Research & Development

Build downward.
Prove more.

Synapse removes layers between model intent and execution: from frameworks and containers down to WebAssembly sandboxes, formal verification, receipts, and eventually sovereign bare-metal compute. The self-hosted runtime is available today. Other areas on this page are research, demonstration, or roadmap work and are labeled accordingly. All claims are backed by reproducible tests, journals, and formal proofs where applicable.

Synapse|Vernon, Canada|All R&D conducted in Canada
The Thesis

Strip every layer between intent and silicon.

Most AI companies build upward — more frameworks, more cloud services, more abstraction. Synapse builds downward — fewer layers, lower overhead, tighter control, and more deterministic execution. The horizon points toward bare metal. The self-hosted runtime is the product available today.

INDUSTRY: Building UPWARD → more frameworks, more cloud, more abstraction
SYNAPSE: Building DOWNWARD → fewer layers, closer to silicon, toward deterministic physics
Layers Removed
#
Layer
Why It Existed
How We Removed It
Phase
1
Containers
Isolation between workloads
Wasm sandboxes: 64KB vs 128MB Docker
1-5
2
Human syntax
Programmers need to read code
.syn: 501-token machine-native language
1-5
3
Framework tax
PyTorch/NumPy for ML
Tape-based autograd as FFI
6-8
4
Safety theater
Runtime testing, evals
Z3 formal proofs at compile time
9-20
5
Human reward
RLHF needs human labelers
Compiler IS the reward model
21-25
6
The OS itself
Unsafe code needs kernel
UEFI bare-metal unikernel
boot/
The Verified AI Loop

A complete pipeline where every stage carries formal guarantees. No external ML frameworks. No containers. No cloud APIs. Each step has working code and passing tests.

01

Train

Transformer training orchestrated by .syn code. Self-attention, layer normalization, residual connections, AdamW optimizer.

02

Verify

Z3 proves training correctness: gradient math, loss convergence, structural integrity. 120+ automated tests.

03

Serve

Self-verifying inference. The program checks its own output against formal bounds before returning results.

04

Improve

Verified self-improvement, composition, and fine-tuning — each with Z3 proofs that safety properties are preserved.

Honest scope:The verified AI training pipeline operates on small models (80–160 parameters) at demonstration scale. Compositional verification scales to 50M+ parameters (8.7ms for 12 layers) but produces conservative bounds due to interval arithmetic. The production execution engine (Rust gateway, Wasm sandboxing, code compilation) is real infrastructure. The verified ML lifecycle (training, self-improvement, federated learning) is research demonstrated via Python scripts, not production gateway features. The contribution is the architecture and the proofs, not the parameter count.

Technological Uncertainties

Each area of investigation addresses a question that could not be resolved through standard engineering practice or reference to publicly available knowledge. Resolution required systematic experimental investigation.

U1

Can formal verification integrate with WebAssembly compilation at production latency?

Resolved

No existing WebAssembly runtime integrates SMT theorem proving at compile time. We investigated whether Z3 proofs could complete within production latency bounds and whether property invariants could be statically verified without full program analysis.

Result

Z3 verification completes in <1ms for property invariants, ~14ms for expression proofs. The verification pipeline runs in the production gateway.

U2

Can a self-hosting compiler operate within Wasm linear memory constraints?

Resolved

Self-hosting compilers exist (GCC, Rust, Go) but all run on a full operating system with virtual memory and dynamic heap allocation. A self-hosting compiler constrained to Wasm linear memory with a fixed arena allocator and a 64KB address space had no documented precedent.

Result

The .syn compiler (~2,000 lines of .syn) compiles to WebAssembly and runs on the same runtime it targets. Bootstrap verification confirms identical output to the reference compiler.

U3

Can automatic differentiation work when orchestrated by a Wasm module across a host FFI boundary?

