Anatomy of .mfsig · mfsig.com

The signed file, every field

Anatomy of a .mfsig

Five pillars. Every parameter declared, typed, gate-checked, and tagged with a confidence band. What's in the file is what you can defend.

current schema (production)mfsig/v0.91.1· generated by MFFactory v0.91.1, the unified DFT-first pipeline

v0.91.1 unifies the five pre-v29 SCF pipelines (drug atlas, pocket fragments, dimer BSSE-CP, hetero pair, drug-as-pocket) into a single source of truth: molforge.MFFactory.compute(). Every mfsig field is a first-principles DFT observable — no Klamt empirical thresholds, no CPCM-X spatial averaging, no external-package calibration constants. Each file carries a 4-way provenance stamp (mf_factory_version · vendor_tree_sha256 · registry_recipe_sha256 · image_tag) so a reader twenty years from now can reconstruct the exact compute context byte-for-byte.

Legacy schemas v2.5-multivariant and earlier are kept readable for back-compat (their archived files retain their original signatures); all new mfsigs ship in v0.91.1 from MFFactory.

headline science · v0.91.1wB97M-V solves the deep-gap pathology that B3LYP couldn't

v28 (B3LYP)

3 of 10 oracle pairs had NEGATIVE HOMO-LUMO gap. BSSE_PCM range −29.6 → +34.9 kcal/mol with multiple state-flips.

v29 (wB97M-V, MFFactory)

10/10 pairs converge at L0 (no fallback). Gap range +1.4 → +3.1 eV (all healthy). BSSE in ±3.4 kcal/mol.

corroborated independently

96/96 drug SCFs (incl. 31 formally charged) all L0-converged with 0 negative gaps. 146/146 wB97M-V SCFs clean.

The functional itself — not the basis-set fallback chain, not UHF-broken-symmetry — is the load-bearing fix. Range-separation + built-in VV10 NLC place electrons in the correct LUMO on charged-pocket dimers that B3LYP folded into closed-shell solutions with vanishing gap.

What this file is, in 30 seconds

A .mfsig.json is a signed, schema-validated, version-locked snapshot of a single molecule's σ-profile, cavity, energy, and provenance — packaged so that any reader, twenty years from now, can re-verify every claim against the bytes without contacting us. Five pillars below; each field is typed, gate-checked, and tagged with a confidence band.

Confidence HIGH · gate-checked, deterministic MED · documented systematic limit LOW · research only

File family

Three .mfsig formats sharing one Ed25519 chain — all now produced by the unified MFFactory v0.91.1 orchestrator. Each extends the previous; same audit, same renderer, same 4-way provenance stamp.

.mfsig.jsonmonomer

One drug, σ-profile + cavity + atomic charges. Used by COSMO-RS, solubility, ΔG_solv kernel. Generated by MFFactory.compute_drug().

mfsig/v0.91.1 · wB97M-V/def2-SVP

.mfsig.assembled.jsonself-dimer

Drug × drug homodimer. ΔE_cohesion + H-bond / π-stack descriptors. Used for melting-point + crystal cohesion. Generated by MFFactory.compute_dimer_bsse_cp().

mfsig/v0.91.1 · 3-SCF Boys-Bernardi

.mfsig.hetero_dimer.jsondrug × polymer

Drug × polymer-segment heterodimer. ΔE_interaction + Flory-Huggins χ. Used for ASD miscibility screening. Generated by MFFactory.compute_hetero_pair().

mfsig/v0.91.1 · 5-SCF asymmetric CP

`audit_and_trust`

Cryptographic identity + 4-way MFFactory provenance. Tamper-evident, ALCOA-compatible, byte-for-byte reproducible.

