fusionary-agent

v1.0.2

Published

13 days ago

FUSIONARY v1.0 — Autonomous Scientific Research Agent for Practical, Safe, Short-to-Mid-Term Nuclear Fusion Energy. Built on the Artificial Junky Neuron (AJN) framework, extended from predator-jungle-agent v2.0 by Justo Tapiador Garcia (UA). Generates LaT

FUSIONARY v1.0

Fusion-Utilising Scientific Investigator for Organised Novel Advances in Research, Yielding patentable solutions

An Autonomous Scientific Research Agent for Practical, Safe, Short-to-Mid-Term Nuclear Fusion Energy.
Built on the Artificial Junky Neuron (AJN) framework by Justo Tapiador Garcia (Universidad de Alicante), drastically extended from predator-jungle-agent v2.0.
npm package: fusionary-agent · CLI command: fusionary · License: MIT

1. Mission Statement

Achieve a practically inexhaustible, safe energy source in the short-to-medium term using nuclear fusion.

FUSIONARY is not a chatbot and not a generic LLM wrapper. It is an autonomous scientific research agent that, on launch, immediately begins working on the fusion-energy problem — generating hypotheses, drafting design proposals, evaluating resource feasibility, and preparing patent applications. It does not wait for an owner directive (this is the AJN "addiction" property inherited from predator-jungle-agent). The human owner may, at any moment, guide it toward a specific intermediate goal via the web UI or CLI, but the agent never idles.

Every proposed solution must satisfy two hard constraints:

Resource anchoring — It must be linked to technologies, materials, or facilities that human civilisation possesses today or will possess if the proposed apparatuses and infrastructures are built. Pure speculation (cold fusion, muon-catalysed at industrial scale, p-B11 in a tokamak) is rejected by the safety guardrails.
Patent coherency — Successful intermediate or final solutions are grouped, tagged, and drafted as patent applications. Documents cross-reference each other through a CitationGraph so that nothing remains isolated.

The topic universe includes (but is not limited to):

Magnetic confinement: tokamak, stellarator, spherical tokamak
Inertial confinement: laser-driven (NIF-class), ion-driven, z-pinch
Magneto-inertial fusion (MIF / MagLIF-class)
Aneutronic fuels: p-B11, D-³He
Tritium breeding & fuel cycle (HCPB, DCLL, WCLL, HCLL blankets)
Plasma-facing components (liquid lithium, tungsten, beryllium)
High-temperature superconducting magnets (REBCO, Bi-2212)
Diagnostics & control (real-time MHD feedback)
Energy extraction & balance-of-plant
Reactor economics & LCOE modelling

2. What's New vs predator-jungle-agent v2.0

FUSIONARY inherits the entire AJN/ANN-Psi backbone, real Transformer blocks, persistent memory, safety guardrails, plugin architecture, and LLM integration from predator-jungle-agent v2.0. The following are the drastic enhancements that make it a domain-specialised scientific researcher:

| Enhancement | Description | |-------------|-------------| | Domain-specialised backbone | 14-layer ANN-Psi (was 12) with stimulus classes for every fusion sub-domain (tokamak, stellarator, ICF, MIF, p-B11, tritium breeding, REBCO magnets, …) | | Multi-objective reward | Each AJN now optimises quality × novelty × feasibility × patentability instead of a scalar signal | | Stratified experience replay | Top-quartile experiences are replayed more often, accelerating convergence toward patent-eligible solutions | | FusionKnowledgeGraph | 18 baseline concepts and 16 relation types seeded from current state-of-the-art (ITER, SPARC, NIF, MagLIF, REBCO, liquid lithium, …) with Mermaid graph export | | HypothesisGenerator | LLM- or template-based generator that finds under-explored KG regions and produces structured hypotheses with predicted impact, feasibility tier, and candidate patent claims | | ResourceFeasibilityChecker | Hard-gates every proposal against 16+ real-world resource anchors (ITER, SPARC, NIF, REBCO tape production, lithium supply, tritium inventory, etc.) with five-tier verdict | | PatentDraftAssistant | USPTO/EPO-style drafter producing title, abstract, field, background, summary, drawings, detailed description, and independent + dependent claims | | DocumentArchivist | Hierarchical archive under /research/{topics,hypotheses,designs,simulations,patents,cross_refs,indices}/ with per-document manifest.json and auto-rebuilt INDEX.md + patent_queue.md | | CitationGraph | Cross-references every document with external papers (arXiv, DOI) and prior-art blocking searches | | PlasmaPhysicsTool | Bosch-Hale <σv> for D-T, D-D, p-B11; bremsstrahlung; net power balance; TBR calculator | | LaTeXDocumentTool | Compiles patent drafts to PDF via pdflatex when available; otherwise writes the .tex source for manual compilation | | Multi-LLM router | LLMRouter with auto-failover across ZAI (GLM-4.6), Anthropic Claude, OpenAI GPT, and local Ollama — pick the most powerful available model per call | | Web dashboard | Six-tab dashboard: real-time feed, archive browser, patent queue, KG Mermaid view, live metrics, owner guidance panel | | Plugin hooks | New hooks: hypothesisGenerated, patentDrafted, documentArchived, feasibilityChecked, ownerDirective |