Resolved

Standard autograd systems (PyTorch, JAX) operate within a single process with shared memory. We investigated whether reverse-mode differentiation could function when the forward pass is orchestrated by a sandboxed Wasm module calling tensor operations through host FFI, with gradient computation managed entirely by the host.

Result

Gradient-verified training at multiple scales. Convergence proven on transformer architectures with self-attention, layer normalization, and residual connections. Results match reference implementations to f32 precision.

U4

Can SMT solvers provide meaningful formal guarantees about neural network training dynamics?

Resolved

Existing neural network verification focuses on inference-time properties (robustness, output bounds). We found no prior work combining SMT solvers with step-level training verification — proving properties about loss monotonicity, gradient correctness, and convergence during training.

Result

Z3 proves training loop structure, gradient formula correctness (8/8 operations, 38ms), output bounds for trained models, and safety preservation across fine-tuning. Three-layer verification: compile-time structural, runtime assertions, and offline formal proofs.

U5

Can independently trained models be composed with formal end-to-end safety guarantees?

Resolved

When chaining neural networks (Model A output feeds Model B input), output bounds typically compound — each model adds uncertainty. We investigated whether transformer architectures enable compositional verification where bounds contract rather than expand.

Result

Z3 proves composed output bounds for multi-model pipelines. Architectural properties of transformer networks cause bounds to contract at each composition stage, making composed systems formally safer than individual models.

U6

Can WebAssembly sandboxing achieve container-equivalent isolation at orders-of-magnitude lower cost?

Demonstrated

Industry standard isolation (AWS Firecracker, Docker) requires 128MB+ per sandbox with 100-200ms cold starts. We investigated whether Wasm linear memory could provide equivalent isolation at dramatically lower resource cost.

Result

64KB per sandbox, <2ms cold start, memory wiped between requests. No guest filesystem, no guest network access. Theoretical density of ~130,000 concurrent sandboxes per 128GB server.

U7

Can differential privacy be formally verified for neural network training inside WebAssembly?

Resolved

DP-SGD (gradient clipping + calibrated Gaussian noise) is the standard for private ML training, but existing implementations rely on empirical testing. We investigated whether Z3 can formally prove that privacy noise calibration, composition bounds, and gradient clipping satisfy their mathematical guarantees.

Result

Z3 proves four properties: noise sigma matches the Gaussian mechanism formula, composition budget is correctly computed, noise is sufficient for claimed epsilon, and gradient clipping holds universally. Training converges despite substantial noise injection.

Technical Achievements

Each achievement has working code, automated tests, and formal proofs where applicable. Classification distinguishes genuine novel R&D from applied research.

Novel R&D
+48.6pp advantage over RNN

.nerve Neural Computation Substrate

Novel typed DAG substrate with 7 biologically-inspired operations. Outperforms ternary weight matrices by 48.6 percentage points on temporal tasks (99.2% vs 50.7%, N=5 seeds, p<0.001). The substrate IS the invention.

Novel R&D
+11%vs float baseline

Ternary Mamba Outperforms Float

Ternary Mamba (State Space Model) variant outperforms float on Shakespeare character-level LM. 458K params, N=5 seeds. First demonstration that ternary SSMs can exceed float accuracy. Scaling to 1B+ params underway.

Novel R&D
9KB receipted Wasm (J149)

Verified SSM Training Loop

World’s first State Space Model (Mamba) training inside a verified, fuel-metered Wasm sandbox. Full forward/backward through selective scan, SiLU gating, RMSNorm, and AdamW. Every weight update is receipted. 254/254 tests passing. Red-team audited (RT8).

Novel R&D
3,031bytes total

Verified Neural Deployment

Trained .nerve DAG compiled to a deterministic integer-only program. Provably terminates, bounded memory, zero floating-point operations. Fits in 2 TCP packets. Runs on any CPU.

Novel R&D
97.8%accuracy (N=5)

97.8% Ternary Spiking MNIST

Scaffold Architecture: ternary convolution + SpikeMIM spiking classification. All weights ∈ {-1,0,+1}. Adversarially audited (RT7: 7/7 PASS, 3 WARN, 0 FAIL).

