Skip to content

Spike #39 — Molecular-modeling AMSC catalog scoping (per Spike #38 follow-up)

Date: 2026-05-17 Research spike artifact. Concertmaster dispatch per user direction following Spike #38's framework-boundary finding (docs/srmech/notes/spike_38_mass_spec_fft_vs_sm_2026-05-17.md): "build fresh, don't transplant cosmic/mechanical SM." Scoping spike — design proposal + small POC catalog (1-2 attested rows for ONE chosen signature), NOT full implementation.

Discipline. Closed-form deterministic chain; NDJSON outputs per [[feedback_ndjson_over_bloated_json]]; NIST CCCBDB autonomously verified; textbook citations flagged honestly per [[feedback_pdf_extraction_citation_discipline]]; no commercial-publisher access per [[reference_autonomous_validation_tos_landscape]]; molecular-modeling is a NEW PARTITION at substrate-binding layer per [[feedback_no_privileged_primitive_classes]] — vocabulary stays at 14 classes A–N.

§1 Bottom line

Vibrational normal modes via Class L reduced-Hessian eigendecomposition chosen for POC. Two attested rows shipped (H₂O bent XY2 + CO₂ linear XY2). Chain-output verified bit-exact against stored row predictions. Falsification gauge in place: predicted-vs-measured residual at ~5% (stretches) / ~5–14% (bends), within expected GVFF-only accuracy floor.

Not a 15th class. Molecular-modeling is a NEW PARTITION at substrate-binding layer (fifth row alongside silicon / bronze / biological / optical from Spike #37). The 14-class vocabulary stays at 14; what changes is the table of substrate-instantiations of those classes.

Catalog structure validates. The descriptor.toml + schema.json + row.ndjson pattern from pi_digits translates cleanly to molecular substrate. The [catalog.operator_chain] step composition (Class J atom-multiset → Class L dynamical-matrix eigvals) is well-formed and TOML-parseable.

Mode H₂O predicted (cm⁻¹) H₂O NIST Residual % Error
ν₁ sym stretch 3839.53 3657.0 +182.5 +4.99%
ν₂ bend 1733.53 1595.0 +138.5 +8.69%
ν₃ asym stretch 3942.18 3756.0 +186.2 +4.96%
Mode CO₂ predicted (cm⁻¹) CO₂ NIST Residual % Error
ν₁ sym stretch 1383.69 1333.0 +50.7 +3.80%
ν₂ bend 573.42 667.0 −93.6 −14.03%
ν₃ asym stretch 2458.99 2349.0 +110.0 +4.68%

CO₂ bend systematic underestimate is the canonical GVFF limit (Urey-Bradley cross-coupling captures the missing carbonyl repulsion across the molecule). H₂O symmetric / asymmetric stretch both overshoot by ~5%, suggesting the f_r value calibrated against slightly different cross-terms than the minimal model includes. Documented honestly in row notes — the POC proves the catalog shape works; better force-field choices ratchet to <1% if needed.

§2 Design proposal for all four candidate signatures

§2.1 Vibrational normal modes (Class L on dynamical matrix) — CHOSEN POC

Substrate primitive: mass-weighted Hessian eigendecomposition. Wilson-Decius-Cross §6 closed-form secular equation for triatomic XY2 (bent + linear branches).

Chain composition: (Class J atom-multiset → mass vector) ∘ (Class L dynamical-matrix eigvals) → predicted frequencies cm⁻¹.

SSoT: Wilson-Decius-Cross Molecular Vibrations (1955) §6; Herzberg Molecular Spectra II (1945) Tables 35–37; Shimanouchi NSRDS-NBS 39 (DOI 10.6028/NBS.NSRDS.39); NIST CCCBDB (cccbdb.nist.gov) — open US-government data aggregator; autonomously verified for H₂O and CO₂ entries 2026-05-17.

