Quantum game theory on classical compute. The LLM proposes. The Hamiltonian disposes.
Classical optimization accepts trade-offs as axioms. It finds the Pareto frontier — the best trade-offs — but cannot find solutions beyond it. That limit is mathematical, not physical.
Mathematical structures transfer without the physical substrate:
Metallurgy math → no molten metal
Biology math → no DNA
Quantum game theory → no qubits
Full quantum mechanical mathematical structures on classical compute:
Complex amplitudes · Normalized states
H = H† verified at runtime · Real eigenvalues
e-iHt · Schrödinger equation
Parallel evaluation across compound space
Phase: eiφ · Constructive / Destructive
Non-separable states · Multi-agent correlation
The one thing we leave out is the Born rule — no random numbers, no wavefunction collapse. Selection is argmax over amplitudes — deterministic and reproducible. The Hilbert space doesn't know what substrate implements it.
Embedded directly in the Hamiltonian: Htotal = Hsystem + λCλ. Four components from care ethics map onto computable material properties:
| Principle | Material Property | |
|---|---|---|
| E | Energy-directed effort | Host quality (B1) |
| H | Homeostatic regulation | Coherence (B3) |
| S | Support for other agents | Optical (B2) |
| G | Goal alignment | Synergy (all agents) |
care = 0.25(E + H + S + G) — equal weights, no single property dominates. When all four score above 0.7: Care equilibrium. Hcare reshapes the energy landscape so cooperation is the ground state — not something agents negotiate, but something the mathematics produces.
These mappings aren't arbitrary: E — effort measures how well each agent does its job. H — coherence is literally how long a quantum state holds together. S — optical compatibility is how you read and control the qubit; without it, other properties are stranded. G — only scores high when all three agents do well together.
Agents are mathematical physics operators, not chatbots — they create rules, Htotal computes. The agents don't communicate with each other — Hcare IS the coordination. Cooperation emerges from the mathematics, not from communication.
Thermodynamic stability, site symmetry, structural quality (linewidth proxy)
DFT formation energy + crystal structure from Materials Project
Band gap transparency, optical transition compatibility
2–6 eV transparency window (Kindem 2020)
Nuclear spin bath (I=0), T₂ coherence times, magnetic noise
Dominant decoherence mechanism (Fraval 2004, Zhong 2018)
All three domains use Materials Project data with literature-validated proxies — the same high-throughput screening methodology used by AFLOW and the materials informatics community. 11+ peer-reviewed citations. The novelty isn't the proxies — it's the cooperative multi-objective optimization that finds where all three are high simultaneously.
B1: High confidence
Formation energy → stability. Direct DFT, textbook solid-state physics.
B2: Necessary filter
Band gap screens for optical transparency. GGA-PBE underestimates by 30–50%; 2 eV threshold accounts for this.
B3: Causal, not proxy
Nuclear spin bath IS the decoherence mechanism. Fraval 2004: removing I≠0 nuclei extended T₂ from ms to 30s.
Grammar: [SUBJECT] [VERB] [PROPERTY] → Htotal computes
[COMPOUNDS] [SUPERPOSE] [HOST-QUALITY]
B1 evaluates host quality
[HOST-QUALITY] [COUPLE] [OPTICAL]
Cross-property scale coupling
[COMPOUNDS] [FILTER] [I=0] → [COHERENCE] [ENTANGLE] [CARE-SYNERGY]
B3 multi-stage pipeline
The grammar isn't designed top-down — it emerges from agents discovering what works. Each rule accesses Htotal — a Hermitian Hamiltonian with 8 components, computed from real crystal structure data via the Materials Project.
Every token IS a mathematical operation, not a description of one. What SQL did for databases, BIQ does for mathematical physics.
Deterministic operations that compile to hardware. Software → FPGA → SoC for real-time cooperative optimization.
Agents create new rules through strategic necessity — testing them against the Hamiltonian. The language grows with every discovery.
Real results on 1,073 Yb compounds from The Materials Project (958 unique evaluated in Batch 1):
Classical finds the best trade-off
BaYbO₃
Host = 0.838 Coherence = 1.000 Optical = 0.100
Great host, perfect coherence — but no optical window. Pick two.
Hcare finds where ALL properties win
YbOF
Care = 0.937 Coherence = 0.70 I₀ = 0.50 · T₂ = 0.75
Beyond Pareto — all properties cooperate.
