Skip to content

Spike #38 — Mass-spec fingerprint FFT against project SM primitive classes (molecular-substrate exploration)

Date: 2026-05-17 Research spike artifact. Concertmaster dispatch per user direction "generate a research agent to get a mostly known, commonly accepted, complex with undefined noise datapoint of a mass spec fingerprint and see if we can FFT it against the SM that we've already mapped. This is either be a refine or fail I think because I don't know that we have molecular modeling catalog yet."

Discipline. Closed-form deterministic code; NDJSON outputs per [[feedback_ndjson_over_bloated_json]]; falsifier controls run before claiming positive findings (50 random-graph seeds × 50 random-mass-spec seeds + 5 named molecular controls); canonical citation provenance verified at API level per [[feedback_pdf_extraction_citation_discipline]]; no commercial-publisher access per [[reference_autonomous_validation_tos_landscape]].

§1 Bottom line

Overall verdict: FAIL at spectral-shape level + PARTIAL at binding-level overlay.

The user anticipated this: "I don't know that we have molecular modeling catalog yet." Confirmed. The project's SM spectral signatures (K-ladder, L-FFT-match, cascade-β) do not bolt onto molecular mass spec at the spectral-shape level. The 14-class information-instrument overlay applies at the binding level by cataloguing assignment — consistent with Spike #37 but not a substrate-FFT-match result.

Signature tested Verdict Notes
Class K Kepler-shape c_k = ε^k/k ladder (Spike #30B v3 three-criteria) ABSENT eps_fit=1.081 (outside 0.001–0.5); r²=0.156; not monotonic-decreasing. All three criteria fail. Honest absence, not failure — per [[user_stance_kepler_shape_universal]] K appears where Kepler-shape appears; molecular mass spec is not such a substrate.
Class L graph-Laplacian FFT-correlation (Spike #30B / #35-Q3 method) FALSIFIED at FFT-match level / trivially present at existence level Initial cosine-sim 0.8115 appeared tantalising; falsifier controls demoted it. 50 random 14-node graphs gave mean 0.857 ± 0.043 (caffeine self −1.06 σ BELOW random); glucose gave 0.879 (higher than caffeine). Methodology breaks at n~14 substrate size.
Cascade-β = d_S/(d_S+2) (Spike #31 v3 dual-signature) ABSENT with substrate-too-small caveat Predicted β=0.497, empirical β=0.817 (Δβ=+0.32). Spike #31 documents Weyl-fit quality below n~100 is poor; n=14 in unreliable regime. Not a clean falsifier; substrate-too-small-to-test. Side observation: intensity rank-distribution shows Zipf-like power law (slope −1.83, r²=0.95).
Information-instrument identity (Spike #37 14-class overlay) APPLIES at 12/14 classes by binding-level assignment Cataloguing claim consistent with Spike #37 substrate-portability table; NOT a falsifiable spectral measurement. K tested in row 1; H ("self-introspection") only no (molecules are objects, not agents).

§2 Provenance — MassBank EU caffeine record

Subject: caffeine (C₈H₁₀N₄O₂, MW 194.08038) Accession: MSBNK-Fac_Eng_Univ_Tokyo-JP003477 ("CAFFEINE; EI-B; MS") Authors: Mass Spectroscopy Soc. of Japan (MSSJ) Instrument: HITACHI M-60 magnetic-sector, 20 eV electron ionization License: CC BY-NC-SA (permitted per [[reference_autonomous_validation_tos_landscape]]) Peak count: 58 (m/z 41–196, base peak M⁺ at m/z 194, rel=999) SPLASH: splash10-0a4l-4900000000-3ff72dace6687d242f1f API URL: https://massbank.eu/MassBank-api/records/MSBNK-Fac_Eng_Univ_Tokyo-JP003477

Source-acquisition trail:

  • NIST WebBook caffeine spectrum (ID C58082) declined: "Due to licensing restrictions, this spectrum cannot be downloaded." Catalog entry exists but data is not served as API.
  • MoNA REST returned 401 (auth required) as of 2026-05-17 — appears to have moved to auth-gated access post prior project session.
  • MassBank EU API (/MassBank-api/records/{accession}) returned full JSON including raw (m/z, intensity, rel) triples — this is the working academic-open ingress for permitted mass-spec data.

