One factory. One recipe. Every mfsig reproducible byte-for-byte.

Five input types, one factory, five output schemas

Every cohort flows through the same MFFactory.compute() dispatcher. No cohort-specific recipe overrides, no parallel pipelines, no copy-paste forks. Same recipe → same provenance leg → same kernel-input contract downstream.

The factory's only conditional logic is around object_kind + phase_label: monomers run 1 SCF, BSSE-CP dimers run 5 SCFs in v28-compliant Boys-Bernardi order. Everything else (functional, basis, solvent, grid, convergence) is read from the registry and is identical across cohorts.

MF_FACTORY_REGISTRY.json · versions.0.92.0

A single JSON file holds every knob the SCF pipeline reads. The file is SHA-256-hashed; that hash becomes the third leg of the four-way provenance stamp on every mfsig produced.

Source of truth: refs/MF_FACTORY_REGISTRY.json · v0.92.0 SHA-256 cd88bda5… · predecessor: 0.91.1 (def2-SVP).

Four-level fallback chain — NEVER silent

The default DIIS solver at the production basis handles 96/96 production drugs at L0. When it fails, the orchestrator records the fallback level in the mfsig so the reader can audit WHICH solver produced the number — no opaque retries, no quietly-different basis.

Raw σ, no empirical thresholds

The legacy σ-profile workflow ran q_sym through a 0.537 Å r_av spatial average then binned with a 0.0084 e/Ų H-bond threshold — both are empirical fits to small-molecule training sets, both inflate the distribution variance by ~30× over the raw electrostatic signal. v0.91.1+ removed both. The new sigma_moments are pure area-weighted statistics on the raw per-segment σ.

σ_i  = q_sym_i / area_i                # raw, per-segment, no spatial average
w_i  = area_i / Σ area_j               # tile-area weight
mean      = Σ w_i · σ_i
variance  = Σ w_i · (σ_i − mean)²
skewness  = m₃ / variance^(3/2)
kurtosis  = m₄ / variance²  −  3       # excess

Four-leg stamp on every mfsig

Every artefact carries four independent SHAs that, together, let a reader twenty years from now reconstruct the exact compute context byte-for-byte. Break any one, the file is rejected.

Plus an Ed25519 signature over the canonicalised payload (legs 1–4 are inputs to the SCF; the signature authenticates that the file was actually sealed by molforge's signing key). Verification is offline — no network call required to confirm authenticity, only the public key at /.well-known/mfsig-keys.json.

Five object kinds, one schema family

MFFactory.compute() accepts five values for object_kind. Each cohort has its own orchestrator that calls compute() one or more times with the appropriate phase_label + ghost_atoms. The output schema is identical across cohorts — only the suffix changes.

What's in every .mfsig.json

Five top-level blocks, the same structure on every cohort. The Anatomy page documents each field; this is the executive summary.

How the factory actually runs

The pipeline runs on H100 SECURE pods orchestrated through a single REST-API spawner. One generic runner script per cohort, one shared spawn script, one SHA-pinned image.

Verification you can run yourself

Every claim on this page is auditable from the bytes — no internal logs required.

# 1. Verify the recipe SHA matches what was used at compute time
sha256sum refs/MF_FACTORY_REGISTRY.json
# expect : cd88bda5… (v0.92.0)

# 2. Verify every mfsig in a cohort matches the registry recipe
python refs/_v091_regen_audit.py
# expect : ALL GREEN — AST + orchestrator alignment + parse + extract/ignore

# 3. Verify a single mfsig's 5-tier quality gates
python refs/_audit_v091_1_n100_quality.py
# tier 1: schema integrity   (5/5 expected)
# tier 2: physics sanity     (5/5)
# tier 3: klamt-purge enforcement
# tier 4: cross-file consistency (uniform recipe SHA)
# tier 5: sigma_moments recomputation (machine-epsilon)

# 4. Verify a downstream prediction comes from the right mfsig
jq '.provenance.{mf_factory_version,vendor_tree_sha256,registry_recipe_sha256,image_tag}' \
   refs/atlas_v031_v091_1/<inchi_key_14>.mfsig.json