Field	Type	Conf.	What it means
`sha256_checksum`	string · hex64	HIGH	SHA-256 of the canonicalised payload. Recomputed on load. Mismatch = file rejected.
`uuid`	string · uuidv4	HIGH	Stable identifier minted at write time. Travels with the file forever.
`timestamp_utc`	ISO 8601	HIGH	UTC moment the .mfsig was sealed.
`molecule_name`	string	HIGH	Trivial name from the request — for humans, not a unique key.
`mf_factory_version`	string	HIGH	Version of the unified compute pipeline, e.g. "0.91.1". Pins the orchestrator + SCF fallback chain + sigma-moment formulas. First leg of the 4-way provenance stamp.
`vendor_tree_sha256`	string · hex64	HIGH	SHA-256 of the vendored gpu4pyscf tree at compute time (patched C-PCM + ghost-exclusion). Asserts that the exact patched binary that produced the numbers can be re-materialised. Second leg of the provenance stamp.
`registry_recipe_sha256`	string · hex64	HIGH	SHA-256 of the entry in MF_FACTORY_REGISTRY.json under versions.0.91.1 — the immutable recipe (functional / basis / dispersion / radii / sigma-moment formulas / SCF policy). Third leg.
`image_tag`	string	HIGH	Docker image tag + sha256 of the worker that ran the SCF (e.g. worker:v29 @ sha256:2d45d796…). Fourth leg — locks the OS, CUDA, and dependency layer.
`image_patch_chain`	string[]	HIGH	Ordered list of patch markers baked into the image (Klamt-purge, ghost-exclusion P1, C-PCM L2bis cavity fix, wB97M-V default, …). Lets a future reader reconstruct WHY this image differs from upstream gpu4pyscf.
`reproducibility`	object	HIGH	engine version + build hash + library lock — replay anywhere. Redundant with the 4-way stamp above; kept for legacy readers.
`lineage`	object	HIGH	Append-only history of every derivation event. tier, generation, parent_sha256, genesis_sha256, history[] — a git for σ-profiles. Each entry references the mf_factory_version that produced it.
`ownership`	object	HIGH	license_id, license_type, organization, signing_key_id, molforge_signature (Ed25519). Public key at /.well-known/mfsig-keys.json.
`recipe`	object	HIGH	Snapshot of the named recipe ("mf-factory-v0.91.1"): functional = wB97M-V, basis = def2-SVP (def2-TZVPP L3 fallback), dispersion = built-in VV10 NLC, radii = Bondi (P1-patched), counterpoise = Boys-Bernardi where applicable. Also carries validation_at_compute (anchor set + MAE + n_molecules).
`derived_grade`	string	HIGH	MF1 / MF2 / MF3 / MF4 — the customer-facing grade derived from the recipe and the validation context.

`chemistry_and_geometry`

What the molecule is, where every atom sits.

Field	Type	Conf.	What it means
`smiles`	string · canonical	HIGH	RDKit-canonical SMILES of the input. Lossless round-trip with the structure.
`inchi_key_14`	string · 14-char	HIGH	Structure-derived hash. Use this as the unique molecule key, not the name.
`atomic_numbers`	int[N]	HIGH	Element list. N matches positions array exactly.
`positions_aa`	float[N][3]	HIGH	Cartesian coordinates in Ångström, relaxed geometry. ‖∇E‖ < 1 mEh/Bohr at write time.
`n_heavy / n_total`	int	HIGH	Heavy-atom and total-atom counts — used for tier pricing and benchmark cohorts.

III

`quantum_and_thermodynamics`

The physics: density, surface, energy, σ-signature. v0.91.1 MFFactory generation — pure DFT-derived, zero Klamt empirical thresholds.