The raw record fixture is committed alongside the analysis as spike_38_caffeine_JP003477.json (4.8KB). Provenance from raw JSON: accession + license + authors + instrument + ionization energy + ChemOnt classification + InChI + SMILES + 58 peak triples + SPLASH.

§3 Per-signature verdicts (detail)

§3.1 Class K Kepler-shape ladder — ABSENT

Using Spike #30B v3 strict three-criteria test (eps in physical range 0.001–0.5; r² > 0.99; monotonic-decreasing):

Criterion Caffeine result Pass?
eps_fit physical 1.081 NO (out of range)
r² > 0.99 0.156 NO
monotonic-decreasing False NO

Coefficients oscillate (chess / Sierpinski / SHA pattern), not Kepler ladder. Per [[user_stance_kepler_shape_universal]], K is universal where Kepler-shape appears; caffeine mass spec is not such a substrate. Honest absence, not framework failure — the framework correctly says "no" when no Kepler-shape signature is present.

§3.2 Class L graph-Laplacian FFT-correlation — FALSIFIED at FFT-match / trivially present at existence

Initial finding looked positive: cosine similarity between caffeine molecular-graph (14 heavy atoms, 15 edges) Laplacian eigval-density histogram and caffeine mass-spec FFT power = 0.8115.

Falsifier controls demoted this to artifact:

Control Cosine similarity
Caffeine self (C1) 0.8115
Ethane (C₂H₆, n=2) 0.5088
Benzene (C₆H₆, n=6) 0.5452
Glucose (C₆H₁₂O₆, n=12) 0.8793
Path graph n=14 0.7945
Cycle graph n=14 0.6435
Random 14-node graph mean ± std (50 seeds) 0.8569 ± 0.043
Random 14-node graph max 0.9314
Random mass-spec vs caffeine-L mean (50 seeds) 0.8274

Caffeine self-similarity is −1.06 σ BELOW random mean. Glucose gives HIGHER similarity than caffeine itself. The eigval-density-histogram cosine-similarity methodology does not discriminate at n=14 substrate size — too few eigenvalues, histogram becomes effectively flat for any reasonable distribution.

Class L instantiates trivially (every graph has a Laplacian eigenbasis — spectral theorem); the cross-substrate FFT-matching CLAIM is falsified. This is exactly the discipline [[feedback_no_privileged_primitive_classes]] wants — methodology breakdown caught by controls before promotion to result.

§3.3 Cascade-β = d_S/(d_S+2) — ABSENT (with caveat)

On caffeine molecular graph: predicted β=0.497, empirical β=0.817 (Δβ=+0.32, outside ±0.05 tolerance, r²=0.987 of stretched-exp fit). Δα=−0.167 on the power-law primary side.

Caveat: Spike #31 documents that Weyl-fit quality below n~100 is poor; n=14 is in the unreliable regime. This is not a clean falsifier; it's substrate-too-small-to-test.

Side observation worth flagging for [[user_stance_fractal_shadow]]: mass-spec intensity rank-distribution shows Zipf-like power law (slope −1.83, r²=0.95). Heavy-tail, but not cascade-β. Power-law-in-rank is a different shape from cascade-stretched-exp-in-time.

