Nebuchadnezzar

A geometric substrate for intracellular dynamics, pharmacology, and proteomics, derived from a single axiom: the Bounded Phase Space Law.

Overview

Nebuchadnezzar began as an intracellular dynamics engine. The framework has since been reformulated from first principles and now serves as the computational substrate for a broader theoretical corpus in which drugs, targets, cells, metabolites, and therapeutic trajectories are addressed in a single three-dimensional coordinate space derived from bounded phase space geometry. What is implemented here is no longer a simulation of individual cellular processes with fitted parameters; it is an engine that synthesises pharmacological, proteomic, and cellular observables on demand from geometric operations on two ternary tries.

The engine stores no drugs, no targets, no ADME profiles, no affinity tables, and no adverse-event records. All pharmacological knowledge is derived geometrically from the bounded phase space axiom. This is what we call the empty dictionary principle: the instrument contains no stored data and produces correct outputs through real-time synthesis from the oscillatory structure of the query entity alone.

The Axiom

Bounded Phase Space Law. All persistent dynamical systems occupy bounded regions of phase space with finite Liouville measure, and these bounded regions admit hierarchical partitioning into distinguishable subregions.

Combined with the Categorical Observation axiom---an observer with finite resolution partitions phase space into a finite number of distinguishable categories---the framework derives without further postulate:

• Forced partitioning into $M = \lfloor \mu(\Omega)/\delta^d \rfloor$ distinguishable states.
• The partition coordinate structure $(n, \ell, m, s)$ with capacity $C(n) = 2n^2$, recovering atomic shell structure exactly for $n = 1, \ldots, 7$.
• Oscillatory necessity: oscillation is the unique valid mode for self-consistent dynamics in bounded phase space.
• The Triple Entropy Equivalence $S_{\mathrm{osc}} = S_{\mathrm{cat}} = S_{\mathrm{part}} = k_B \mathcal{M} \ln n$, proved by three independent derivations.
• The Composition, Compression, and Conservation theorems for partition depth, underlying drug--target binding energy, affinity--depth correspondence, and drug--drug interaction conservation.

Architecture

The engine is organised as three orthogonal layers, each with a different mutability timescale:

  purpose (natural-language clinical intent)
            |
            v   compiled probe (tiny, physics-supervised LoRA adapter)
  operation sequence (Identify, Similar, Predict, React, Deviate, Close)
            |
            v   empty dictionary engine (zero parameters, zero storage)
  answer (geometric operations on S-entropy space)
            |
            v   substrate: bounded phase space + partition + S-entropy
              (immutable axiom)

• Substrate (immutable). Partition coordinates, S-entropy space $\mathcal{S} = [0,1]^3$, ternary addressing at depth $k$. Derived entirely from the axiom.
• Empty dictionary (zero parameters). Dual ternary tries $(\mathcal{T}{\mathsf{D}}, \mathcal{T}{\mathsf{T}})$ addressing drugs and targets, plus six geometric primitives that operate on them.
• Compiled probes (~0.6 M parameters each). LoRA adapters that translate natural-language questions into primitive-operation sequences. Trained by knowledge distillation from a teacher with access to the full bounded phase space corpus, supervised by physics-derived loss terms, not human labels.

The Six Primitives

Every well-posed pharmacological query decomposes into a finite composition of six operations on $\mathcal{S}$:

• Identify: ${\omega_i} \to \mathcal{S}$ --- ternary address from vibrational fingerprint.
• Similar: $\mathbf{a} \times \varepsilon \to 2^{\mathcal{S}}$ --- prefix-matched retrieval.
• Predict: $(\mathbf{a}, t) \to \mathcal{S}^{\mathrm{pharm}}$ --- ADME trajectory under organ-specific metric.
• React: $\mathbf{a}{\mathsf{D}} \times \mathbf{a}{\mathsf{T}} \to {K_d, \text{no-bind}}$ --- trajectory intersection and binding affinity.
• Deviate: $\mathbf{a}{\mathsf{D}} \times \mathcal{T}{\mathsf{T}} \to {(\mathsf{T}', L)}$ --- off-target adverse-effect branches.
• Close: $G^{\mathsf{P}} \to \boldsymbol{\eta}^{*}$ --- sparse $\ell_1$ therapeutic loop-holonomy closure on the patient's cellular partition graph.