Falsification gauge: measured-vs-predicted residual. POC at ~5% (stretches) / ~10% (bends) is canonical GVFF floor; row notes flag this honestly. Production catalog can ratchet to Urey-Bradley (~1–2%) or DFT-computed Hessian (<1%).

Scope expansion path: triatomic XY2 (POC) → triatomic XYZ asymmetric (e.g., HCN) → tetraatomic (NH₃, CH₄) → polyatomic (full 3N × 3N Cartesian Hessian; numerical eigendecomposition).

§2.2 ECFP hyperdimensional fingerprints (Class M)

Substrate primitive: Extended-Connectivity Fingerprint (Rogers & Hahn 2010). Iterative atom-environment hashing producing fixed-length binary vectors. Already named in Spike #37 Class M row.

Chain composition: (SMILES parse) ∘ (atom-environment iterative hashing) ∘ (Class M VSA bundle/bind) → fingerprint vector.

SSoT: Rogers & Hahn 2010, Extended-Connectivity Fingerprints, J. Chem. Inf. Model. 50(5), 742–754. DOI 10.1021/ci100050t — canonical ECFP reference (ACS publication; PDF behind paywall — DOI metadata only, per [[reference_autonomous_validation_tos_landscape]] commercial-publisher exclusion). RDKit (open-source BSD) is the canonical implementation; PubChem (NIH) is the open SMILES corpus.

Catalog row shape: InChIKey + SMILES + ECFP-radius + ECFP-bit-length + chain-computed-fingerprint-hash. Cross-implementation parity test against srmech.amsc.m.bind/bundle is the structural verification.

Caveat (Fermata-1): ECFP uses MurmurHash; srmech Class M may use a SHA-based VSA — cross-implementation parity is structural-not-byte-exact. Does srmech ship a "molecular-ECFP-compatible" Class M variant, or document structural equivalence using its own VSA scheme? Catalog design choice — affects whether ecfp_fingerprints rows store byte-exact fingerprint hashes.

Scope path: small molecules (POC) → PubChem-30M coverage.

§2.3 McLafferty rearrangement enumeration (Class C streaming)

Substrate primitive: mass-spec fragmentation rule enumeration. Each cleavage rule is one streaming-iterator step (Class C primitive).

Chain composition: (SMILES → molecular graph) ∘ (Class C stream cleavage-rule applications) ∘ (collect fragments) → predicted fragment-mass list.

SSoT: McLafferty Interpretation of Mass Spectra (4th ed., University Science Books, 1993). Textbook canonical; not arXiv-mirrored; not PDF-extractable autonomously per [[feedback_pdf_extraction_citation_discipline]]. Cross-validate against MassBank EU records (CC BY-NC-SA permitted) — Spike #38 caffeine fixture is a ready test substrate.

Caveat: McLafferty rules are heuristic (textbook-codified rules of thumb, not first-principles). Falsification gauge is "expected" not "deterministic" — rules generate fragment candidates; observed mass spec confirms a subset. Structurally weaker than vibrational modes (closed-form) or ECFP (canonical algorithm). The chain works; the falsification is statistical.

Scope path: caffeine POC reuse from Spike #38 → 100-molecule MassBank curated set.

§2.4 Isotope-ratio rational approximation (Class N)

Substrate primitive: isotope-pattern prediction via best-rational-approximation. The 12C/13C ratio in a mass-spec isotope cluster IS a Class N rational-approximation problem.

Chain composition: (Class J atom-multiset of molecular formula) ∘ (Class N rational-approximation of natural-abundance ratios) → predicted isotope-pattern relative intensities.

SSoT: Beynon 1960 Mass Spectrometry and Its Applications to Organic Chemistry (textbook canonical); IUPAC isotope natural-abundance tables (open); NIST atomic-weights table (open). Caffeine fixture from Spike #38 has M / M+1 ratio observable: M (m/z 194, rel 999) vs M+1 (m/z 195, rel 167) — ratio 0.167, with natural-abundance prediction ~0.087 for 12C/13C alone (M+1 also fed by ¹⁵N + ²H contributions).

