every evaluated density, with numerical provenance
  in the accuracy receipt. Whole-method accuracy scope is unchanged.

Added: overlap-derived periodic RI accuracy limits

• Automatic periodic RI-BUPO coupling selection can derive its density bound
  from the fixed canonical overlap and spin occupations. HF/KS jobs no longer
  require a manual AO density limit; explicit positive overrides remain
  supported. Output records the resolved bound and its numerical provenance.
  Coupling-only accuracy scope and per-density enforcement are unchanged.

• Pin the QVF 0.5.0 toolkit contract and refresh its byte-identical reference pages. Links to the new CLI, C and Fortran guides follow the same release pin. QVF archives remain format version 1.

---

Download Source code (zip) or Source code (tar.gz) under Assets. These are source archives, not prebuilt installers; follow the included installation instructions. Managed Git-based installers need a clone checked out at this tag.

Documentation: https://vibe-qc.com/ · Desktop downloads: https://vibe-qc.com/download/