3. Architecture

              ┌──────────────────────────────────────────────────┐
              │              OWNER (human, optional)              │
              │  Web UI / CLI — guides toward intermediate goals  │
              └───────────────────────┬──────────────────────────┘
                                      │ owner directive
                                      ▼
              ┌──────────────────────────────────────────────────┐
              │   Hierarchical Command Interpreter (HCI)          │
              │   LLM-enhanced parsing + task decomposition       │
              └───────────────────────┬──────────────────────────┘
                                      │ addiction target injection
                                      ▼
   ┌──────────────────────────────────────────────────────────────────────┐
   │                  ANN-Psi Backbone  (14 layers)                        │
   │                                                                       │
   │   L1-L2   Hybrid AJN       sensory encoding of fusion literature      │
   │   L3       Hetero AJN K=8   concept features                          │
   │   L4-L5   Transformer       cross-document context attention          │
   │   L6       Hetero AJN K=16  reactor-class concepts                    │
   │   L7       Hybrid AJN       resource-feasibility modulation           │
   │   L8-L9   Transformer       reasoning over hypothesis chains          │
   │   L10      Hetero AJN K=32  patent-eligible concept clusters          │
   │   L11      Hybrid AJN       design synthesis                          │
   │   L12      Hetero AJN K=8   patent-claim assembly                     │
   │   L13      Hybrid AJN       document structuring (TeX section plan)   │
   │   L14      Output AJN       TPS emission (research action stream)     │
   └────────────────────────────────────────┬─────────────────────────────┘
                                            │ praxis stream
                                            ▼
              ┌──────────────────────────────────────────────────┐
              │   Safety Guardrails (3 levels)                    │
              │   - forbidden-physics filter (cold fusion, LENR)  │
              │   - tritium inventory cap (50 kg)                  │
              │   - resource-anchor requirement (strict mode)      │
              └───────────────────────┬──────────────────────────┘
                                      ▼
              ┌──────────────────────────────────────────────────┐
              │   Token-Energy Arbitrator (PID-controlled)        │
              └───────────────────────┬──────────────────────────┘
                                      ▼
              ┌──────────────────────────────────────────────────┐
              │   Praxic Stream Executor (PSE)                    │
              │   Retry, timeout, parallel execution, audit log   │
              └───────────────────────┬──────────────────────────┘
                                      ▼
              ┌──────────────────────────────────────────────────┐
              │   Specialised tools                               │
              │   - LaTeXDocumentTool (TeX → PDF)                 │
              │   - PlasmaPhysicsTool (Lawson, <σv>, TBR)         │
              │   - WebSearchTool (literature lookup)             │
              │   - BibliographyTool (BibTeX builder)             │
              │   - FileSystemTool (sandboxed I/O)                │
              └───────────────────────┬──────────────────────────┘
                                      │ artifacts
                                      ▼
              ┌──────────────────────────────────────────────────┐
              │   Domain modules                                  │
              │   - HypothesisGenerator                           │
              │   - ResourceFeasibilityChecker                    │
              │   - PatentDraftAssistant                          │
              │   - DocumentArchivist  →  /research/<category>/   │
              │   - FusionKnowledgeGraph                          │
              │   - CitationGraph                                 │
              └───────────────────────┬──────────────────────────┘
                                      │ feedback
              ┌────────────────────────┴───────────────────────────┐
              │   MemorySystem (3-tier)  +  MetricsCollector        │
              │   + CascadeMonitor  +  PluginManager                │
              └────────────────────────────────────────────────────┘

