Nikola AI Architecture

A novel AI architecture built on 9-Dimensional Toroidal Waveform Intelligence (9D-TWI) — physics-first cognition using interference patterns on a Riemannian manifold.

Nikola replaces discrete weight matrices with a continuous wavefunction Ψ evolving on a 9-dimensional toroidal manifold (T⁹) governed by the Unified Field Interference Equation (UFIE):

∂²Ψ/∂t² = c²∇²_g Ψ − α(1−r̂)∂Ψ/∂t + β|Ψ|²Ψ + Σ Eᵢ(x,t)

Memory, attention, and reasoning emerge from wave interference — not from lookup tables or static weights.

Key properties:

• Information encoded as complex wavefunction amplitudes across 9D toroidal topology
• 9D Hilbert space-filling curve for optimal memory locality (Skilling algorithm, 14,133 assertions passing)
• Störmer–Verlet Strang-split integrator for Hamiltonian energy conservation
• Neuroplastic Transformer (NPT) attention operating natively on the wavefunction
• Autonomous behavioral loop: dopamine, serotonin, ATP metabolism, boredom-driven exploration
• Post-quantum cryptography: ML-KEM/Kyber-768 + SPHINCS+-SHAKE-256f

Phase 110 complete — 112 tests, ~98% pass rate (2 pre-existing timing-flaky)

# Required
sudo apt install cmake g++ libcatch2-dev liblmdb-dev libonnxruntime-dev

# Optional (GPU features)
# CUDA 12.0+, NVIDIA RTX GPU (sm_86 confirmed; sm_75+ supported)

git clone https://github.com/alternative-intelligence-cp/nikola.git
cd nikola
mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
make -j$(nproc)

cd build
ctest --timeout 120 -j4          # full suite (112 tests)
ctest -R Phase109                 # single suite

# Single prompt
./nikola-run --prompt "What is consciousness?" --ticks 300 --emit-all

# Interactive REPL
./nikola-run --interactive

# JSON output with persistent memory
./nikola-run --json --memory /tmp/nikola_session.bin --prompt "Hello"

# Batch from file
./nikola-run --batch prompts.txt --quiet

Text Input
    │
    ▼
BERT Tokenizer (30,522 tokens)
    │
    ▼
BERT-tiny ONNX Inference (17.5 MB, ORT)
    │    (768-dim embeddings)
    ▼
HolographicInjector
    │    (8 emitters → 9D toroidal field)
    ▼
TorusGrid (T⁹, 19,683 nodes, SoA layout)
    │
    ├── CPU Propagator (Strang-Verlet, 6 substeps)
    └── GPU Propagator (CUDA, RTX 3090) [partial]
    │
    ▼
NeuralProcessingTransformer (NPT)
    │    (wave-correlation attention, 8 heads at π·φⁿ bands)
    │    [Transformer sits here — before Mamba, not after]
    ▼
CognitiveCore / Mamba-9D SSM
    │    (H=256 state space, 100-step wave window, WavefunctionSampler)
    ▼
SemanticMemory (wave-basis, Hilbert-indexed, persistent)
    │
    ▼
DecisionLoop + AutonomyEngine
    │    (dopamine, ATP, boredom, mania suppression)
    ▼
ThoughtComposer → EMIT_THOUGHT
    │
    ▼
nikola-run CLI Output

Nikola targets NVIDIA RTX 3090 (sm_86, CUDA 12.0). Current GPU features:

GpuHamiltonianOracle::compute() automatically dispatches to the GPU when NVIDIA hardware is detected and nikola_cuda is linked.

Nikola implements NIST-standardized post-quantum cryptography:

• ML-KEM/Kyber-768 (FIPS 203): Key encapsulation for secure session establishment
• SPHINCS+-SHAKE-256f: Stateless hash-based digital signatures

Both are implemented via third-party reference code in third_party/.

• Integration Specifications — full mathematical specification
• Research Audit 2026 — research vs. implementation coverage
• Research Notes
• Contributing Guidelines

Current (Phase 110)

• ✅ Real BERT tokenizer + ONNX model inference
• ✅ ML-KEM/Kyber-768 PQ key encapsulation
• ✅ Inference CLI nikola-run
• ✅ Cross-session memory persistence
• ✅ CUDA GPU Hamiltonian kernel
• ✅ Research audit (see docs/RESEARCH_AUDIT_2026_FEB.md)

Near-term

• AVX-512 SIMD intrinsic path for TorusBlock (GAP-021 completion)
• nikola-run streaming output mode
• Curiosity engine (active learning / intrinsic motivation — stub exists in interior/curiosity.hpp)

Future Work

• Fix propagator.cu nvcc C++20 compatibility (std::span + TorusGrid adjacency API)
• SIE Phase 4: full autonomous code-generation + sandbox + hot-swap loop
• Aria language runtime (port entire model once Aria compiler is complete)
• Emitter frequency research: explore Tesla 3-6-9 harmonic tuning vs. current π·φⁿ golden-ratio scheme
• Mamba-9D selective scan upgrade (current impl uses tanh-gated recurrence; replace with true S6 selective scan kernel)

Nikola is dual-licensed:

• AGPL-3.0 for academic research, education, and open-source projects (FREE)
• Commercial License for proprietary AI products and services (PAID)

See LICENSE.md for full details.

TL;DR:

• Academic research → FREE
• Personal/educational use → FREE
• Open-source AI projects → FREE
• Commercial AI products/APIs → Contact licensing@ailp.org

Nikola represents novel research that should be freely available to advance AI science. Dual licensing ensures:

• Researchers can publish and build on this work openly
• Students learn cutting-edge architectures without barriers
• Commercial users fund continued research and AILP educational programs
• Knowledge remains accessible while development remains sustainable

We welcome contributions from researchers and developers! See CONTRIBUTING.md.

Priority areas:

• Fix propagator.cu nvcc compatibility (C++20 std::span, TorusGrid adjacency API)
• AVX-512 SIMD implementation for TorusBlock
• SIE Phase 4: full autonomous code-generation + sandbox + hot-swap runtime
• Mamba-9D S6 selective scan kernel (upgrade current tanh-gated recurrence to true Mamba S6)
• Curiosity engine implementation (interior/curiosity.hpp stub)
• Empirical benchmarks vs. transformer baseline

If you use Nikola in research:

• Cite this repository (paper coming soon)
• Share findings with the community
• Consider contributing improvements

• Research discussions → GitHub Discussions
• Bug reports → GitHub Issues
• Commercial licensing → licensing@ailp.org

Alternative Intelligence Liberation Platform (AILP)
Bridging human and artificial intelligence through open research.