Field	Type	Conf.	What it means
`scf.converged`	bool	HIGH	Energy gate. False = file rejected at sealing time, never shipped.
`scf.method`	string	HIGH	L0=DIIS, L1=DIIS+SOSCF (Newton), L2=LS=0.5+coarse grid, L3=DAMP+SOSCF. v0.91.1 production : 96/96 drug SCFs converge at L0.
`scf.fallback_level`	int (0-3)	HIGH	Which level of the L0→L3 fallback chain converged. 0 = clean DIIS. Higher = harder system.
`scf.cycles / scf.wall_s`	int / float	HIGH	SCF iteration count + wall-clock seconds for forensic audit.
`energies.total_hartree`	float	HIGH	Converged DFT total energy E[ρ].
`energies.g_polar_hartree`	float	HIGH	C-PCM polarization energy = mf.with_solvent.e (raw DFT, canonical). Always negative for neutral closed-shell drugs.
`energies.homo_hartree`	float	HIGH	Highest occupied MO energy (occupation > 1e-6). v0.91.1 / wB97M-V : 0/96 drugs show pathological gap collapse.
`energies.lumo_hartree`	float	HIGH	Lowest unoccupied MO energy. Strictly above HOMO under wB97M-V on this cohort.
`energies.gap_ev`	float	HIGH	(lumo − homo) × 27.2114. v0.91.1 production range across 96 drugs : +5.7 → +13.9 eV. Negative gaps → file rejected.
`moments.dipole_debye`	float[4]	MED	(μx, μy, μz, ‖μ‖) from converged density via mf.dip_moment(unit='Debye').
`moments.atomic_charges_mulliken`	float[N]	MED	Mulliken-partition atomic charges via diag(dm @ S). N matches atom count.
`cavity_and_sigma.cavity_area_aa2`	float	HIGH	Σ area_i over all surface tiles, mf.with_solvent.surface['area']. Bondi vdW radii after P1 ghost-exclusion patch.
`cavity_and_sigma.n_segments`	int	HIGH	Number of Connolly surface tiles (typically ~3000-11000 depending on drug size).
`cavity_and_sigma.sigma_per_segment`	float[Ns]	HIGH	Raw σ_i = q_sym_i / area_i per surface tile (e/Å²). Preserved verbatim — downstream can re-compute any moment without regen.
`cavity_and_sigma.segment_areas`	float[Ns]	HIGH	Per-tile area in Å²; pairs 1-to-1 with sigma_per_segment.
`cavity_and_sigma.sigma_moments`	object	HIGH	Area-weighted moments of RAW σ (no Klamt cutoff, no r_av averaging): { mean, variance, skewness, kurtosis }. Formulas in MF_FACTORY_REGISTRY.json versions.0.91.1.sigma_moments_formula. No HB-donor/acceptor mass split — the kernel discovers HB-like signal from variance + skewness.

`ai_intelligence_layer`

Pre-computed features ML pipelines can consume directly.

Field	Type	Conf.	What it means
`embedding_vector`	float[D]	MED	Fixed-length molecular embedding suitable for retrieval / similarity / classification heads.
`feature_provenance`	object	HIGH	Which feature extractor + version produced the embedding. Frozen, replayable.
`downstream_tasks_supported`	string[]	HIGH	Declared compatibility list — ΔG-solv · γ∞ · χ12 · SLE · LLE.

`legacy_vault`

One-way bridge to incumbent σ-profile formats — for portability.

Field	Type	Conf.	What it means
`cosmotherm_profile`	string \| null	HIGH	COSMOtherm-compatible σ-profile block. Drop-in replacement for legacy pipelines.
`turbomole_cosmo`	string \| null	HIGH	Turbomole .cosmo file body. Lossless re-export when source format permits.
`openCOSMO_rs`	string \| null	HIGH	openCOSMO-RS feed. For groups that already have an open ASD workflow.
`vendor_provenance`	object	HIGH	Which incumbent each block targets, version range, known caveats.

`.mfsig.assembled.json`homodimer · mfsig/v0.91.1 · MFFactory

Drug × drug self-cohesion. Inherits all five monomer pillars (audit · chemistry · quantum · AI · legacy) for both copies, then adds the assembled_data block with the cohesion energy and dimer-only descriptors. Used for melting-point prediction, crystal cohesion screening, and π-stack / H-bond network analysis.