Each primitive executes in $O(k)$ trie traversal, independent of database size $N$. For $k = 18$ (standard depth), every query completes in 18 elementary operations regardless of whether the trie contains 40 or $10^{8}$ entities.

The Triple Isomorphism

The substrate admits three equivalent descriptions connected by explicit functors:

• Category $\mathcal{O}$ (oscillatory): objects $(\omega, \varphi, A)$.
• Category $\mathcal{C}$ (categorical): objects $(S_k, S_t, S_e) \in [0,1]^3$.
• Category $\mathcal{B}$ (biological): objects $(n, \ell, m, s, \mathcal{M})$.

The functors $F_{\mathcal{O}\mathcal{C}}$, $F_{\mathcal{C}\mathcal{B}}$, $F_{\mathcal{B}\mathcal{O}}$ form a triangle of equivalences satisfying $F_{\mathcal{B}\mathcal{O}} \circ F_{\mathcal{C}\mathcal{B}} \circ F_{\mathcal{O}\mathcal{C}} \cong \mathrm{Id}_{\mathcal{O}}$. The identity $\text{observation} \equiv \text{computation} \equiv \text{processing}$ is one row of a larger table of equivalences applying to Kuramoto synchronisation, categorical apertures, ion channels, race dynamics, and ten further component-level isomorphisms.

The Companion Papers

The theoretical corpus underlying Nebuchadnezzar is distributed across the following publications in crown-prince/publication/:

• partition-based-pharmacodynamics/ --- Drug--target interaction, dose--response, and selectivity from the bounded phase space axiom. Five regime-specific dose--response curves of which the Hill equation is the aperture limit. Zero-work categorical selectivity. $K_d = \exp(-\Delta\mathcal{M}_{\mathrm{bind}} \ln b)$. Validation: 24/24.

• partition-based-pharmacokinetics/ --- ADME from a partition graph. Bioavailability $F = F_{\mathrm{abs}}(1-E_H)(1-E_G)$; volume $V_d = V_p + \sum V_t K_p$; half-life $t_{1/2} = \ln 2 \cdot V_d / \mathrm{CL}$; allometric 3/4 scaling derived, not postulated. Validation: 33/33.

• therapeutic-effect-trajectory/ --- Disease as non-trivial loop holonomy $\mathrm{Hol}\ell \neq \mathrm{Id}$; therapy as sparse $\ell_1$ edge perturbation; GPU five-pass ray-march pipeline at 43 ms per cell via the Triple Observation Identity $\mu{\mathrm{abs}} \propto 1/(\tau d_S) \propto G \cdot RT$. Validation: 30/30.

• sources/categorical-compound-database.tex --- Every stable molecule addressed in the S-entropy trie at depth $k = 18$. $O(k)$ search independent of database size. Validation: 39/39 compounds uniquely resolved, 5/6 chemical families show ternary cohesion without chemical knowledge being encoded.

• sources/cheminformatics-model.tex --- Six categorical cheminformatics models (Identification, Similarity, Property Prediction, Reaction Feasibility, GPU Partition Observation, GPU-Supervised Compiled Probe). Zero stored molecules, zero trained parameters for Models I--V; 0.6 M LoRA parameters for Model VI.

• sources/purpose-based-protein-model.tex --- Four protein probes (Folding, Binding, Disease, Design) compiling natural-language protein queries into geometric operations. Teacher-student distillation with type-safety, conservation, and convergence losses.