Demonstrated
3independent nodes

3-Node Cross-Continental Verification

Germany, Finland, and client browser independently execute identical computation and produce bit-identical SHA-256 receipts. Live at synapserun.dev/demo. No blockchain required.

Novel R&D
8/8ops proved (38ms)

Z3-Proved Gradient Correctness

Every differentiable operation in the autograd engine formally verified by Z3. Dual validation with numerical finite differences confirms agreement at machine precision.

Novel R&D
~2,000lines of .syn

Self-Hosting Compiler in Wasm Linear Memory

A compiler that compiles its own source language, constrained to Wasm linear memory with a fixed arena allocator. No OS, no heap, no string libraries. No documented precedent.

Novel R&D
8.7ms for 50.7M params

Compositional Verification at Scale

Layer-by-layer Z3 verification that scales linearly. Synapse-50M verified across all 12 transformer layers. Bounds stabilize after layer 2.

Novel R&D
3principled negative results

3 Backprop-Free Approaches Falsified

Hebbian, Correlation-Guided Wiring, and Discrete Target Propagation systematically tested and falsified with mechanistic diagnoses. Narrows the research frontier for future work.

Applied R&D
87%acceptance rate

Self-Distillation Loop

Model generates .syn programs, runs them through verified execution, and accepts programs meeting quality thresholds as new training data. Bootstraps improvement from its own outputs.

Novel R&D
7stages, one receipt

Sovereign AI Node

Train, self-improve, fine-tune, merge, verify, RL reward, and receipt — all in a single .syn execution. 28 weight tensors, fully deterministic.

Applied R&D
88,000evals/sec/core

Compiler-Stage GRPO Training

Novel reward signal derived from compiler pipeline stages. In-process Wasm evaluation achieves 797x acceleration over CPython for reinforcement learning.

Efficiency Thesis

For deterministic computation workloads, formally verified small models on commodity hardware can achieve equivalent correctness at a fraction of the compute cost.

Factor
Industry Standard
Synapse Approach
Gain
Correctness method
Empirical evaluation
Z3 formal proofs
Proven
Training verification
None (loss curves)
Gradient + convergence proofs
Formal
Execution sandbox
128MB+ container
64KB Wasm arena
2000x
Framework dependency
PyTorch + CUDA stack
Single binary, no frameworks
Eliminated
Hardware required
NVIDIA H100 ($25K+)
Commodity CPU ($60/mo)
GPU eliminated

Comparison applies to deterministic computation tasks, not general-purpose AI. The verified AI pipeline operates at demonstration scale. Production scaling is an active research direction.

Environmental Impact Potential
7-9x
Less energy per execution vs Docker
0%
GPU dependency for verified workloads
15MB
Total runtime binary size

AI infrastructure consumes approximately 10 TWh of electricity annually, growing 30%+ year-over-year. For deterministic workloads where correctness can be formally proven, Synapse eliminates the need for GPU hardware entirely by running verified models on commodity CPUs inside Wasm sandboxes.

Measured energy reduction (RAPL instrumentation, AMD EPYC) is 7-9x per execution vs Docker containers on identical hardware. All research conducted in Canada by Synapse

Measurement Methodology
Hardware
Primary Node (Germany)
Hetzner AX102, AMD Ryzen 9 7950X3D, 128GB DDR5
Verification Node (Finland)
Hetzner CX23, AMD EPYC, Helsinki
Research Hardware
Apple M4 Pro, local training + .nerve experiments
Verified Measurements
+48.6pp.nerve DAG vs ternary RNN (99.2% vs 50.7%, N=5 seeds, p<0.001)
+11%Ternary Mamba vs float baseline (Shakespeare char-level LM, N=5 seeds)
9KBVerified SSM training loop — receipted Mamba in .syn (J149, RT8)
7-9xEnergy reduction vs Docker (RAPL measurement, matched workloads)
38msZ3 proof time for all 8 autograd gradient formulas
<1msZ3 property verification latency (compile-time invariants)

All benchmarks are automated and reproducible. Methodology details are documented alongside each measurement.