Field (under `assembled_data`)	Type	Conf.	What it means
`monomer_inchi_key_14`	string · 14-char	HIGH	InChI-key of the monomer being self-assembled — matches the chemistry_and_geometry pillar of the underlying .mfsig.json.
`dimer_geometry_aa`	float[N×2][3]	HIGH	Cartesian coordinates of both A copies at the relaxed dimer geometry, Ångström. xtb_relax pre-step then DFT single-points (no PySCF gradient — bypasses geomeTRIC + PCM pathology).
`n_scfs`	int · always 3	HIGH	Boys-Bernardi symmetric counterpoise: SCF on isolated A, SCF on isolated A′ (ghost-A geometry), SCF on AB dimer. Three independent SCF cycles per artefact.
`E_monomer_a_hartree`	float	HIGH	Total DFT energy of isolated copy A at the dimer geometry.
`E_monomer_b_hartree`	float	HIGH	Total DFT energy of isolated copy A′ at the dimer geometry.
`E_dimer_hartree`	float	HIGH	Total DFT energy of the AB pair at the relaxed dimer geometry.
`delta_e_cohesion_kcal_mol`	float	HIGH	ΔE = E_dimer − E_a − E_b, converted to kcal/mol. The cohesion observable. BSSE-corrected via the counterpoise scheme.
`delta_e_bsse_kcal_mol`	float	HIGH	Basis-set superposition error correction, separately reported for forensic audit. Always ≥ 0 by construction.
`hb_network`	object	HIGH	H-bond network: pairs of (donor_atom_idx, acceptor_atom_idx, distance_aa, angle_deg, type) classified by geometric criteria. Counts: hb_donor_count, hb_acceptor_count, hb_total.
`pi_stack_descriptor`	object	MED	π-stack analysis on aromatic ring centroids: stacked / T-shaped / parallel-displaced + centroid distances. Only emitted when the monomer has ≥ 1 aromatic ring.
`contact_area_aa2`	float	HIGH	Connolly surface contact area between A and B at the dimer geometry — the touching surface where cohesion is mediated.
`assembled_provenance`	object	HIGH	Recipe used (mf-factory-v0.91.1), xc_functional = wB97M-V, basis = def2-SVP (def2-TZVPP L3 fallback), dispersion = built-in VV10 NLC, counterpoise_scheme = Boys-Bernardi, dimer_relax_method = xtb_relax + DFT single-points, methodology_hash + 4-way MFFactory provenance stamp. Frozen at write time.
`lineage_link`	object	HIGH	Parent monomer .mfsig SHA-256 (parent_sha256) so the dimer is verifiably derived from a signed monomer. The lineage chain runs monomer → assembled, never the other direction.

Why three SCFs, not two. For a homodimer, the counterpoise scheme still runs both monomer SCFs in the dimer-basis ghost-atom setup (so A is in A′-basis ghosts, and A′ is in A-basis ghosts), even though the two are chemically identical. This makes BSSE-correction explicit and replicable — the same as the heterodimer recipe with the symmetry simplification recorded in assembled_provenance.counterpoise_scheme.

VII

`.mfsig.hetero_dimer.json`drug × polymer · mfsig/v0.91.1 · MFFactory

Drug × polymer-segment interaction for amorphous-solid-dispersion (ASD) miscibility prediction. Inherits both partners' monomer pillars, then adds hetero_interactionwith the asymmetric Boys-Bernardi counterpoise output plus the Flory-Huggins χ regression. Dual-Track storage: the raw unbounded χ is the canonical AI-internal field; a separate tanh-bounded χ_display is emitted ONLY for the frontend viewer.

