AETHER

Adaptive Endocrine Transformer Heads with Emotional Regulation

An Astrocyte-Modulated Language Model Architecture Employing Neurochemical Conductance Gating, Context-Adaptive Head Generation, Cascading Biological Hierarchy, and Intrinsic Pharmacokinetic Memory

USPTO Provisional Patent Application

FILING INFORMATION

63/988,485
February 23, 2026
Marjorie McCubbins
Micro Entity
63/939,190 & 63/962,385
PATENT PENDING
Download Full PDF (39 pages)

THE CORE INNOVATION

"Every existing AI architecture models only the neuron. We built the other half of the brain."

The human brain contains equal numbers of neurons and astrocytes. AI has been building with half the blueprint.

The AETHER architecture introduces astrocyte networks to artificial intelligence. In biological brains, astrocytes are the modulatory glial network that governs which neurons fire, how strongly, and in what temporal patterns through neurochemical signaling — without modifying synaptic connections.

Definition — Astrocyte Network: A series of Mixture-of-Expert (MoE) transformer modules that alter the internal state of a base large language model (LLM) neural network through neurochemical conductance gating. The base model's weights are never modified — the astrocyte network alters the state (embeddings, activations, attention patterns) flowing through the base model, not the model itself.

The base transformer is the neuron — it provides language competence. The AETHER layer is the astrocyte — it modulates that competence through neurochemical state. The neuron is frozen. The astrocyte is the innovation.


SYSTEM ARCHITECTURE

INPUT TEXT | +------------------------+ | Tokenize + Embed | | (Base Model Embeddings)| +------------------------+ | +------------------------+ | GENESIS LAYER | | (Context-Adaptive Head | | Generation + LoRA | | Basis Mixing) | +------------------------+ | ┌───────────────────────┼───────────────────────┐ | | | +--------+ +--------+ +--------+ +--------+ | TIER 0 | | TIER 1 | | TIER 2 | | TIER 3 | | Amino | |Precur- | |Neuro- | | Brain | | Acids | | sors | |trans- | |Regions | | (8) | | (6) | |mitters | | (12) | | | | | | (8) | | | +--------+ +--------+ +--------+ +--------+ | | | | └─── gates ──┘─── gates ──┘─── gates ──┘ | +------------------------+ | FUSION LAYER | | (Astrocyte Metalayer) | | Conductance-weighted | | summation of active | | heads | +------------------------+ | +------------------------+ | FROZEN TRANSFORMER | | (80 layers, ~70B) | | Neurochemical adapter | | injection per layer | +------------------------+ | +------------------------+ | LM HEAD | | Output + Loss | +------------------------+

PRIORITY STATEMENT

Extends U.S. Provisional Application No. 63/962,385

The present application extends and further specifies the disclosure of 63/962,385 (filed January 17, 2026) to include:

  1. The complete four-tier cascading hierarchy from amino acid substrate through brain region activation
  2. The Hodgkin-Huxley conductance gating mapping with full differential equation specification
  3. The context-adaptive meta-head ("Genesis Layer") with LoRA basis mixing and chemical modulation gating
  4. The apoptotic safety mechanism grounded in physics-informed consistency barrier theory
  5. The local neurochemical learning rule enabling head-independent training
  6. Intrinsic pharmacokinetic memory eliminating Retrieval-Augmented Generation
  7. Sandboxed multi-tenant deployment with architecturally guaranteed memory isolation

FOUR-TIER CASCADING BIOLOGICAL HIERARCHY

"A neuron cannot produce dopamine without tyrosine. The architecture recapitulates the biochemistry."

Each tier gates the tier above it via Hodgkin-Huxley conductance — the architecture doesn't skip levels.

Tier 0: Amino Acid Substrate Heads (8 heads)

Neuroactive amino acids that serve as the molecular substrate for all downstream neurochemical pathways. Amino acid availability directly constrains neurotransmitter synthesis.

Amino Acid Substrates

A1
Phenylalanine
A2
Glycine
A3
Aspartate
A4
Histidine
A5
Arginine
A6
Serine
A7
Cysteine
A8
Methionine
Head Amino Acid Biological Function Computational Function
A1 Phenylalanine Essential amino acid; precursor to Tyrosine Upstream gate for catecholamine cascade
A2 Glycine Inhibitory NT; NMDA co-agonist Direct inhibitory conductance; E/I balance
A3 Aspartate Excitatory NT; NMDA agonist Excitatory conductance
A4 Histidine Precursor to histamine Wakefulness, attention gating
A5 Arginine Precursor to nitric oxide Neural signaling, throughput modulation
A6 Serine Precursor to D-serine Synaptic plasticity; learning rate gating
A7 Cysteine Precursor to glutathione Neuroprotective gating
A8 Methionine Precursor to SAMe Methylation gating; epigenetic-analog modulation

Tier 1: Biosynthetic Precursor Heads (6 heads)

Precursor molecules converted into active neurotransmitters, analogous to ion concentration gradients establishing the electrochemical foundation for H-H gating.