Systematic Investigation

Each investigation follows a hypothesis-experiment-result cycle documented in dated journal entries. 145+ entries spanning 45+ research phases, with 8 adversarial red team audits.

Automatic differentiation across Wasm-host boundary

Hypothesis

Reverse-mode backpropagation can function when the forward pass is orchestrated by a sandboxed Wasm module and gradient computation is managed by the host.

Method

Implemented autograd system with gradient-verified training at multiple scales. Tested against reference implementations at f32 precision.

Result

Exact match with reference at 2x4, 4x8, and 4x16 dimensions. Loss convergence verified across all scales.

Formal verification of training dynamics

Hypothesis

Z3 can prove meaningful properties about neural network training: gradient correctness, loss monotonicity, convergence, and safety preservation.

Method

Built three-layer verification system (compile-time, runtime, and offline formal proofs). Applied across MLP and transformer architectures.

Result

All 8 backward formulas Z3-proved in 38ms. Training convergence verified. Fine-tuning safety preservation proved with three simultaneous Z3 properties.

Compositional neural network verification

Hypothesis

Independently trained transformer models can be composed with formal end-to-end safety guarantees.

Method

Trained two transformers on different data, composed their inference, and applied formal bound propagation through the composition.

Result

Z3 proves composed output bounds in 23ms. Architectural properties cause bounds to contract at each composition stage.

Compiler-stage reward shaping

Hypothesis

Compiler pipeline stages can serve as a structured training signal for reinforcement learning.

Method

Trained Synapse-50M using reward signals from each compiler stage. Measured evaluation throughput via in-process Wasm.

Result

88,000 evals/sec/core. Model generates compilable .syn at >90% first-attempt success rate.

Development Roadmap
Phases 1-5

Execution Engine & Compiler

complete
Self-hosting .syn compiler in Wasm linear memory
Z3 formal verification at compile time
Python-to-.syn transpiler
Bare-metal UEFI unikernel
GRPO training with compiler-stage rewards
Phase 6

ML Type System & Neural Inference

complete
f32 type inference and arrays in .syn
Matrix operations and GPU FFI bridge
Transformer layer forward pass in .syn
Compiler maintained at ~2,000 lines
Phases 7-8

Sovereign Neural Network Training

complete
Automatic differentiation across Wasm-host boundary
Complete transformer training with self-attention and residual connections
AdamW optimizer with persistent state
Gradient-verified at three scales against reference implementations
Phases 9-10

Verified Training & Inference

complete
Z3 structural verification of training loop correctness
Runtime training assertions (loss monotonicity, convergence)
Self-verifying inference with output gating
Complete sovereign AI loop: train, export, serve, verify
Phases 11-14

Verified AI Properties

complete
Verified self-improvement with Z3-proved loss reduction
All 8 autograd backward formulas Z3-proved correct
Multi-model composition with formal end-to-end safety
Verified fine-tuning: bounds preserved, no forgetting, effective transfer
Phase 15

Verified Differential Privacy

complete
DP-SGD with gradient clipping and calibrated Gaussian noise
Z3-proved privacy budget: sigma formula, composition bounds, noise sufficiency
Universal gradient clipping proof for all possible gradient norms
Training converges under formal noise injection
Phases 16-20

Verified AI Runtime & Sovereign Node

complete
Cryptographic execution receipts with SHA-256 hash chains
Self-hosting ML pipeline: train, infer, fine-tune, evaluate, receipt — one execution
Verified reinforcement learning (REINFORCE with reward-scaled gradients)
Verified model merging with bounded degradation
Sovereign AI node: 7-stage pipeline, 28 tensors, one receipt
Phases 21-25