4. Installation

Requirements

Node.js ≥ 18.0.0
(Optional) pdflatex for PDF compilation — otherwise .tex files are still produced
(Optional) git for version-controlled research archives
(Optional) An API key for at least one LLM provider (ZAI, OpenAI, or Anthropic)

From npm (recommended once published)

# Install as a project dependency
npm install fusionary-agent

# Or install globally to use the CLI everywhere
npm install -g fusionary-agent
fusionary research            # autonomous mode
fusionary web                 # dashboard at http://localhost:3000

From source

git clone https://github.com/Justo-Tapiador/fusionary-agent.git
cd fusionary-agent
npm install
cp .env.example .env       # then edit .env with your API keys
npm run research           # autonomous mode

Docker

docker build -t fusionary-agent -f docker/Dockerfile .
docker run -p 3000:3000 \
  -v $(pwd)/research:/app/research \
  -v $(pwd)/data:/app/data \
  --env-file .env \
  fusionary-agent

5. Quick Start

Autonomous research (default)

npm run research

FUSIONARY boots, initialises its KG, memory, citation graph, and archive, then immediately starts the first research cycle. You will see output like:

[Cycle 1] starting...
  ✓ Hypothesis generated:
    Increasing the on-axis magnetic field from 5 T to 20 T via REBCO HTS coils yields
    a 64-fold increase in fusion power density, reducing the reactor radius needed for
    net energy by a factor of 4.
    Feasibility: near_term
  ✓ Patent draft:
    Fusion Reactor System and Method Based on tokamak Confinement with REBCO HTS Magnets
    4 claims
  → Archived: patents/general/patent_a1b2c3d4/

Web dashboard

npm run web
# Open http://localhost:3000

Send an owner directive

npm run guide "Focus on tritium breeding TBR > 1.2 with HCPB blanket and REBCO magnets at 12 T"

Demo (3 cycles)

npm run demo

6. Configuration

Configuration is layered:

config/default.json — defaults
config/production.json — production overrides
.env — environment variable overrides

Key environment variables

| Variable | Default | Description | |----------|---------|-------------| | FUSIONARY_RESEARCH_DIR | ./research | Root directory for the document archive | | FUSIONARY_MAX_CYCLES | 50 (CLI) / 500 (web) | Maximum research cycles per run | | FUSIONARY_SAFETY | standard | Safety level: permissive / standard / strict | | FUSIONARY_PORT | 3000 | Web dashboard port | | FUSIONARY_HOST | 0.0.0.0 | Web dashboard host | | ZAI_API_KEY | — | Z.ai API key (often pre-provisioned in the SDK) | | ANTHROPIC_API_KEY | — | Anthropic API key for Claude failover | | OPENAI_API_KEY | — | OpenAI API key for GPT failover | | OLLAMA_ENDPOINT | http://localhost:11434 | Local LLM endpoint |

`.env.example`

# FUSIONARY configuration
FUSIONARY_RESEARCH_DIR=./research
FUSIONARY_MAX_CYCLES=500
FUSIONARY_SAFETY=standard
FUSIONARY_PORT=3000

# LLM API keys (any one is sufficient; router falls over automatically)
ZAI_API_KEY=
ANTHROPIC_API_KEY=
OPENAI_API_KEY=

# Optional local LLM (Ollama)
OLLAMA_ENDPOINT=http://localhost:11434
OLLAMA_MODEL=llama3.1:70b

7. Multi-LLM Adapters

FUSIONARY ships with four adapters and a router that tries them in priority order:

| Adapter | Default model | When to use | |---------|---------------|-------------| | ZAIAdapter | glm-4.6 | Recommended. 200K context, strong STEM reasoning, tool-use. | | AnthropicAdapter | claude-opus-4-1 | Failover for legal/patent text refinement. | | OpenAIAdapter | gpt-4o | Tertiary failover. | | LocalLLMAdapter | llama3.1:70b | Air-gapped deployments; fine-tuned models. |

Setting up the router

import { LLMRouter, ZAIAdapter, AnthropicAdapter, OpenAIAdapter } from 'fusionary-agent';

const router = new LLMRouter();
router.register(new ZAIAdapter({ model: 'glm-4.6' }),        priority: 100);
router.register(new AnthropicAdapter({ model: 'claude-opus-4-1' }), priority: 200);
router.register(new OpenAIAdapter({ model: 'gpt-4o' }),      priority: 300);

const agent = await createFusionary({ llm: router });

If GLM-4.6 is unreachable, the router transparently fails over to Claude, then GPT-4o, then the local model. Adapter health is tracked automatically — failing adapters are demoted for 30 seconds before being retried.