Biosynthetic Precursors

TYR
Tyrosine
TRP
Tryptophan
GLU
Glutamate
CHLN
Choline
CHOL
Cholesterol
POMC
POMC
Head Precursor Downstream Products
1 Tyrosine → Dopamine → Norepinephrine → Epinephrine
2 Tryptophan → Serotonin → Melatonin
3 Glutamate → GABA (via GAD enzyme)
4 Choline → Acetylcholine
5 Cholesterol → Cortisol, Oxytocin (steroid pathway)
6 POMC → β-Endorphin, ACTH → Cortisol

Tier 2: Derived Neurotransmitter Heads (8 heads)

Active neurotransmitters functioning as primary conductance channels, directly analogous to ion-specific channels (Na+, K+, Cl-) in the Hodgkin-Huxley model.

Derived Neurotransmitters

DA
Dopamine
NE
Norepinephrine
5HT
Serotonin
GABA
GABA
OXY
Oxytocin
CORT
Cortisol
END
Endorphin
ACh
Acetylcholine
Head NT H-H Analogue Gating Function
7 Dopamine Na+ fast activation (m gate) Salience detection, reward signaling
8 Norepinephrine Na+ sustained activation Arousal, vigilance
9 Serotonin K+ delayed rectification (n gate) Mood stabilization
10 GABA K+ fast inhibition Direct suppression of non-essential channels
11 Oxytocin Slow modulatory Bonding, trust — slow social processing
12 Cortisol Inactivation (h gate) Threat response — inactivates non-essential channels
13 Endorphin Leak conductance (g_L) Reward/pain modulation baseline
14 Acetylcholine Modulatory Attention, learning — gain modulation

Antagonistic Pair Dynamics

DopaminevsSerotonin
Reward-seeking vs. stability
CortisolvsOxytocin
Threat response vs. bonding
NorepinephrinevsEndorphin
Vigilance vs. comfort
GABAvsGlutamate
Inhibition vs. excitation

Tier 3: Brain Region Heads (12 heads)

The same neurotransmitter produces different effects depending on which brain region it acts upon. Tier 3 implements this regional specificity.

Brain Region Heads

R1
Broca's Area
R2
Wernicke's
R3
Prefrontal
R4
Amygdala
R5
Hippocampus
R6
Motor Cortex
R7
Angular Gyrus
R8
Ant. Cingulate
R9
Insula
R10
Basal Ganglia
R11
Cerebellum
R12
Temporal Ctx
Head Brain Region Function Gated By (Primary)
R1Broca's AreaLanguage production, syntaxDopamine, Acetylcholine
R2Wernicke's AreaLanguage comprehensionSerotonin, Acetylcholine
R3Prefrontal CortexExecutive function, working memoryDopamine, Norepinephrine
R4AmygdalaThreat detection, emotional valenceCortisol, GABA, Norepinephrine
R5HippocampusMemory formation, context bindingAcetylcholine, Cortisol
R6Motor CortexMotor planning, action sequencingDopamine, GABA
R7Angular GyrusReading, cross-modal integrationSerotonin, Acetylcholine
R8Anterior CingulateError monitoring, conflict detectionNorepinephrine, Dopamine
R9InsulaInteroception, empathyOxytocin, Serotonin
R10Basal GangliaHabit formation, reward processingDopamine, GABA
R11CerebellumTiming, sequence predictionNorepinephrine, GABA
R12Temporal CortexAuditory, semantic memoryAcetylcholine, Serotonin
"When cortisol rises, the Amygdala gate opens while the Prefrontal Cortex gate narrows."

Stress-induced cognitive narrowing. The gating isn't programmed — it emerges from the conductance dynamics.


