mfsig.com
Releases

Changelog

What shipped, when. Validated on a CPU pod for every release.

v0.92.0 — TZVPP (in validation)

wB97M-V / def2-TZVPP · registered, head-to-head validating · not yet at atlas scale2026-06-15
  • RECIPE — v0.92.0 (wB97M-V / def2-TZVPP + C-PCM ε=80 + Bondi cavity) upgrades the basis def2-SVP → def2-TZVPP, the basis wB97M-V was actually calibrated for. One knob moves.
  • STATUS — registered and being head-to-head validated against the v0.91.1 production recipe; NOT yet at atlas scale, so v0.91.1 remains the production recipe every order is generated with.
  • PIPELINE — same canonical path: MFFactory.compute() → vendor_guard (SHA-256 of the vendored PCM module) → feature extraction. Both recipes can coexist in one signed file via the physics_variants block (schema v2.5).

v0.91.1 — production atlas

Single source of truth · MFFactory + vendor-guard SHA-pin2026-05-31
  • ATLAS — 6,876 signed fingerprints at v0.91.1 (wB97M-V / def2-SVP + C-PCM ε=80 + Bondi cavity), all computed via the canonical MFFactory write path. Recipe SHA-pinned, vendor patches SHA-validated on every compute. The full atlas regenerates from raw data in ≤ 2 days for < $50.
  • ARCHITECTURE — single source of truth: every production .mfsig flows through MFFactory.compute() → vendor_guard (SHA-256 of the vendored PCM module) → feature extraction. No parallel compute paths in production; legacy writers are quarantined behind a raising stub.
  • VENDOR PATCHES — gpu4pyscf forked at a SHA-pinned commit + a local C-PCM cavity patch (ghost-atom exclusion on the surface). Validated by an A/B benchmark and a CPU-side logic test, both kept under vendor/patches/.

mfsig v2.5-multivariant

Schema · physics_variants · chem-aware moments · same file holds three recipes2026-05-14
  • SCHEMA — new physics_variants block under quantum_and_thermodynamics. One .mfsig now carries results for multiple recipes side-by-side, each with its own methodology, results, validation, and methodology_hash. Backward compatible: legacy v2.4 files remain valid v2.5.
  • CHEM-AWARE MOMENTS — sigma_moments now splits by chemical role of the surface segments: hb_donor_mass · hb_acceptor_mass · pi_negative_mass · polar_area_aa2 · halogen_mass · sigma_negative_mass · sigma_positive_mass. Bond-table classification (segment_chemistry v2.2), not σ-threshold-only. Legacy donor/acceptor masses remain available as *_legacy fields.
  • LINEAGE — unchanged from v2.4. The git-for-σ-profile derivation chain (parent_sha256 + history[]) carries forward. Tier upgrades and recipe enrichments append to the chain without rebuild.
  • VIEWER — 3D viewer and /anatomy page updated to render the new fields and the multi-variant layout.

v0.29 — audit tiers

Three audit tiers · Standard · Compatibility · Premium2026-05-13
  • Compatibility tier — drop-in feed for legacy COSMO-RS pipelines. Reference-aligned on the public overlap set. Maps to the MF3 audit tier.
  • Premium tier — pure-physics, zero empirical calibration on the benchmark. Same audit grade, intended for R&D engagements. Maps to the MF1 audit tier.
  • Standard tier (preserved) — production audit grade, reference-validated. Maps to the MF2 audit tier.
  • REGISTRY — versioned recipe registry now exposes all three tiers under a single schema; each entry carries methodology hash, validation context, and pricing tier. /recipes renders from the registry.

v0.28.3

D3BJ + Bondi · 84/84 PASS · HB signal restored · charge conservation at machine ε2026-05-13
  • COHORT — 58 / 58 drugs PASS 7/7 gates · 26 / 26 polymers PASS 7/7 gates · 84/84 overall. Charge conservation ∫σ·dA verified to machine ε (max |Q| ≈ 1.6 × 10⁻¹⁵ e) on the entire cohort.
  • HB SIGNAL RESTORED — reverted to Bondi cavity radii after the v0.28.2 regression suppressed donor / acceptor σ-amplitudes. Donor/acceptor σ-amplitudes back within physical expectation across the cohort.
  • DISPERSION — Grimme D3BJ correction wired into the DFT path. Improves London dispersion against the def2-SVP baseline; no impact on σ-surface invariants.
  • DIPOLE — direct DFT dipole archived in every .mfsig (a known basis-set bias of def2-SVP, addressed by the Reference-tier def2-TZVPD path).
  • DELIVERY — drug + polymer bundles + atlas snapshot. Frontend viewer + anatomy page updated to the new fields.

May 2026

100-drug Ground-Truth cohort published2026-05-12
  • 100-drug reference cohort delivered in 38.6 minutes — 99/100 PASS all four audit gates
  • σ-profile Pearson r vs CPU ground truth: +0.9998 on aspirin and atorvastatin
  • Energy fidelity |ΔE| ≤ 1.1 × 10⁻⁹ H across the cohort
  • Every entry signed with SHA-256 trust hash — tamper-evident by construction
  • Reference tier now includes a DOI-backed Zenodo deposit and an automated 21 CFR Part 11 ALCOA audit pack

Apr 2026

DrugCentral bulk cohort2026-05-12
  • 389-drug DrugCentral coverage running on the production stack
  • .mfsig.tar.gz cohort delivery for batch customers
  • Webhook-on-completion for hands-off cohort runs
  • Per-drug acceptance gates published with every cohort

Mar 2026

Analytical cavity shipped2026-05-12
  • Full Connolly Solvent-Excluded Surface — contact, saddle, reentrant — composed analytically
  • 5 of 5 cavity unit tests pass at machine ε precision
  • No fitted parameters per molecule — same cavity construction across all drugs
  • Same cavity used in Pro, Platinum, and Reference tiers

Feb 2026

.mfsig.json open format · free converter2026-05-12
  • Five-pillar canonical file format for σ-profile artefacts
  • SHA-256 trust hash — 21 CFR Part 11 ALCOA-compatible
  • Universal converter reads 7 vendor formats (Turbomole, COSMOtherm, Gaussian, NwChem, ORCA, xtb, MolForge CSV)
  • CLI and web UI both shipped — free, no API key required
  • Batch migration tool for legacy archives — idempotent and resumable

See the full CHANGELOG.md.