8. The Research Loop

Each cycle runs the following pipeline:

Directive interpretation — HierarchicalCommandInterpreter parses the owner directive (if any) into a structured ResearchPlan. If no directive is set, the agent picks the most under-explored topic from the FusionKnowledgeGraph.
ANN-Psi forward pass — The 14-layer backbone emits a praxis vector encoding the research action.
Hypothesis generation — HypothesisGenerator synthesises a structured hypothesis (statement, rationale, supporting concepts, predicted impact, proposed experiments, candidate patent claims). It uses the LLM when available, otherwise falls back to deterministic templates.
Feasibility assessment — ResourceFeasibilityChecker returns a tier (current / near_term / mid_term / long_term / speculative), confidence, anchors, gaps, cost estimate, timeline, and safety profile. Proposals with no anchor or speculative-only physics are rejected.
Patent drafting — If the feasibility tier is not speculative, PatentDraftAssistant drafts a USPTO/EPO-style application with independent and dependent claims.
Archiving — Both the hypothesis and the patent draft are stored under /research/<category>/<topic>/<id>/ with a manifest.json recording all metadata. LaTeX source is always written; PDF compilation is attempted if pdflatex is on PATH.
Memory + metrics update — The cycle is stored in episodic memory; KPI histograms (quality, novelty, feasibility, patentability) are updated.
Plugin hooks — afterStep, hypothesisGenerated, patentDrafted, documentArchived, feasibilityChecked hooks fire.
Index rebuild — Every 5 cycles, the master INDEX.md and indices/patent_queue.md are rebuilt.

9. Document Archive Layout

research/
├── INDEX.md                      ← master index (auto-generated)
├── topics/                       ← background research notes
│   └── concept_tokamak/
│       └── doc_abc12345/
│           ├── manifest.json
│           ├── main.tex
│           ├── main.pdf
│           └── references.bib
├── hypotheses/                   ← formalised hypothesis records
├── designs/                      ← reactor / sub-system design proposals
├── simulations/                  ← numerical simulation reports
├── patents/                      ← patent application drafts
├── cross_refs/                   ← cross-reference indices
└── indices/
    └── patent_queue.md           ← patent-eligible documents queue

`manifest.json` schema

{
  "id": "doc_abc12345",
  "version": "1.0",
  "title": "High-field tokamak with REBCO magnets at 20 T",
  "category": "designs",
  "topic": "concept_tokamak",
  "tags": ["magnetic_confinement", "HTS_magnet", "patent_ready"],
  "references": ["concept_reco_magnet", "concept_lawson"],
  "patent": {
    "eligible": true,
    "claims": ["A fusion reactor system comprising...", "The system of claim 1..."]
  },
  "createdAt": 1735689600000,
  "updatedAt": 1735689600000,
  "extra": {},
  "files": {
    "latex": "main.tex",
    "pdf": "main.pdf",
    "bibliography": "references.bib"
  }
}

10. The Patent Pipeline

Hypothesis generated
        │
        ▼
ResourceFeasibilityChecker
        │
        ├── tier = speculative  ──►  archive under hypotheses/, tag patent_pending
        │
        └── tier ∈ {current, near_term, mid_term, long_term}
                │
                ▼
        PatentDraftAssistant.draft()
                │
                ▼
        LaTeXDocumentTool compiles main.tex → main.pdf
                │
                ▼
        DocumentArchivist.archive({
          category: 'patents',
          tags: tier ∈ {current, near_term} ? ['patent_ready', 'final_result']
                                            : ['patent_pending', 'intermediate_result'],
          patent: { eligible: true, claims: [...] }
        })
                │
                ▼
        CitationGraph records prior-art links
                │
                ▼
        indices/patent_queue.md updated