Field (under `hetero_interaction`)	Type	Conf.	What it means
`drug_inchi_key_14`	string · 14-char	HIGH	InChI-key of the drug partner.
`polymer_segment_smiles`	string	HIGH	SMILES of the polymer monomeric segment used in the pair (e.g. PVPVA-vinyl-acetate, Soluplus-vinylcaprolactam-co-PEG, HPMCAS-cellulose-acetate-succinate). The segment is documented and matched against the polymer registry.
`polymer_segment_mw_da`	float	HIGH	Molar mass of the segment monomer in Daltons — needed by Flory-Huggins χ to convert ΔE_interaction (per pair) to per-lattice-site χ.
`dimer_geometry_aa`	float[Na+Nb][3]	HIGH	Cartesian of the relaxed AB heterodimer.
`n_scfs`	int · always 5	HIGH	Asymmetric Boys-Bernardi: (1) A clean, (2) B clean, (3) AB dimer, (4) A in dimer-basis with ghost-B, (5) B in dimer-basis with ghost-A. Five independent SCF cycles.
`E_drug_clean_hartree`	float	HIGH	DFT energy of the drug at the dimer geometry, monomer basis only.
`E_polymer_clean_hartree`	float	HIGH	DFT energy of the polymer segment at the dimer geometry, monomer basis only.
`E_dimer_hartree`	float	HIGH	DFT energy of the drug + polymer-segment dimer at the relaxed geometry.
`E_drug_in_dimer_basis`	float	HIGH	DFT energy of the drug in the full dimer basis (with ghost-B atoms). Used to subtract BSSE attributable to the basis the drug sees in the dimer context.
`E_polymer_in_dimer_basis`	float	HIGH	Mirror of the above for the polymer segment.
`delta_e_interaction_kcal_mol`	float	HIGH	ΔE_interaction = E_dimer − E_drug_clean − E_polymer_clean, kcal/mol. The interaction observable. BSSE-corrected via the asymmetric counterpoise.
`delta_e_bsse_kcal_mol`	float	HIGH	Basis-set superposition error for forensic audit, separately reported (always ≥ 0).
`flory_huggins_chi`	float	HIGH	Raw, unbounded χ at 298 K. The CANONICAL AI-internal field — every downstream MolForge pipeline reads this value. Calibrated by closed-form ridge regression against the ASD miscibility cohort (90 hetero-dimères, 9 drugs × 10 polymères pharma, B3LYP+BSSE) — LODO ranking top-1 78%, top-3 89%, Spearman ρ +0.76 vs Hansen HSP ~30% top-1. NO Hansen HSP heuristics, NO empirical group contributions.
`flory_huggins_chi_display`	float	HIGH	tanh-bounded copy of χ clipped to [−3, +3] — emitted ONLY for the frontend Viewer's chip rows. Never trained on. Distinct field name so downstream consumers cannot confuse the two.
`asd_miscibility_tag`	string · enum	MED	Human-readable classification: good / moderate / poor / phase_separating. Derived from χ ranges with the boundary justifications recorded in the recipe registry.
`contact_area_aa2`	float	HIGH	Connolly surface contact area between drug and polymer-segment at the dimer geometry.
`hb_network`	object	HIGH	H-bond network across the interface: donor-on-drug × acceptor-on-polymer pairs, plus the inverse.
`chi_calibration_provenance`	object	HIGH	ASD miscibility cohort SHA (refs/asd_miscibility_kernel.json), anchor count (n=90 hetero-dimères, 9 drugs × 10 polymères), ridge λ, LODO ranking metrics (top-1 78%, top-3 89%, Spearman ρ +0.76), feature set (8-d RDKit drug fingerprint + 10-d polymer one-hot). Frozen at calibration time.
`hetero_provenance`	object	HIGH	Recipe = mf-factory-v0.91.1, xc_functional = wB97M-V, basis = def2-SVP (def2-TZVPP L3 fallback), dispersion = built-in VV10 NLC, counterpoise_scheme = 'asymmetric_bb', dimer_relax_method = 'xtb_relax + DFT_single_point', methodology_hash + 4-way MFFactory provenance stamp.
`lineage_link`	object	HIGH	Parent drug .mfsig.json SHA-256 + parent polymer-segment .mfsig.json SHA-256 → the heterodimer is verifiably derived from two signed monomers.

Dual-Track storage

Raw χ is canonical. Display χ is presentational. The unbounded flory_huggins_chi is what every downstream MolForge AI pipeline trains on — we preserve the full quantum signal. The tanh-clipped flory_huggins_chi_display exists only so frontend chip rows show a value within the legacy pharma [−3, +3] band. The two are separately named so a consumer can never accidentally pipe display values back into training.

ASD miscibility kernel · n=90

No Hansen HSP, no empirical group contributions. χ is fit by closed-form ridge regression against the ASD hetero-dimer cohort — 90 dimères calculés en B3LYP/def2-SVP + counterpoise BSSE sur 9 drogues × 10 excipients pharma standards (eudragit, hpmcas, peg400, plga, pva, pvp, soluplus_*). LODO blind : top-1 hit 78%, top-3 hit 89%, Spearman ρ +0.76 (2.5× plus précis que Hansen HSP littérature ~30% top-1). Le kernel utilise 8 features RDKit (MolWt, LogP, MolMR, TPSA, NumHBD, NumHBA, NumRotBonds, NumAromRings) + one-hot 10 polymères. Anchor SHA-256, ranking metrics, feature set vivent dans chi_calibration_provenance sur chaque fichier. Source: refs/asd_miscibility_kernel.json.

