zkML + Circuit Stack

zkde.fi uses zero-knowledge machine learning to turn AI model outputs into verifiable, policy-aware signals. The system includes 31 Circom circuits, 2 EZKL ML models, and a bridge layer (ModelBridge) that converts arbitrary ONNX model outputs into on-chain-verifiable commitments via Garaga.

The result: AI outputs become evidence-bearing, policy-aware signals — not black-box recommendations.

The EZKL Pipeline

flowchart LR
  ONNX[ONNX model] --> EZKL[EZKL: quantize + compile]
  EZKL --> KZG[KZG commitment to model weights]
  KZG --> INFER[Private inference on user data]
  INFER --> G16[Groth16 proof of correct inference]
  G16 --> BRIDGE[ModelBridge circuit]
  BRIDGE --> GAR[Garaga KZG pairing check in Cairo]
  GAR --> CHAIN[On-chain verified result]

1. Model training — a standard ML model is trained and exported as ONNX
2. EZKL compilation — EZKL quantizes the model and compiles it into a Halo2 circuit with KZG commitments to the model weights
3. KZG commitment — the model's weight commitment is registered on-chain, binding the model's identity
4. Inference — the model runs on private user inputs; the user's data never leaves their control
5. Groth16 proof — EZKL generates a proof that the model ran correctly on the given inputs and produced the given output
6. ModelBridge — a Circom circuit that takes the EZKL proof output and bridges it into a format verifiable by Garaga's BN254 pairing check
7. On-chain verification — Garaga verifies the Groth16 proof in Cairo on Starknet

EZKL ML Models

Model	Architecture	Classes	Purpose
yield_forecast	Linear(12→32)→ReLU→Linear(32→16)→ReLU→Linear(16→4)	declining / stable / growing / surging	Predict yield trajectory for pool allocation decisions
anomaly_detector	Linear(8→24)→ReLU→Linear(24→12)→ReLU→Linear(12→3)	safe / warning / critical	Detect anomalous pool behavior before capital enters

Both models are compiled via EZKL into Halo2 circuits with KZG polynomial commitments. Proofs are verified on Ethereum Sepolia via the Halo2Verifier and bridged to Starknet via L1→L2 messaging or the ModelBridge circuit path.

The 13-Circuit Agent Screening Bundle

Before any execution path is surfaced in Trade Desk, the relevant circuits from this bundle evaluate the action:

#	Circuit	What It Proves
1	RiskScore	Aggregate risk rating for a pool or position
2	AnomalyDetector	Whether pool behavior deviates from safe baselines
3	YieldOptimality	Whether the yield justifies the risk at current conditions
4	StrategyIntegrity	Whether the proposed strategy matches the committed parameters
5	ExecutionIntegrity	Whether the execution respected declared constraints (slippage, timing)
6	ImpermanentLossPredictor	Estimated IL exposure for LP positions
7	SlippageBound	Whether expected slippage stays within acceptable bounds
8	MEVResistanceProof	Whether the execution route has MEV protection characteristics
9	LiquidationRisk	Proximity to liquidation threshold for leveraged positions
10	CorrelationRisk	Cross-asset correlation risk in the portfolio context
11	CreditEligibility	Whether collateral posture meets lending thresholds
12	SolvencyProof	Verifiable solvency attestation
13	AgentReputationScore	Composite trust score from execution history

Each circuit is compiled to WASM + zkey for fast client-side proof generation. The FICO-pack subset (Solvency, RiskPassportTier, TraderPerformance, StrategyIntegrity, ExecutionIntegrity) has dedicated on-chain verifier contracts:

Verifier	Address
SolvencyProofVerifier	0x043b253e3f2fcac35eef0b08fd2f8f4ff81aeb52848f11640d62879854329c9b
RiskPassportTierVerifier	0x05e71cc0c4b87908230414644d675164fb90cd6d8cfafeae87198241e60eb788
TraderPerformanceVerifier	0x04c8087855dd0812042de58b2a3f3838d3cea45118c86f07d32ac87648e90769
StrategyIntegrityVerifier	0x00c9478f355bdad25caf13899a0d5bf2ee1accb1678e9934ebeda40f2653e549
ExecutionIntegrityVerifier	0x03bb26a38ea2d8e4bd21895f665d0056a5496f31ad84f4d77e040d9e63e6873b

Additional Circom Circuits (Full 31-Circuit Set)

Beyond the 13 agent screening circuits, the system includes circuits for privacy, governance, and bridge operations:

Category	Circuits
Privacy	FullPrivacyWithdraw, FullPrivacyWithdrawHashed, FullPrivacyWithdrawWithChange, PrivateDeposit, PrivateWithdraw, PoolMembership
Bridge	ModelBridge, ModelBridgeHeavy
Governance	private_vote
Position	TWAPPosition, RebalanceTimingCommitment
Attestation	HistoricalPerformanceAttestation, RobustnessCertificate, SafetyDiversification
Threshold	BalanceAboveThreshold, TenureAboveThreshold
DeFi	CrossProtocolArbitrage, RiskPassportTier, TraderPerformanceProof

Plus a Noir circuit (noir_ezkl_bridge/) for the Noir HONK proving lane.

ModelBridge vs ModelBridgeHeavy

Variant	Artifact Size	Proving Time	Use Case
ModelBridge	Standard (~2MB zkey)	~3-5s	Default bridge for standard model outputs
ModelBridgeHeavy	Heavy (~8MB zkey)	~10-15s	Complex model outputs requiring more constraint capacity

Both variants produce Groth16 proofs verified by the same Garaga verifier infrastructure on Starknet.

ModelBridge Verifier Contract

The on-chain entry point for verifying bridged EZKL proofs:

0x037c42e8734271aca0c3c1bdf1746d9ccc098ddfd5ee211c94bbb8786fa4626f (Starknet Sepolia)

zkML API Endpoints

Method	Endpoint	Purpose
POST	/api/v1/zkdefi/zkml/risk_score	Risk score proof/signal generation
POST	/api/v1/zkdefi/zkml/anomaly	Anomaly detection signal
POST	/api/v1/zkdefi/zkml/combined	Combined risk + anomaly path
GET	/api/v1/zkdefi/zkml/status	zkML subsystem status
GET	/api/v1/zkdefi/zkml/pool-safety	Pool safety snapshot
POST	/api/v1/zkdefi/zkml/scan	Scan-oriented model path
GET	/api/v1/zkdefi/zkml/circuits	Circuit metadata

---

Next: Proof Pipeline | Privacy Rails | API Overview