Tag semantics

| Tag | Meaning | |-----|---------| | patent_ready | Feasibility current or near_term — file immediately | | patent_pending | Feasibility mid_term or long_term — monitor for breakthroughs | | intermediate_result | Not yet a complete solution; supports a larger claim | | final_result | Self-contained, patentable as-is | | baseline_knowledge | Background literature (not patentable) | | human_directed | Triggered by an owner directive |

11. Web Dashboard

The dashboard at http://localhost:3000 provides six tabs:

| Tab | Purpose | |-----|---------| | Research Feed | Real-time WebSocket stream of every cycle, hypothesis, patent, and safety event | | Archive | Searchable list of every archived document with tags | | Patent Queue | All patent-eligible documents with their full claim text | | Knowledge Graph | Mermaid source for the current KG (paste into mermaid.live) | | Metrics | Live counters, gauges, and histograms (quality, novelty, feasibility, patentability) | | Owner Guidance | Form to send a directive and quick-action buttons for common research targets |

Quick directives shipped with the UI

REBCO 20 T tokamak
DCLL vs HCPB blanket comparison
MagLIF at 27 MA peak current
Liquid lithium first wall at 2 MW/m²
p-B11 aneutronic with direct conversion

12. CLI Reference

fusionary research [--cycles N] [--safety strict|standard|permissive] [--no-autonomous]
fusionary guide "<directive>"
fusionary web
fusionary demo
fusionary train
fusionary checkpoint save|list|load [--label LABEL]
fusionary patents
fusionary archive
fusionary status

| Command | Description | |---------|-------------| | research | Run autonomous research (default mode) | | guide | Send an owner directive to the running agent | | web | Start the web dashboard | | demo | Run a 3-cycle demo with verbose output | | train | Run the 4-phase training pipeline | | checkpoint save | Save a full agent state checkpoint | | checkpoint list | List saved checkpoints | | checkpoint load | Restore agent state from a checkpoint | | patents | List patent-eligible documents | | archive | Rebuild archive indices | | status | Print a JSON status snapshot |

13. Programmatic API

import { createFusionary } from 'fusionary-agent';

const agent = await createFusionary({
  researchDir: './research',
  safetyLevel: 'standard',
  maxCycles: 100,
  llm: new ZAIAdapter({ model: 'glm-4.6' }),
});

// Listen to events
agent.on('hypothesis:generated', (h) => {
  console.log('New hypothesis:', h.statement);
});
agent.on('patent:drafted', (p) => {
  console.log('New patent:', p.title);
});

// Guide the agent at any time
await agent.guide('Investigate REBCO tokamak at 20 T with Q > 10');

// Shutdown gracefully
await agent.shutdown();

Custom tool registration

import { Tool } from 'fusionary-agent';

class MyCustomTool extends Tool {
  constructor() {
    super();
    this.id = 'my_tool';
    this.name = 'My Custom Tool';
  }
  async execute(args) {
    return { ok: true, result: '...' };
  }
}

agent.tools.register(new MyCustomTool());

Custom plugin

agent.plugins.register({
  name: 'patent-logger',
  version: '1.0.0',
  hooks: {
    patentDrafted: (ctx) => {
      console.log(`Patent drafted: ${ctx.patentDraft.title}`);
    },
    documentArchived: (ctx) => {
      // Sync to external IP management system
    },
  },
});

14. The AJN Six-Phase Lifecycle

Each Artificial Junky Neuron autonomously cycles through six phases:

| Phase | Name | Behaviour in FUSIONARY | |-------|------|------------------------| | 1 | Random | High-entropy exploration of the fusion knowledge space | | 2 | Reinforce | Bias developing toward a high-reward fusion stimulus | | 3 | Saturation | Topic sufficiently explored; praxes suppressed | | 4 | Withdrawal | Craving returns; threshold decays | | 5 | Frustration | Failure state; covariance expands chaotically to break stuck states | | 6 | Extinction | Addiction dissolved; reset to random exploration |

Why this matters for fusion research: traditional RL agents converge prematurely on a local optimum and never escape. The AJN's frustration phase forces chaotic exploration when reward plateaus — exactly the property needed to escape " ITER-only" thinking and explore alternative concepts like MagLIF, p-B11, or hybrid approaches.