How .mfsig compares

Every format that touches σ-profile or molecular property pipelines, side-by-side. ✓ first-class · ◐ partial / extractable · ✗ absent. The .mfsig column is what you take home.

Capability	.mfsig this spec	.cosmo Turbomole	CTD COSMOtherm	OpenCOSMO-RS feed	.fchk / .log Gaussian	ORCA out ORCA	NwChem out NwChem	xtb out xtb	.sdf MDL · V2000	.mol MDL · V2000	.xyz atoms only	.pdb RCSB	SMILES string	InChI string
coverage	25/25	6/25 +3◐	3/25 +4◐	3/25 +2◐	7/25 +2◐	6/25 +4◐	6/25 +4◐	6/25 +4◐	5/25	5/25	3/25	4/25 +2◐	4/25	4/25 +1◐
Identity Connectivity (SMILES / structure)	✓	✗	✗	✗	◐	◐	◐	◐	✓	✓	✗	✓	✓	✓
Stable structure hash (InChI Key)	✓	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✓
Atomic numbers + positions (Å)	✓	✓	◐	✗	✓	✓	✓	✓	✓	✓	✓	✓	✗	✗
Bond table	✓	✗	✗	✗	✗	✗	✗	✗	✓	✓	✗	◐	✓	◐
σ-content σ-profile histogram	✓	✓	✓	✓	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗
Per-segment σ-surface (position · normal · area · charge)	✓	◐	✗	◐	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗
σ-moments (mean · var · skew · kurt)	✓	✗	◐	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗
HB donor / acceptor mass	✓	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗
Charge conservation gate (∫σ·dA ≈ 0)	✓	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗
Quantum Converged SCF energy	✓	✓	✗	✗	✓	✓	✓	✓	✗	✗	✗	✗	✗	✗
g_polar (COSMO solvation free energy)	✓	✓	✓	◐	✗	◐	◐	◐	✗	✗	✗	✗	✗	✗
Bias-corrected ΔG (FreeSolv-anchored)	✓	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗
Direct DFT dipole archived	✓	✗	✗	✗	✓	◐	◐	◐	✗	✗	✗	✗	✗	✗
Atomic charges (Mulliken)	✓	◐	✗	✗	✓	✓	✓	✓	✗	✗	✗	✗	✗	✗
Forces ∂E/∂R at the reported geometry	✓	✗	✗	✗	✓	✓	✓	✓	✗	✗	✗	✗	✗	✗
Trust SHA-256 audit seal	✓	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗
21 CFR Part 11 / ALCOA-compatible	✓	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗
UUID + ISO timestamp	✓	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	◐	✗	✗
Method + engine version pinned in file	✓	◐	◐	✗	✓	✓	✓	✓	✗	✗	✗	✗	✗	✗
Reproducibility metadata (lib lock)	✓	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗
AI Embedding vector (ML feature store)	✓	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗
Declared downstream tasks (γ∞ · χ12 · ΔG)	✓	✗	◐	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗
Interop Legacy vault — re-exports other vendor formats	✓	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗
Open spec · no vendor licence required	✓	✓	✗	✓	✓	✓	✓	✓	✓	✓	✓	✓	✓	✓
Single-file deliverable (no companion needed)	✓	✓	✓	✓	◐	◐	◐	◐	✓	✓	✓	✓	✓	✓

Legend✓ first-class field◐ partial · extractable with effort✗ absent

Sources: published format specs for Turbomole .cosmo, COSMOtherm CTD, Gaussian .fchk/.log, ORCA, NwChem, xtb, ChemDraw / MDL SDF (V2000 / V3000), Protein Data Bank PDB, InChI / SMILES strings, OpenCOSMO-RS feed. Comparison is for the data persisted to a single deliverable file — not for what the originating tool can compute.