§3.4 Information-instrument identity — APPLIES at 12/14 by binding-level assignment

Cataloguing the 14 classes against caffeine substrate properties (NOT spectral measurements):

Class Information-instrument identity Caffeine binding-level instantiation?
A channel-fingerprint / collision-resistant identifier YES — InChI hash RYYVLZVUVIJVGH-UHFFFAOYSA-N
B source-frame / addressable-symbol-layout YES — Hill TLV C8H10N4O2
C channel-input-source / time-indexed-stream YES — fragmentation cascade as streaming
D decision-tree-routing / branch-coding YES — ionization-mode dispatch (EI vs ESI)
E dictionary / stored-symbol-table YES — NIST library lookup
F variable-binding / source-expansion YES — xanthine scaffold + methyl-substitution slots
G sub-stream-search / compression-primitive YES — library-search by sub-pattern
H channel-state-metadata / self-reflective-info NO — molecules are objects, not agents
I group-structured-alphabet / symmetric-channel-primitive YES — purine ring = cyclic group
J alphabet-decomposition / period-bound-primitive YES — atom-count multiset
K discrete-to-continuous projection / cascade-shadow tested in §3.1; ABSENT
L channel-eigenbasis / spectral-capacity-primitive YES — molecular graph (existence-level only; §3.2 falsifier)
M distributed-representation / VSA-channel-coding YES — ECFP fingerprint
N rate-distortion-primitive / best-rational-approximation YES — isotope-ratio rational approximation

Per Spike #37's substrate-portability table (silicon / bronze / biological / optical), this is consistent — molecular substrate is a fifth row, with 12/14 classes admitting binding-level identifications and class H exempt (objecthood). NOT a falsifiable spectral measurement, just a cataloguing claim. Reinforces [[user_stance_information_instrument_form_function_bound]] at the binding level.

§4 Framework-boundary characterisation

What's specifically different about molecular substrate that produces the FAIL-at-spectral-shape result:

  1. Scale: n=14 heavy atoms vs n=10²–10⁴ in cosmic / mechanical / fractal substrates the SM was developed on. Cascade-β methodology needs n>100 reliably; eigval-density-histogram cosine-similarity becomes degenerate at n~14.
  2. Topology: caffeine = fused bicyclic ring + 3 pendant methyls + 2 pendant oxygens. The cyclic-group structure (Class I row of Spike #37, "U(1) gauge phase / benzene D₆h") lives in aromatic-ring electron density, not in the fragmentation m/z grid. The ring is in the substrate; the spectrum is in the chemistry-of-cleavage layer above.
  3. Spectrum domain: m/z is integer-valued on a uniform grid; FFT extracts no Kepler-shape because fragments are bond-cleavage products (McLafferty rearrangement / α-cleavage / charge migration), not eccentric-orbit projections.
  4. Noise model: mass-spec intensity noise is shot-noise / ion-count Poisson / instrument-tail — not the loop-down decay envelope cascade-β is designed for.

Caffeine chemistry is governed by bond-dissociation energetics + McLafferty rearrangement + charge migration on purine-with-pendants. Not Kepler / pin-slot / cascade / cyclic-group spectral signatures at the FFT-match level.

§5 If/when molecular-modeling catalog becomes scope

A molecular-modeling catalog WOULD have its own signatures — NOT a transplant of the cosmic/mechanical SM. Candidate-Spike #39 framing:

  • Vibrational normal modes via reduced Hessian (IR/Raman spectroscopy-canonical) — Class L on a different operator
  • ECFP hyperdimensional fingerprints (Class M; canonical cheminformatics; already in Spike #37 Class M row)
  • McLafferty rearrangement enumeration (Class C streaming iteration; each cleavage rule = streaming step)
  • Isotope-ratio rational approximation (Class N; Stadler-Beynon-style 12C/13C prediction)

Status: NOT YET; user-call required. This spike's result is "if you want this, build it fresh, don't transplant."