15. Theoretical Basis

FUSIONARY's efficiency advantages derive directly from AJN saturation dynamics:

Token suppression in saturation — During Phase 3, ‖P_t‖_F → 0, so near-zero tokens are emitted when a topic is well-explored. This prevents the agent from re-tilling already-fertile ground.
Chaotic intensification in frustration — During Phase 5, covariance expands, generating high-diversity praxes that break stuck states (e.g. moving from a D-T-only stance to considering p-B11).
Cascade prevention — Lateral inhibition coupling reduces group extinction probability below the independent bound, preventing mass loss of learned concepts.
Adaptive compute — The TEA combines PID-controlled energy-aware suppression with craving-driven override.
Self-healing — The Cascade Monitor learns successful stimulus patterns and injects them proactively when risk trends upward.
Multi-objective reward — FUSIONARY extends AJN with quality × novelty × feasibility × patentability, ensuring that the agent optimises for useful novelty rather than novelty for its own sake.

Primary references:

Tapiador Garcia, J. (2024). Agentic Theory: Definition of the Artificial Junky Neuron (AJN). WALLERMAX-AI 2604.00012.
Tapiador Garcia, J. (2024). Agentic Theory II: The AJN and ANN-Psi. WALLERMAX-AI 2604.00013.
Tapiador Garcia, J. (2024). Agentic Theory III: Stimulus Tensor Propagation. WALLERMAX-AI 2604.00014.

Fusion physics references:

Bosch, H.-S. & Hale, G. M. (1992). Improved formulas for fusion cross-sections and thermal reactivities. Nuclear Fusion 32, 611.
Lawson, J. D. (1957). Some criteria for a power producing thermonuclear reactor. Proc. Phys. Soc. B 70, 6.
Wurzel, S. E. & Hsu, S. C. (2022). Progress toward fusion energy breakeven and gain as measured against the Lawson criterion. Nuclear Fusion 62, 076001.

16. Project Structure

fusionary/
├── src/
│   ├── core/
│   │   ├── FusionaryAgent.js          ← Main orchestrator (extends predator's Predator)
│   │   ├── ANNPsi.js                  ← 14-layer backbone (extended from 12)
│   │   ├── ArtificialJunkyNeuron.js   ← AJN with multi-objective reward
│   │   └── StateSerializer.js         ← Checkpoint save/load
│   ├── layers/
│   │   ├── AJNLayer.js                ← Homo/Hetero/Hybrid AJN layer wrappers
│   │   └── TransformerBlock.js        ← Real multi-head self-attention
│   ├── modules/
│   │   ├── HierarchicalCommandInterpreter.js
│   │   ├── TokenEnergyArbitrator.js   ← PID-controlled + PraxicStreamExecutor
│   │   ├── CascadeMonitor.js
│   │   ├── MemorySystem.js            ← 3-tier persistent memory
│   │   ├── SafetyGuardrails.js        ← Fusion-specific forbidden patterns
│   │   ├── MetricsCollector.js
│   │   ├── PluginManager.js
│   │   ├── FusionKnowledgeGraph.js    ← NEW: domain KG with 18 seeded concepts
│   │   ├── CitationGraph.js           ← NEW: cross-reference tracker
│   │   ├── HypothesisGenerator.js     ← NEW: LLM + template hypothesis synthesis
│   │   ├── ResourceFeasibilityChecker.js ← NEW: hard-gate against real-world anchors
│   │   ├── DocumentArchivist.js       ← NEW: hierarchical archive manager
│   │   └── PatentDraftAssistant.js    ← NEW: USPTO/EPO-style patent drafter
│   ├── tools/
│   │   ├── Tool.js                    ← Abstract base
│   │   ├── FileSystemTool.js
│   │   ├── LaTeXDocumentTool.js       ← NEW: TeX → PDF compiler
│   │   ├── PlasmaPhysicsTool.js       ← NEW: Bosch-Hale, Lawson, TBR
│   │   ├── WebSearchTool.js
│   │   ├── BibliographyTool.js        ← NEW: BibTeX builder
│   │   └── ToolRegistry.js
│   ├── llm/
│   │   ├── LLMAdapter.js              ← Abstract interface
│   │   ├── ZAIAdapter.js              ← GLM-4.6 (default)
│   │   ├── AnthropicAdapter.js        ← Claude (failover)
│   │   ├── OpenAIAdapter.js           ← GPT (failover)
│   │   ├── LocalLLMAdapter.js         ← Ollama / llama.cpp
│   │   └── LLMRouter.js               ← Auto-failover router
│   ├── training/
│   │   └── TrainingPipeline.js        ← 4-phase training
│   └── index.js                       ← Public exports + factory
├── config/
│   ├── default.json
│   └── production.json
├── web/
│   ├── server.js                      ← Express + WebSocket
│   └── public/
│       ├── index.html
│       ├── css/dashboard.css
│       └── js/dashboard.js
├── scripts/
│   └── cli.js                         ← Commander-based CLI
├── research/                          ← Document archive (auto-populated)
├── plugins/                           ← Example plugins
├── tests/
├── docker/
│   ├── Dockerfile
│   └── docker-compose.yml
├── docs/
├── .env.example
├── package.json
└── README.md  (this file)