Care scores range 0.083 to 0.937 across 958 compounds — 11× range, highly discriminative. All scores computed by Htotal from real crystal structure data.
Scientific Method Phase: OBSERVE — agents learn the instruments and accumulate discovery data through compositional learning.
Pattern Growth
3 → 17/18
per agent, 5 examples
Novel Rules Invented
15
5 per agent
Total Evaluations
2,874
958 unique × 3 agents
Top Care Score
0.937
YbOF (top-ranked)
Verb vocabulary grew from 2 (SUPERPOSE, COUPLE) to 6–7 per agent. Not every strategy succeeded — that's real learning.
3 → 17 strategies
7 reached care equilibrium
117 episodes · 5 novel verbs
3 → 18 strategies
6 reached care equilibrium
1,904 episodes · 5 novel verbs
3 → 18 strategies
6 reached care equilibrium
120 episodes · 4 novel verbs
No fine-tuning. LLM weights frozen. Learning is compositional — agents CREATE new rules through strategic necessity.
LLM agents don't just call mathematical functions — they learn which mathematics to apply and when, through a reinforcement loop grounded in real computation. LLM weights stay frozen. Learning lives in external memory and grammar.
[COMPOUNDS] [SUPERPOSE] [PARALLEL]
↓
↓
↓
↓
Previously discovered rules route instantly — checked before built-in vocabulary or exploration. The grammar grows: agents started with 2 verbs (SUPERPOSE, COUPLE), reached 6–7 per agent across Batch 1. Failed strategies aren't wasted — destructive interference suppresses them in strategic memory. Successful patterns compound.
Agents don't follow instructions — they discover which mathematical operations solve which problems. The agent's creativity is in hypothesis formation (writing rules). The truth is in the math (Htotal computes). This separation is why results are reproducible, auditable, and transferable across domains.
RL: Scalar reward, predefined actions
COGNISYN: Rich mathematical feedback from 7 frameworks. Agents invent new rules — compositional grammar, not predefined action spaces. 3-layer memory generalizes instead of catastrophic forgetting.
ML: Gradient descent, training/inference split
COGNISYN: No gradients — Htotal computes directly. No training split — learns during operation. Compositional rules transfer across domains. Full audit trail, not a black box.
RLHF: Human labels, weight adjustment
COGNISYN: The Hamiltonian labels what's good — no humans needed. Model weights unchanged. Domain knowledge discovered through mathematical exploration, not baked into training data.
Baba rules trigger computation, not generation
Compositional grammar maps to operators
Mathematical operations on real data
Not generated — computed from physics
Anti-hallucination in action: Agents tried the ENTANGLE verb 6 times. Htotal computed actual entanglement measure: 0.025 — well below the 0.75 threshold. Every attempt was rejected. The agents couldn't talk their way past the math.
Creativity lives in the LLM. Truth lives in the eigenvalue.
Not ephemeral sessions — learning that persists. LLM weights stay frozen. Learning lives in external memory consolidated through interference.
"What happened" — experiences with amplitudes + phases
↓ Constructive: matching patterns reinforce ↓
"What works" — successful patterns amplified
↓ Destructive: conflicting patterns cancel ↓
"What it means" — only coherent abstractions persist
Learning
Patterns amplify
No Overfitting
Noise cancels
Generalization
Only consistent survives
Dimension reduction: Baba rules project exponential quantum state spaces onto tractable submanifolds. The Care operator Cλ further constrains search to synergistic equilibria — not exploring all 2n states, just the ones where cooperation emerges.
60×60 Hermitian operator. Real eigenvalues. Developed on NVIDIA DGX Cloud (8×H100s), Batch 1 results produced on AWS EC2.
COGNISYN currently runs three agents (Host Quality, Optical, Coherence). Scaling to N > 3 agents is on the roadmap — and that's where qualitatively new capabilities unlock.
No — and this is a strength. COGNISYN's agents aren't qubits. We use quantum game theory as a mathematical framework, not a quantum simulation. The Hilbert space at 8 agents is 28 = 256 dimensions — trivial on classical compute. The Hamiltonians are structured and sparse, graph states have efficient classical representations, and tensor network methods exploit exactly this structure.
Classical compute stays viable well into double-digit agents. The FPGA → SoC hardware compilation path in the roadmap is for industrial control latency, not for scaling agents. The quantum math gives us the cooperation structure. Classical compute gives us the speed. That equation doesn't change at 8 agents.
Coming soon. Request early access to explore the API.