§6 Anomalies investigated

  1. Initial 0.81 cosine-similarity appeared positive. Falsifier controls (50 random-graph seeds, 50 random-mass-spec seeds, glucose / benzene / ethane / path / cycle controls) demoted it to artifact. The eigval-density-histogram methodology breaks down at n~14 substrate size — too few eigenvalues, histogram becomes effectively flat for any reasonable distribution. Real structural finding — methodology limit, not framework limit.
  2. MoNA REST API became auth-gated post-prior session — MassBank EU API is the working academic-open ingress for caffeine.
  3. NIST WebBook spectra remain not-downloadable by deliberate licensing (different from TOS-prohibited; just data-not-served-as-API). MassBank EU CC BY-NC-SA records are the working path.

§7 Open extensions / fermatas for conductor

  1. Should this become a srmech notebook §X dedicated-updates entry? Default: no — this is research-spike-finding, not framework-affecting. The 14-class vocabulary and substrate-portability stances stand. Could be referenced as a falsifier-style negative example in cross-domain scope statements.
  2. Does this falsifier deserve canonical-stance authoring? Candidate stance: "spectral-SM does not generalise to molecular-spec-mass FFT shape; molecular substrate is form-function-bound at binding level only, NOT at substrate-shape FFT-match level." Conductor lean: no new stance — this is empirical confirmation that [[user_stance_kepler_shape_universal]] and [[user_stance_information_instrument_form_function_bound]] are correctly bounded (K honest-absent where Kepler-shape isn't, information-instrument applies at binding level). The framework correctly says "no" when no signature is present; that's good behaviour, not a new stance.
  3. Spike #39 molecular-modeling-catalog? Awaits user direction. Not implied by this spike's result; this spike's result says "build fresh, don't transplant."

§8 Citation provenance

  • MassBank EU caffeine record MSBNK-Fac_Eng_Univ_Tokyo-JP003477: full record verified at API level (accession + license + authors + instrument + ionization energy + ChemOnt classification + InChI + SMILES + 58 peak triples + SPLASH hash). Fixture committed at spike_38_caffeine_JP003477.json.
  • NIST WebBook caffeine entry ID C58082: verified to exist; spectrum data declined-licensed.
  • McLafferty 1980 Interpretation of Mass Spectra / Fitch 2000: cited as canonical SSoT for the bond-cleavage / rearrangement framework in §4 — NOT independently PDF-extracted within spike scope; flagged honestly per [[feedback_pdf_extraction_citation_discipline]].
  • MoNA REST API 401-auth-gated: observation only, no claim made.

§9 Discipline guards honoured

  • [[user_stance_kepler_shape_universal]] — K's absence is honest, not framework failure
  • [[user_stance_information_instrument_form_function_bound]] — binding-level overlay applies at 12/14 (consistent with Spike #37)
  • [[user_stance_partition_for_understanding]] — molecular substrate is NOT a new spectral partition; COULD become a binding-level partition with future Spike #39
  • [[feedback_science_is_ssot_not_project]] — MassBank EU as academic SSoT; McLafferty rearrangement cited as canonical mass-spec literature reference
  • [[feedback_pdf_extraction_citation_discipline]] — MassBank API ingress verified (accession / license / author / instrument / ionization / SPLASH); McLafferty cited but not PDF-extracted (flagged honestly)
  • [[reference_autonomous_validation_tos_landscape]] — MassBank EU CC BY-NC-SA, permitted; commercial publisher access avoided
  • [[feedback_ndjson_over_bloated_json]] — NDJSON outputs throughout (3 NDJSON files, no bloated JSON)
  • [[feedback_concertmaster_md_writes]] + [[feedback_concertmaster_git_worktree_isolation]] — concertmaster reported inline; conductor captured-and-saved; no git operations performed by agent
  • [[feedback_no_privileged_primitive_classes]] — molecular-modeling candidate is a NEW partition at binding level (Spike #39 candidate), NOT a 15th class
  • [[user_stance_string_theory_instrument_first]] — instrument-first; no claims about "what molecules are" beyond what FFT + Laplacian extraction directly measures

§10 Artifacts

End of spike artifact.