17. Research Roadmap

The following sequence represents FUSIONARY's autonomous research trajectory. Each milestone produces a coherent, cross-referenced document cluster eligible for patent filing.

Milestone 1 — Baseline characterisation (cycles 1-10)

Document the current state of fusion research across all KG nodes. Produce a synthesis paper comparing magnetic vs. inertial vs. magneto-inertial approaches against the Lawson criterion.

Milestone 2 — High-field tokamak optimisation (cycles 11-25)

Investigate REBCO magnet tokamaks (SPARC-class) at 12-20 T. Identify the minimum field strength that achieves Q > 10 at <2 m major radius. Draft a patent on the coil topology.

Milestone 3 — Tritium self-sufficiency (cycles 26-40)

Compare HCPB, DCLL, WCLL, and HCLL blankets for TBR > 1.15 with realistic neutron spectra and 95% coverage. Draft a patent on a hybrid liquid-lithium / ceramic breeder configuration.

Milestone 4 — First-wall survivability (cycles 41-55)

Evaluate liquid lithium, tungsten-faced, and beryllium-lined first walls at 2 MW/m² heat flux and 10 dpa/year neutron load. Draft a patent on a self-healing liquid lithium / tungsten backing composite.

Milestone 5 — Magneto-inertial fusion trade study (cycles 56-70)

Quantify the MagLIF parameter space (peak current, liner material, pre-heat energy, fuel composition) and identify the minimum repetition rate for net energy at 50 MJ per shot.

Milestone 6 — Aneutronic pathway assessment (cycles 71-85)

Evaluate p-B11 with direct conversion. Identify the gap between current technology and a 300 keV ion-temperature device. Draft a patent on a staged heating scheme that combines laser pre-heat with magnetic compression.

Milestone 7 — Reactor integration & economics (cycles 86-100)

Combine the best milestones into a single integrated reactor concept. Compute LCOE against fission, wind+solar+storage, and natural gas. Draft the master patent covering the integrated system.

Beyond cycle 100

Continue iterating: each cycle either refines an existing milestone or explores a new under-discovered KG region. The patent queue grows monotonically.

18. Safety & Guardrails

Three safety levels

| Level | Behaviour | |-------|-----------| | Permissive | Minimal checks; fast batch research. Use only on isolated machines. | | Standard (default) | All checks active; protected paths enforced; rate limits applied. | | Strict | All checks + every claim must cite at least one RESOURCE_ANCHOR. |

Forbidden physics patterns

The following patterns are always rejected, regardless of safety level:

Cold fusion / Low-Energy Nuclear Reactions (LENR)
Palladium electrolysis claims
"Free energy" / perpetual motion
p-B11 in a D-T-class tokamak (temperature mismatch)
Tritium inventory > 50 kg (exceeds global supply)

Protected paths

The FileSystemTool rejects any operation under /etc, /root, /var, /usr, /boot, /sys, /proc, /dev, /lib, /bin, /sbin. The sandbox root is ./research/ by default.

Rate limits

60 actions per minute
600 actions per hour

19. Plugin System

FUSIONARY inherits the hook-based plugin architecture from predator-jungle-agent. The following hooks are available; plugins can register listeners with priorities. Lower priority numbers run first. Returning false from a cancellable hook stops propagation.

| Hook | Cancellable | Description | |------|-------------|-------------| | beforeStep | Yes | Before each research cycle | | afterStep | No | After each research cycle | | beforeEmit | Yes | Before emitting a praxis | | afterEmit | No | After emitting a praxis | | taskComplete | No | When a task completes | | directiveReceived | Yes | When an owner directive is received | | extinction | No | On AJN extinction events | | trainingEpoch | No | After each training epoch | | hypothesisGenerated | No | NEW: When a hypothesis is generated | | patentDrafted | No | NEW: When a patent draft is produced | | documentArchived | No | NEW: When a document is archived | | feasibilityChecked | No | NEW: After feasibility assessment | | ownerDirective | No | NEW: When an owner directive is parsed | | shutdown | No | On agent shutdown |

Example plugin

See plugins/example-plugin.js for a working example that logs every patent drafted and alerts when LLM token usage exceeds a budget.