Caveat (Fermata-2): srmech Class N primitives currently target pi_cascade / continued-fraction-convergent shapes. Isotope-ratio framing IS Class N rational-approximation, but may need a different entry point — srmech.amsc.n.isotope_ratio_predict() or similar. Catalog design choice.

Scope path: small organics (POC) → metabolite-class coverage.

§3 POC results — vibrational normal modes for H₂O and CO₂

Computed via Wilson-Decius-Cross §6 closed-form GF eigendecomposition (Class L on mass-weighted Hessian; NOT molecular-graph Laplacian — same primitive, different operator, per [[user_stance_information_instrument_form_function_bound]] substrate-portability).

Chain bit-exact reproducibility verified — running normal_mode_frequencies(spec) over each row's stored inputs produces predicted_frequencies_cm-1 bit-identical to stored values. The falsification infrastructure works as designed (per [[feedback_every_doc_edit_faces_falsification]]).

Results table see §1. CO₂ bend (−14% residual) is the GVFF-limit anomaly documented in §6.

§4 Citation provenance

Source Type Verified? Status
Wilson-Decius-Cross Molecular Vibrations (1955) §6 textbook WorldCat author+title only; full PDF not autonomously fetched flagged honestly — textbook, not arXiv
Herzberg Molecular Spectra II (1945) Tables 35–37 textbook citation through NIST CCCBDB flagged honestly — textbook attribution via aggregator
Shimanouchi NSRDS-NBS 39 (DOI 10.6028/NBS.NSRDS.39) NIST gov publication WebFetched PDF >10 MB; one-shot verify failed NIST-attested via CCCBDB; direct PDF extraction failed (size limit)
NIST CCCBDB H₂O entry gov web WebFetched 2026-05-17 — ν₁=3657, ν₂=1595 cm⁻¹ confirmed verified ✓
NIST CCCBDB CO₂ entry gov web WebFetched 2026-05-17 — ν₁=1333, ν₂=667, ν₃=2349 cm⁻¹ confirmed verified ✓
Huber-Herzberg 1979 H₂O ν₃ textbook (via CCCBDB) cited as source for ν₃=3756; CCCBDB attribution confirmed flagged honestly — textbook

The two NIST CCCBDB entries are autonomously verified. Textbook citations (Wilson-Decius-Cross, Herzberg, Shimanouchi PDF, Huber-Herzberg) are anchored via CCCBDB metadata but not independently PDF-extracted — matches [[feedback_pdf_extraction_citation_discipline]] honesty discipline (empirical anchors verified; textbook anchors flagged).

§5 Anomalies investigated

  1. CO₂ bend frequency underestimated by 14% with canonical Herzberg force constants. Stretches reproduce at ~4%, but the bend f_θ = 0.5712 mdyne·Å/rad² predicts 573 cm⁻¹ vs measured 667 cm⁻¹. Known GVFF limitation — the CO₂ bend benefits significantly from Urey-Bradley terms (carbonyl repulsion across the molecule). Not a chain bug; an FF-modeling limit. Documented in row notes.

  2. H₂O sym/asym stretches BOTH overshoot by ~5%. Suggests the canonical f_r = 8.454 mdyne/Å is calibrated for slightly different f_rr / f_rθ cross-terms than the minimal model includes. Stronger-than-expected systematic offset — single-figure adjustment would tune both. Documented in row notes.

  3. NIST Shimanouchi PDF was >10 MB and failed WebFetch one-shot verification. Citation via CCCBDB metadata is solid (NIST aggregator confirms attribution); direct PDF extraction not feasible within spike scope. Flagged honestly.

  4. Python 3.14 default cp1252 console encoding caused initial display errors with UTF-8 special chars. Required python -X utf8 invocation. Minor — does not affect bit-exact NDJSON output (UTF-8 LF on disk).