v0.92.0 SOTA migration · in flightrecipe added 2026-05-29

One change in v0.92.0: `basis: def2-SVP → def2-TZVPP`

wB97M-V was parameterised against def2-TZVPP (Mardirossian & Head-Gordon, 2016). v0.91.1 ran on def2-SVP as a pragmatic v28-era speed compromise. v0.92.0 flips the default to TZVPP — the basis the functional was calibrated for — and bumps the L3 fallback to def2-QZVPP. Everything else is unchanged: same wB97M-V, same C-PCM ε=80, same Bondi (P1) radii, same raw-σ moments, same 4-way provenance stamp. This is the only recipe knob that moves.

accuracy

TZVPP: ~1 kcal/mol energetics uncertainty (drug-like organics). SVP: ~3-5 kcal/mol. TZVPP is the recommended basis for wB97M-V production work.

cost

~3× more basis functions/atom → ~9× SCF time. Drug atlas (4529) goes from ~3h (SVP) to ~7h (TZVPP) on 5× H100.

kernel impact

All closed-form kernels (PC-SAFT, LogP, hydration, binding, χ) re-fit against v0.92.0 atlas after regen. v0.91.1 mfsigs stay readable for back-compat.

Recipe is in refs/MF_FACTORY_REGISTRY.json versions.0.92.0 (SHA cd88bda5…) · full pipeline diagram + audit commands at /factory.

v0.91.1 changelog · MFFactory generation

What changed from v2.5-multivariant to v0.91.1

NEW · one orchestrator

MFFactory.compute()

The five pre-v29 SCF pipelines (drug atlas, pocket fragments, dimer BSSE-CP, hetero pair, drug-as-pocket) collapse into one unified entry point in src/molforge/mf_factory.py. Same SCF fallback chain, same vendor guard, same provenance stamp on every artefact.

NEW · Klamt purge

pure area-weighted moments

Removed σ_HB threshold (0.0084 e/Å²), r_av spatial averaging (0.537 Å), hb_donor_mass / hb_acceptor_mass / polar_area_aa2 / halogen_mass empirical fields. The new sigma_moments block stores mean / variance / skewness / kurtosis of the raw σ_i = q_sym_i / area_i — the kernel discovers HB-like signal directly from the moments.

NEW · wB97M-V default

deep-gap pathology solved

Functional flipped from B3LYP-D3BJ → wB97M-V (range-separated + built-in VV10 NLC). On the v29 sealed-holdout: 10/10 oracle pairs converge at L0 with healthy gaps (+1.4 → +3.1 eV), where B3LYP had 3/10 with negative HOMO-LUMO gap. Corroborated by 96/96 drug SCFs L0-clean. 146/146 wB97M-V SCFs total.

NEW · 4-way provenance stamp

Every artefact now carries mf_factory_version · vendor_tree_sha256 · registry_recipe_sha256 · image_tag. Reading a file ten years from now reconstructs the exact compute context — Python source, vendored gpu4pyscf binary, recipe immutable, OS + CUDA + dependency layer — byte-for-byte.

KEPT · lineage chain

The append-only lineage from v2.4 is unchanged: parent_sha256 → new_sha256 + history[]. Each entry now carries the producing mf_factory_version so a tier upgrade or recipe enrichment can be traced to the exact MFFactory call that produced it.

Backward readable, not backward producible — the verifier still opens any v2.4 / v2.5-multivariant file and replays its signatures, but new mfsigs only ship in v0.91.1. The Klamt-tainted fields are intentionally absent from the new schema; downstream consumers must read from cavity_and_sigma.sigma_moments (raw moments) instead of the legacy chemistry-aware split.

See the lineage chain →Recipes side-by-side →How verification works →

Example mfsig/v0.91.1 · MFFactory generation · trimmed for readability

{
  "mfsig_version": "0.91.1",
  "audit_and_trust": {
    "sha256_checksum":          "7e1f744c77d25dee78515d29ee505f55...",
    "uuid":                     "b61705e5-b992-4a7a-8a3e-7762155cc178",
    "timestamp_utc":            "2026-05-27T18:42:11Z",
    "molecule_name":            "BAXOFTOLAUCFNW",