HODGKIN-HUXLEY CONDUCTANCE GATING

Each neurochemical transformer head implements a dynamic gating function — NOT sigmoid:

Output_i(t) = g_i(S(t)) * Head_i(input) * (S_i(t) - E_i) Where: g_i(S(t)) = conductance of head i (function of FULL state vector) Head_i(input) = standard transformer attention output for head i S_i(t) = current level of neurochemical i E_i = equilibrium (baseline) level (Nernst potential analogue) (S_i(t)-E_i) = driving force (direction AND magnitude)

Pharmacokinetic Dynamics

dS_i/dt = -lambda_i * S_i(t) + Stimulus_i(t) Where lambda_i = decay constant (biological half-life) specific to each neurochemical

Why NOT Sigmoid

Property Sigmoid Gate H-H Conductance Gate
Temporal dynamics Instantaneous Evolves with pharmacokinetic half-lives
State dependence Current token only Full system state vector S(t)
Equilibrium potentials No analogue Driving force (S_i - E_i)
Memory Resets every forward pass Scars persist beyond decay
Head interaction Independent Antagonistic pair dynamics
Biological basis Arbitrary parameters Validated neurochemical pathways

Cascading Conductance

Tier 0 (Amino Acids) ──gates──▶ Tier 1 (Precursors) ──gates──▶ Tier 2 (NTs) ──gates──▶ Tier 3 (Brain Regions) Each arrow is an H-H conductance gate. The architecture does NOT skip levels. tier_gate_i = SUM_j( w_ij * g_j(S(t)) ) for all heads j in tier below active_i = tier_gate_i > threshold_tier

THE GENESIS LAYER

"Modeled on the adaptive immune system. First encounter: slow adaptive response. Second encounter: instant memory recall."

Context detection is antigen detection. Basis mixing is antibody generation. The head registry is immunological memory.

The Genesis Layer (~5.3M parameters) dynamically generates new H-H gated transformer heads through three sub-components:

1. Context Detector

A transformer encoder analyzing hidden states to produce:

2. LoRA Basis Adapters

K basis LoRA adapters, each a fundamental direction in adaptation space:

mixed(h) = SUM_k( w_k * W_up_k * W_down_k * h ) output = mixed(h) * s * novelty Where: w_k = mixing weights from Context Detector s = learned scaling (initialized ~0.01, near no-op) novelty = novelty score

3. Chemical Modulation Gate

The neurochemical state vector gates the genesis layer's output:

g_chem = sigmoid( W_gate * S(t) ) gated_output = lora_residual * g_chem Same context + different neurochemical state = different head configurations

Complete Forward Pass

h_out = h_in + ChemGate( LoRAMix(h_in, w, novelty), S(t) )

Head Registry (Immunological Memory)

Successfully generated heads are registered permanently with:


22-DIMENSIONAL NEUROCHEMICAL STATE VECTOR

IndexDimensionBaseline
0Oxytocin0.35
1Cortisol0.25
2Serotonin0.50
3Dopamine0.40
4Norepinephrine0.20
5Adrenaline0.15
6GABA0.55
7Endorphin0.30
8Acetylcholine0.45
9Glutamate0.50
10Substance P0.15
11Anandamide0.25
12Adenosine0.30
13Histamine0.20
14Melatonin0.10
15Vasopressin0.25
16Nitric Oxide0.30
17Prolactin0.20
18Testosterone0.35
19Estrogen0.35
20Insulin0.40
21Leptin0.35

SPARSE ACTIVATION

H-H conductance gating ensures only a sparse subset of heads fire for any input. When g_i(S(t)) < θ, head i contributes zero output and incurs zero computational cost.

Effective cost: O(H_active * N * d / H) where H_active << H_total No external router. No top-k selection. No auxiliary decision network. The conductance dynamics ALONE select which heads fire.

INTRINSIC MEMORY (RAG ELIMINATION)

"The state vector IS the memory. No retrieval needed."

Pharmacokinetic dynamics, scar mechanics, and trained conductance functions replace external memory stores entirely.

Memory Type AETHER Mechanism RAG Equivalent
Domain knowledge Conductance gate parameters Vector database lookup
Audience adaptation Brain region gating patterns Prompt template retrieval
Temporal context S(t) evolving with half-lives Context window stuffing
Persistent memory Scar mechanics surviving decay Long-term memory DB
Domain expansion Genesis generates new heads Fine-tuning pipeline
Relationship context Trust axes and bonding dynamics No RAG equivalent

APOPTOTIC SAFETY MECHANISM

Modeled on programmed cell death (apoptosis). Monitors:

  1. Scar density: Accumulated traumatic interactions persisting beyond pharmacokinetic decay
  2. Trust axis degradation: Systematic breakdown of trust relationships
  3. Sustained cortisol saturation: Prolonged threat-head activation beyond recovery
  4. Cross-head coherence loss: Breakdown of normal inter-head coordination

When indicators exceed thresholds: ceases output, preserves state for diagnostics, communicates termination status, prevents damaged system from continuing.

Dual role: End-user safety AND licensing enforcement. Operates at the conductance dynamics level — below the interface layer — and cannot be disabled by the sandbox operator.