Code Generation & Continuous Training

complete
Verified execution bridge: model generates .syn → compile → verify → receipt
Training-loop reward experiments and compiler-ranked evaluation
Training data expansion: 68+ FFI functions covered
Continuous training daemon with pause/resume
Synapse-50M deployed for .syn generation experiments
Phases 26-30

Scalable Verification & Sovereign Network

complete
Compositional verification: Synapse-50M (50.7M params) verified in 8.7ms
Self-distillation: model generates its own verified training data (87% acceptance)
Multi-tenant receipt chains with per-tenant Merkle linkage
Federated verified training: 3 simulated nodes with verified weight sharing
Sovereign AI Network: research simulation of an autonomous verified AI network
Phases 31-38

Spiking & Ternary Neural Research

complete
Scaffold Architecture: 97.8% MNIST with ternary weights + LIF spiking (N=5 seeds)
SpikeMIM local learning: 88.0% MNIST with zero backpropagation
6 WGSL compute shaders validated on Apple M4 Metal
Ternary attention wall: 3 approaches systematically falsified
Adversarial audit RT7: 7/7 PASS, 3 WARN, 0 FAIL
Phases 39-45

.nerve Substrate & Ternary Mamba

complete
.nerve substrate: DAG outperforms RNN by 48.6 pp (99.2% vs 50.7%, p<0.001)
Shakespeare scaling: 68.8% bit accuracy at 256 nodes
Verified deployment: 3,031-byte integer-only program with 4 proven invariants
Ternary Mamba: outperforms float by 11% on Shakespeare (458K params, N=5)
3 backprop-free approaches falsified with mechanistic diagnoses
Phase 46+

Scale to Language Intelligence

active
Scale .nerve DAGs: 256 → 1,024 → 10,000+ nodes
Scale ternary Mamba: 458K → 3.5M → 25M → 1B+ parameters
Cross-entropy loss for character-level language generation
Mesh compute: scale 3-node verification to 10+ consumer devices
NoProp/SpikeMIM gradient-free structural learning research
Repository Statistics
145+
Journal Entries
R&D documentation
45+
Research Phases
Complete
254+
Test Cases
Automated
8
Red Team Audits
Adversarial review
827+
Knowledge Records
Knowledge graph
25+
Novel Breakthroughs
No known prior art
15MB
Gateway Binary
Single Rust binary
~2,000 lines
Compiler
.syn (self-hosted)
Commercial Applications

Verified AI addresses a gap in regulated industries: the ability to formally prove that AI models preserve safety properties after adaptation. No existing technology provides this guarantee.

Healthcare

Health Canada, FDA

Hospitals fine-tune diagnostic models for local populations. Verified fine-tuning proves original safety properties are preserved after adaptation.

Financial Services

OSFI Guideline E-23

Banks adapt credit scoring models for new products. Formal verification satisfies model risk management requirements.

Defense

DND

Military AI adapted for new environments needs formal guarantees that existing capabilities are preserved.

AI Safety

Emerging standards

Safe self-modification — the alignment community's core concern — implemented as verified self-improvement with Z3 proofs.

Canadian Intellectual Property

Synapse is developed by Synapse, incorporated in British Columbia, Canada. All research and development is conducted in Canada by Canadian residents.

The following are novel Canadian IP developed through systematic R&D:

The .syn programming language and specification
Self-hosting compiler in Wasm linear memory
Z3 formal verification for Wasm compilation
Verified neural network training pipeline
Self-verifying inference architecture
Multi-model compositional safety verification
Verified fine-tuning with safety preservation
Formally verified differential privacy for training
Compiler-stage GRPO reward shaping algorithm
Sovereign AI node: full ML lifecycle in one execution
Compositional verification scaling to 50M+ parameters
Federated verified training with receipt attestation

By targeting open hardware (AMD, open-source drivers) instead of proprietary ecosystems (NVIDIA CUDA), Synapse avoids dependency on a single foreign hardware vendor, aligning with Canadian digital sovereignty objectives.