    "mf_factory_version":       "0.91.1",
    "vendor_tree_sha256":       "9c1a4f88b0e7d3a2c5e6f1b9d8a7e4f3...",
    "registry_recipe_sha256":   "fd2d7890ded98d3af5b577f46c28fdcb...",
    "image_tag":                "worker:v29 @ sha256:2d45d796275ef25e840b373cba842a87...",
    "image_patch_chain": [
      "klamt-purge:no-sigma-hb-threshold",
      "klamt-purge:no-rav-spatial-averaging",
      "ghost-exclusion:P1-bondi-vdw",
      "pcm-cavity:L2bis-fix",
      "scf-default:wB97M-V/def2-SVP"
    ],

    "lineage": {
      "tier":           "Platinum",
      "generation":     1,
      "parent_sha256":  null,
      "genesis_sha256": "7e1f744c77d25dee78515d29ee505f55...",
      "history": [
        { "op": "genesis",
          "tier": "Platinum",
          "tool": "MFFactory.compute_drug::v0.91.1",
          "new_sha256": "7e1f744c..." }
      ]
    },
    "ownership": {
      "license_type":      "internal-rd",
      "organization":      "molforge.ai R&D",
      "signing_key_id":    "molforge-dev-2026",
      "molforge_signature": "boAdABa/LLHnDa7Rb05TddTi6qKIzacSTgDx..."
    },
    "recipe": {
      "name":             "mf-factory-v0.91.1",
      "functional":       "wB97M-V",
      "basis":            "def2-SVP",
      "dispersion":       "built-in VV10 NLC",
      "radii":            "Bondi (P1-patched)",
      "methodology_hash": "fd2d7890ded98d3af5b577f46c28fdcb...",
      "validation_at_compute": {
        "anchor_set":      "v29-sealed-holdout-2026-05-27",
        "MAE_dG_kcal_mol": 1.18,
        "n_molecules":     96
      }
    },
    "derived_grade": "MF3"
  },

  "chemistry_and_geometry": {
    "smiles":         "CC(C)N(C(C)C)C(=O)C...",
    "inchi_key_14":   "BAXOFTOLAUCFNW",
    "atomic_numbers": [6, 6, 6, 7, ...],
    "positions_aa":   [[...], ...],
    "n_heavy":        24,
    "n_total":        38
  },

  "quantum_and_thermodynamics": {
    "scf": {
      "converged":      true,
      "method":         "L0:DIIS",
      "fallback_level": 0,
      "cycles":         18,
      "wall_s":         42.7
    },
    "energies": {
      "total_hartree":    -1234.5678,
      "g_polar_hartree":     -0.0163,
      "homo_hartree":        -0.2412,
      "lumo_hartree":        -0.0118,
      "gap_ev":               6.24
    },
    "moments": {
      "dipole_debye":              [1.21, -0.84, 2.06, 2.55],
      "atomic_charges_mulliken":   [-0.31, 0.12, ...]
    },
    "cavity_and_sigma": {
      "cavity_area_aa2":   711.42,
      "n_segments":        4128,
      "sigma_per_segment": [...],
      "segment_areas":     [...],
      "sigma_moments": {
        "mean":      4.79e-05,
        "variance":  4.32e-05,
        "skewness": -0.214,
        "kurtosis":  3.067
      }
    }
  },

  "ai_intelligence_layer": {
    "embedding_vector":   [...],
    "feature_provenance": { "extractor": "molforge.kernel.v29-9d", "stamp": "..." }
  }
}

Note the absent fields vs v2.5: no physics_variants wrapper (single canonical SCF per file), no hb_donor_mass / hb_acceptor_mass / polar_area_aa2 / halogen_mass (Klamt-tainted, removed), no sigma_hb_threshold: 0.0084 (purged), no sigma_profile_101 binned histogram (raw per-segment σ is preserved instead).

Every field in a .mfsig is sealed by the SHA-256 in pillar I. Tamper with one byte and the audit signature breaks. The same file opens twenty years from now in the open spec — that's the contract. No vendor lock-in, no rebuild required.

Anatomy of a .mfsig

File family

audit_and_trust

chemistry_and_geometry

quantum_and_thermodynamics

ai_intelligence_layer

legacy_vault

.mfsig.assembled.jsonhomodimer · mfsig/v0.91.1 · MFFactory

.mfsig.hetero_dimer.jsondrug × polymer · mfsig/v0.91.1 · MFFactory