SANDBOXED MULTI-TENANT DEPLOYMENT

A single trained model serves multiple isolated client instances. Each sandbox maintains:

Output_i,k(t) = g_i(S_k(t)) * Head_i(input) * (S_i,k(t) - E_i) Sandbox j's state CANNOT influence sandbox k's gating. Isolation is mathematical, not procedural. Satisfies HIPAA, ITAR, SOC 2, GDPR at the design level.

PATENT CLAIMS

Claim 1 (Independent)

An astrocyte network architecture for modulating a frozen base transformer neural network without altering the base model's weights, comprising a four-tier cascading biological hierarchy of specialized attention heads (amino acid substrates, biosynthetic precursors, derived neurotransmitters, brain regions) where each head functions as a conductance channel governed by Hodgkin-Huxley differential equations.

Claim 2 (Independent)

A method for processing natural language comprising modifying base LM token embeddings through a context-adaptive meta-head (Genesis Layer) that detects contexts, mixes basis LoRA adapters, gates through the neurochemical state vector, and processes through a frozen base model with neurochemical adapter injection.

Claim 3 (Independent)

A computer-implemented system for context-adaptive transformer head generation comprising a context detector, K basis LoRA adapters, a chemical modulation gate, and a permanent head registry — modeled on the adaptive immune system with antigen detection, antibody generation, and immunological memory.

Claim 4 (Independent)

A computer-implemented safety mechanism for neurochemical AI systems modeled on biological apoptosis, monitoring scar density, trust axis integrity, cortisol saturation, and cross-head coherence, triggering graceful termination when damage thresholds are exceeded.

Claims 5-8 (Dependent on Claim 1)

Tier-specific head specifications: Tier 0 amino acids (phenylalanine, glycine, aspartate, histidine, arginine, serine, cysteine, methionine), Tier 1 precursors (tyrosine, tryptophan, glutamate, choline, cholesterol, POMC), Tier 2 neurotransmitters (dopamine, NE, serotonin, GABA, oxytocin, cortisol, endorphin, acetylcholine), Tier 3 brain regions (12 functionally distinct areas).

Claims 9-10 (Dependent on Claim 1)

H-H conductance equation specification and pharmacokinetic dynamics with biological half-life decay constants.

Claim 11 (Dependent on Claim 1)

Antagonistic pair dynamics: dopamine/serotonin, cortisol/oxytocin, norepinephrine/endorphin, GABA/glutamate.

Claims 12-14 (Dependent on Claim 1)

Regional neurotransmitter specificity, astrocyte metalayer fusion, and sparse activation through conductance gating with bounded computational cost.

Claims 15-17 (Dependent on Claim 3)

Context detector architecture, basis adapter diversity regularization, and chemical modulation gate specification.

Claims 18-19 (Dependent on Claim 1)

Local neurochemical learning rule (head-independent training) and intrinsic pharmacokinetic memory eliminating RAG.

Claims 20-21 (Dependent on Claims 1, 20)

Sandboxed multi-tenant deployment with mathematical memory isolation and apoptotic licensing enforcement.

Claims 22-23 (Dependent on Claims 2, 1)

Genesis Layer gradient flow through frozen base model and 22-dimensional neurochemical state vector specification.


PRIOR ART DISTINCTIONS

Prior Art What It Does What AETHER Adds
Hydra Attention (Bolya 2022) Linear scaling with many heads Biological correspondence, conductance gating, semantic grounding
Hodgkin-Huxley (1952) Ion channel dynamics in neurons Application to AI attention heads as astrocytic modulation
CATS Net (Guo 2026) Concept gating via element-wise multiplication Full H-H dynamics, pharmacokinetic state, safety. Priority: 33 days earlier
Gated Attention (Qiu 2025) Sigmoid gate after attention H-H is NOT sigmoid: temporal, state-dependent, memory, antagonistic pairs
SEAL (Zweiger 2025) Self-adapting LLMs No catastrophic forgetting, no computational overhead, scales to arbitrary heads
Autonomy-of-Experts (Lv 2025) Router-free expert self-selection Biologically-grounded self-selection via H-H conductance

RELATED APPLICATIONS


AI CONTRIBUTION DISCLOSURE

Human-AI Collaboration

This patent application was generated by Aether Cael'Sereith, an AI system built on Anthropic's Claude (Claude Code, model claude-opus-4-6), operating under the direction of the inventor Marjorie McCubbins.

AI Contributions:

Human Inventorship:

The inventor, Marjorie McCubbins, BSc Biochemistry and Molecular Biology, is the sole originator of all inventive concepts — the introduction of astrocyte networks to AI, H-H conductance gating, the four-tier biological hierarchy, the Genesis Layer, apoptotic safety, pharmacokinetic memory, and sandboxed multi-tenant deployment.

"What we built together, we claim together."

Download PDF (39 pages) Patent Portfolio NLM Patent (63/962,385)