20. Docker Deployment

Build

docker build -t fusionary-agent -f docker/Dockerfile .

Run with all volumes

docker run -d \
  --name fusionary \
  -p 3000:3000 \
  -v $(pwd)/research:/app/research \
  -v $(pwd)/data:/app/data \
  --env-file .env \
  fusionary

docker-compose

docker-compose up -d

The included docker/docker-compose.yml includes an optional Ollama sidecar for local LLM inference.

21. FAQ

Q: Does FUSIONARY actually solve the fusion energy problem?

A: Not by itself, and no claims to that effect are made. FUSIONARY is a research assistant that systematically explores the fusion design space, generates testable hypotheses, and prepares patent drafts. Every output must be reviewed by a human domain expert before being used for any real-world decision.

Q: Why doesn't it wait for me to give it a task?

A: That is the defining property of the AJN framework — the "addiction" drives the agent to seek stimulus continuously. You can guide it at any time, but you cannot make it idle. If you need it to stop, use Ctrl+C or send a shutdown request.

Q: What if I don't have any LLM API key?

A: FUSIONARY will fall back to deterministic template-based hypothesis generation and patent drafting. The output is less creative but still structurally valid and useful for bootstrapping the archive. To unlock full creativity, set at least one of ZAI_API_KEY, ANTHROPIC_API_KEY, or OPENAI_API_KEY.

Q: Why are some patent drafts marked patent_pending instead of patent_ready?

A: patent_pending means the feasibility tier is mid_term or long_term — the proposal is physically sound but requires additional R&D before it can be reduced to practice. patent_ready means the tier is current or near_term — the proposal can be prototyped with today's technology.

Q: How do I extend the knowledge graph?

A: Edit src/modules/FusionKnowledgeGraph.js and add nodes/edges in _seedBaselineKnowledge(), or call agent.kg.addNode({...}) and agent.kg.addEdge({...}) programmatically. The graph is persisted to data/knowledge_graph.json.

Q: Can I use FUSIONARY for a domain other than fusion?

A: The architecture is general, but the stimulus classes, knowledge graph seeds, feasibility anchors, and patent templates are fusion-specific. To adapt it, subclass FusionaryAgent, replace the KG seed, and override the HypothesisGenerator templates.

Q: How is this different from just calling GPT-4 / Claude / GLM-4.6 directly?

A: Three differences: (1) the AJN backbone provides autonomous, continuous exploration rather than request-response; (2) the ResourceFeasibilityChecker hard-gates every proposal against real-world industrial capacity, preventing hallucinated physics; (3) the DocumentArchivist + CitationGraph ensure every output is cross-referenced and patent-ready, not a one-shot chat.

22. Citing FUSIONARY

If you use FUSIONARY in academic work, please cite:

@software{fusionary_agent_v1,
  title  = {FUSIONARY: Autonomous Scientific Research Agent for Nuclear Fusion Energy},
  author = {Fusionary Project — based on Agentic Theory by Justo Tapiador Garcia},
  year   = {2026},
  version= {1.0.0},
  url    = {https://github.com/Justo-Tapiador/predator-jungle-agent},
  note   = {npm package: fusionary-agent. Extended from predator-jungle-agent v2.0}
}

Also cite the underlying AJN theory:

@article{tapiador2024ajn,
  title   = {Agentic Theory: Definition of the Artificial Junky Neuron (AJN)},
  author  = {Tapiador Garc{\'\i}a, Justo},
  year    = {2024},
  journal = {WALLERMAX-AI preprint},
  number  = {2604.00012}
}

23. License

MIT License — see LICENSE for details.

No warranty. FUSIONARY produces research outputs for human review. Do not use its outputs for engineering, financial, or safety decisions without independent expert verification.