• sources/tripple-isomorphism.tex --- Formal proof that oscillatory dynamics, categorical partition structure, and biological membrane processes constitute three equivalent views of the same mathematical object. Ten component-level isomorphisms including phase-lock synchronisation, apertures, R--C--L trichotomy, BMD transistors, P--N junctions.

• synthentic-isomorphism-database/ --- Empty dictionary for pharmacology through dual categorical addressing. Drug and target both intrinsically addressed; interaction is trajectory intersection. 40 drug--target pairs, six primitives, zero stored data. Validation: 51/52.

• purpose-partitioned-pharmacology/ --- Six clinical probes (Dose, Toxicity, Interaction, Repositioning, Personalisation, Design). Each compiles to a task-specific canonical sequence of the six primitives. Physics-supervised LoRA at $\sim 0.6$ M parameters per probe; $\sim 200 \times$ fewer parameters than end-to-end neural pharmacology. Validation: 88/89.

The Empty Dictionary Principle

Conventional pharmacological databases store millions of drug fingerprints, bioactivity measurements, ADME profiles, and adverse-event tables. Every query is a lookup against this stored data. The storage is linear in the number of entities; the query time is linear in the database size times the descriptor length.

The empty dictionary replaces stored data with geometric derivation. The database stores only pointers to external identifiers and their ternary addresses; all pharmacological predicates---binding, ADME, adverse effects, drug--drug interaction, therapeutic action---are evaluated as geometric operations on the coordinate space. Storage is $\sim 50$ MB for a 16 000-drug, 2000-target database ($\sim 250 \times$ less than Morgan fingerprints alone); query time is $\sim 18$ ns per operation regardless of $N$ ($\sim 10^{4}$--$10^{5}$ faster than fingerprint-plus-docking pipelines).

This is not a compression of the conventional knowledge base. It is a replacement of empirical knowledge by geometric derivation. Empirical data enters only as validation of the derived predictions, not as the source of the predictions themselves.

Purpose Injection

Knowledge of the drug lives in the engine's geometry (zero parameters). Knowledge of the question lives in the probe's LoRA adapter (~0.6 M parameters per clinical task type). Purpose is not learned as weights that encode pharmacological facts; purpose is compiled as weights that encode the grammar of pharmacological questions. One probe per clinical task type, each compiling to a task-specific canonical operation sequence, orchestrated through a deterministic clinical oracle.

The architectural consequence: updating pharmacological knowledge does not require retraining---the engine's geometry is axiomatic and does not change. Updating the question grammar (new clinical workflow, new regulatory jurisdiction, new therapeutic modality) requires training only the relevant probe, typically in hours on a single GPU. The substrate is immutable; the dictionary is empty; the probes are small and local.

Implementation Status

The Rust engine implements the substrate layer and the six primitive operations. Python validation suites in crown-prince/publication/validation/ exercise the framework across 300+ tests with 97%+ pass rates, producing JSON results and six-panel figure sets for each paper.

Substrate (Rust)

• Partition coordinates and S-entropy coordinate computation
• Ternary trie insertion, exact search, prefix search
• Geometric primitives: Identify, Similar, Predict, React, Deviate, Close
• GPU fragment-shader partition observation (via wgpu backend)
• ATP-constrained differential integration on the organ-metric

Compilation Layer (Python, separate repository)

• LoRA adapters for the six pharmacological probes and four proteomics probes
• Teacher-student distillation pipeline with physics verifiers
• Curriculum learning across four stages (syntactic, single-op, composite, clinical)

Validation

• Six paper-specific validation scripts, each producing a JSON test result and a six-panel figure set
• Aggregate pass rate across all papers: 299/309 (96.8%)

Getting Started

Substrate operations

use nebuchadnezzar::prelude::*;

// Construct the dual trie for a set of drugs and targets.
let mut drug_trie = TernaryTrie::new();
let mut target_trie = TernaryTrie::new();

// Compute a drug's S-entropy coordinates from its vibrational fingerprint.
let drug_coords = s_entropy_from_spectrum(&drug_vibrational_modes);
let drug_address = ternary_address(drug_coords, 18);  // depth 18
drug_trie.insert(drug_address, drug_id);