§6 Fermatas for conductor

  1. (Fermata-1) ECFP MurmurHash vs srmech.amsc.m VSA hash. Cross-implementation parity is structural-not-byte-exact. Does srmech ship a "molecular-ECFP-compatible" Class M variant, or document structural equivalence and use its own scheme? Conductor lean: structural equivalence + document, don't proliferate Class M variants per [[feedback_no_privileged_primitive_classes]].

  2. (Fermata-2) Class N isotope-ratio entry point. Needs srmech.amsc.n.isotope_ratio_predict() (or similar) added when this catalog ships in production. Conductor lean: add the entry point when §2.4 catalog actually lands; deferred to Phase C2 catalog-expansion work.

  3. (Fermata-3) Force-field SSoT for production vibrational_modes catalog. GVFF (POC; ~5%) → refined GVFF with Urey-Bradley (~1–2%) → DFT-computed Hessian (<1%, per-row method citation required). Conductor lean: GVFF for chain-primitive cleanliness; DFT rows would shift chain shape from "Class L closed-form" to "Class L numerical on stored DFT-Hessian." Both legitimate; the cleanest-primitive choice is GVFF for the closed-form-chain SSoT row; DFT for production-accuracy rows.

  4. (Fermata-4) Caffeine reuse path. Spike #38's caffeine MassBank fixture is the natural test substrate for §2.3 (McLafferty) and §2.4 (isotope-ratio). Conductor lean: reuse the Spike #38 fixture for §2.3 / §2.4 POCs; cite the spike's fixture path as the test-substrate provenance.

  5. (Fermata-5) Catalog wiring scope. The POC files in docs/srmech/notes/ are NOT wired into srmech.amsc.catalog. Actual wiring (placing files under srmech/amsc/attested/vibrational_modes/ + writing Class L dynamical_matrix_eigvals_xy2 primitive in srmech.amsc.l or new submodule + adding tests) is a real engineering ship — Phase C2 catalog-expansion work. Per [[feedback_no_mvp_framing]] full-coverage shipping, the version that lands should cover all four candidate signatures (or explicitly defer §2.3/§2.4 to follow-on tasks with named-event resolution). Awaits user direction on whether to dispatch Phase C2 work now or hold for sprint-cycle planning.

§7 Discipline guards honoured

  • [[feedback_no_privileged_primitive_classes]] — NOT a 15th class; new partition at substrate-binding layer
  • [[user_stance_information_instrument_form_function_bound]] — Class L on different operator (mass-weighted Hessian, NOT molecular graph) demonstrates substrate-portable identity per Spike #37
  • [[user_stance_partition_for_understanding]] — molecular-modeling is a substrate-binding-level partition complementary to existing partitions (algebraic / kinematic / observable / information-instrument)
  • [[reference_autonomous_validation_tos_landscape]] — NIST CCCBDB (gov) verified; textbook citations anchored via CCCBDB; no commercial-publisher autonomous access (ACS DOI metadata only)
  • [[feedback_pdf_extraction_citation_discipline]] — NIST CCCBDB web entries verified by WebFetch; textbook anchors flagged honestly as not-PDF-extracted
  • [[feedback_science_is_ssot_not_project]] — Wilson-Decius-Cross + Herzberg + Shimanouchi + NIST CCCBDB as canonical SSoT; not any srmech sub-project
  • [[feedback_ndjson_over_bloated_json]] — NDJSON outputs; TOML descriptor for descriptor-shaped data
  • [[feedback_concertmaster_md_writes]] — concertmaster returned findings inline; conductor captured-and-saved this note
  • [[feedback_concertmaster_git_worktree_isolation]] — agent performed zero git operations; all work in D:\temp\spike_39\
  • [[feedback_every_doc_edit_faces_falsification]] — chain spec lives in catalog config; chain bit-exact reproducibility verified for both POC rows
  • [[user_stance_string_theory_instrument_first]] — instrument-first; no claims about "what molecules are" beyond what reduced-Hessian eigendecomposition directly computes

§8 Artifacts

End of spike artifact.