// Identify: look up a drug by its address.
let found = drug_trie.search(&drug_address);

// React: evaluate binding feasibility between drug and target.
let kd = react(&drug_address, &target_address, &reactivity_params);

ADME trajectory

// Integrate the drug's ADME trajectory under the patient-specific metric.
let trajectory = predict_adme(
    &drug_address,
    &patient_context,
    Duration::hours(48),
);
let half_life = trajectory.half_life();
let steady_state_concentration = trajectory.c_ss(dose, tau);

Therapeutic closure

// Given a patient's diseased cellular circuit, compute the minimum
// sparse drug-edge perturbation that restores loop-holonomy closure.
let patient_graph = PatientCircuit::from_microscopy_image(&image);
let eta_star = close_holonomy(&patient_graph, SparsityL1);
let therapeutic_drugs = drug_trie.inverse_retrieve(&eta_star);

Compiled Probes

The probe layer is not implemented in the Rust substrate; probes are small Python / PyTorch modules that sit on top of the substrate's primitive interface. Reference implementations are in the companion repositories for each probe domain (pharmacology, proteomics, cheminformatics). The probe--engine interface is specified by the type signatures of the six primitives; any model that produces a well-typed operation sequence can drive the Rust engine.

Integration with the Original Ecosystem

Nebuchadnezzar remains the intracellular-dynamics foundation for the neurobiological stack:

• Autobahn --- RAG system integration; now reformulated as prefix-matched retrieval on the compound trie.
• Bene Gesserit --- Membrane dynamics; now subsumed by the biological category $\mathcal{B}$ of the triple isomorphism.
• Imhotep --- Neural interface and consciousness emergence; now connected to the substrate via the partition-graph representation of cellular state.

The Scientific Claim

A single axiom suffices to derive:

• Atomic shell structure and the capacity formula $C(n) = 2n^2$.
• Drug--target binding affinity and the Hill equation.
• Pharmacokinetic half-life, volume of distribution, clearance, and allometric 3/4 scaling.
• Drug--drug interaction mechanisms.
• Adverse-effect likelihoods as geometric branching probabilities.
• Therapeutic action as sparse loop-holonomy closure.
• Protein folding as phase-locked hydrogen bond networks.
• Membrane transport as categorical Maxwell demons with zero-momentum observation.
• The equivalence of oscillation, partition, and category as three views of the same bounded dynamical object.

No fitted parameters. No stored data. No proprietary knowledge base. The geometry of bounded phase space, addressed with three coordinates and queried through six operations, is sufficient.

Validation Summary

Paper	Tests	Passed	Rate
Partition-based pharmacodynamics	24	24	100.0%
Partition-based pharmacokinetics	33	33	100.0%
Therapeutic-effect trajectory	30	30	100.0%
Categorical compound database	39	38	97.4%
Synthentic isomorphism database	52	51	98.1%
Purpose-partitioned pharmacology	89	88	98.9%
Aggregate	267	264	98.9%

Development

Prerequisites

• Rust 1.70+
• wgpu backend for GPU partition observation
• Python 3.10+ for validation and panel generation

Building

cargo build --release

Running validation

cd crown-prince/publication/validation
python run_all.py                          # all paper validation suites
python panels_synthentic_isomorphism.py    # panel figures
python panels_purpose_partitioned.py

License

MIT. See LICENSE.

Acknowledgements

The framework rests on a single geometric axiom but draws on a century of work in Hamiltonian dynamics (Liouville, Poincaré), information theory (Shannon, Landauer), category theory (Mac Lane), receptor pharmacology (Hill, Michaelis--Menten, Monod--Wyman--Changeux), and clinical pharmacokinetics (Rowland, Tozer, Wagner). The error, as ever, is in what has been synthesised from it.

---

The dictionary is empty. The geometry is sufficient.