Loading…

obsqra.xyz/starkforge.xyz/zkde.fi
Starknet Sepolia live
Open console
Composable ZK execution layer · Starknet

Where proof systems
converge.

Composable. Private. Portable. Trustless by construction.

Every ZK project picks one proof system. StarkForge composes five — because proving AI execution requires guarantees no single system provides: zkML inference, policy enforcement, execution integrity, and privacy, all working together. The output is composable, portable, and private by construction. Proofs settle recursively from Madara L3 to Starknet L2 to Ethereum L1. AI agents prove before they act. Every action produces a receipt. No backend trust required at any layer.

5
Proof systems composed
31
Circom circuits
316
On-chain receipts
14/14
Claims validated live
3
Settlement paths
40+
Deployed contracts
What the composition produces

Composability. Privacy. Portability.
Trustless by construction.

These are not features. They are properties that emerge when multiple ZK proof systems compose correctly. StarkForge is the layer that makes the composition work.

01 — Composability
Five proof systems.
One enforcement primitive.

No single ZK system provides all the guarantees needed for verifiable AI execution. SNARK for model integrity. STARK for execution faithfulness. KZG for weight commitment. Noir for alternative paths. Poseidon for fast fallback. StarkForge composes them so you don't have to choose.

//SNARK-in-STARK dual proving
//Five independent verifier lanes
//ModelBridge as the binding primitive
//No single point of failure
02 — Privacy
Not a feature.
A first-class output.

Privacy works because the ZK composition makes it economically viable at every execution boundary. Commitment/nullifier rails, shielded pools, relayer networks, dark settlement, private governance — all confirmed live. The proof enforces the rule without revealing the data.

//Commitment / nullifier / Merkle tree
//Relayer — breaks tx graph link
//Selective disclosure — prove without revealing
//Dark settlement + private governance live
03 — Portability
Receipts that travel.
Trust that doesn't expire.

A fact registered in ObsqraFactRegistry on Madara L3 can propagate to Starknet L2 and settle to Ethereum L1 — all confirmed. FICO-pack ZK credentials travel cross-protocol. Any contract calls is_valid(fact_hash) without a backend dependency. The receipt is the trust.

//L3 → L2 → L1 tri-chain settlement
//Portable FICO-pack ZK credentials
//is_valid(fact_hash) — no backend trust
//ModelBridge chain-agnostic by design

The composition
is the primitive.

Picking one proof system means accepting its blind spots. Groth16 is fast but has a trusted setup. STARK has no trusted setup but can't verify ML inference natively. EZKL proves the model but can't enforce on-chain policy. Each system is incomplete. StarkForge composes all of them — so the output carries every guarantee simultaneously, and there is no single point of failure.

EZKL
zkML layer
ML inference — proven
ONNX model quantized and compiled into a Halo2 circuit via EZKL. KZG commitment to model weights registered on-chain. Private inputs never leave the user. A different model produces a different commitment — and fails.
KZG commitment
model identity bound
bridges proof formats
ModelBridge
composition layer
Policy enforced — formats bridged
Open-source Circom circuit that takes EZKL's Halo2 output and re-proves the key facts in Groth16 — the format Garaga can verify on-chain. Binds model identity, input hash, policy bounds, and timestamp into one commitment. The bridge that makes the composition possible.
Groth16 commitment
bounds enforced
verifies on-chain
Garaga
SNARK verification
Groth16 verified — execution gated
BN254 elliptic curve pairing check implemented in native Cairo. Execution is blocked until ModelBridge's Groth16 proof passes. This is not monitoring — the proof is the gate. No valid proof, no execution, no exceptions.
Cairo gate passes
execution unlocked
wraps execution integrity
Stone / STARK
integrity envelope
Execution integrity — independently proven
Stone prover wraps the entire execution trace independently of the SNARK lane. No trusted setup — security relies on collision-resistant hashing and DEEP-FRI, not a ceremony. In FULL_DUAL_PROVER mode, both proofs must clear. A compromised Groth16 trusted setup cannot forge a STARK.
STARK envelope
no trusted setup
settles to
Settlement
recursive · tri-chain
Recursive tri-chain settlement — three paths, all live
Path A: ObsqraFactRegistry on Madara L3 — fast, cheap, primary. Path B: L3 block validity proved recursively to Starknet L2 via SNOS — L3 inherits L2 security. Path C: Full L3→L2→L1 Ethereum bridge — confirmed live, l1_status=1, polls=61. Any contract calls is_valid(fact_hash) at any layer.
3 settlement paths
L3 · L2 · L1 all live
Output — one enforcement primitive
One receipt. Five systems. No trust required anywhere in the stack.
ML inference proven. Policy enforced. Execution verified. Settlement immutable. Each layer uses a different cryptographic system with different trust assumptions. Together they produce one receipt that holds even if any single layer is compromised.
97.5% verification rate
316 receipts accumulated
Recursive settlement

Three paths. Every layer of the stack.

Most ZK systems have one settlement path. StarkForge has three — all confirmed live. A proof registered on Madara L3 can propagate to Starknet L2 and settle to Ethereum L1. Recursively. Each layer inheriting the security of the one above it.

Live · Primary path
Path A — Madara L3
ObsqraFactRegistry

Proof facts registered on OBSQRA_PROOF_CHAIN — the dedicated Madara L3 appchain. Cheapest path. Fastest finality. The primary settlement target for every StarkForge proof.

Any contract: is_valid(fact_hash)
Staged · Integration
Path B — L3 → L2
SNOS recursive validity

L3 block validity proved to Starknet L2 via SNOS-style recursive proving. L3 proofs inherit the full security of the Starknet L2 consensus layer. Proof facts become L2-verifiable without re-proving.

L3 security → L2 security recursively
Live · Maximum finality
Path C — L3 → L2 → L1
Ethereum Sepolia bridge

Full tri-chain settlement path, confirmed live. Proof generated on L3, propagates to Starknet L2, bridges to Ethereum L1. Maximum finality — Ethereum's security model, all the way up.

0x6dbcc1a6…876db4 ↗ l1_status=1 · polls=61
Verification path — STARK lane
Herodotus Integrity Verifier

The STARK integrity lane verifies through the Herodotus Integrity verifier on Starknet — the same storage proof infrastructure that underlies Starknet's historical state verification. StarkForge's STARK proofs settle through this path independently of the SNARK lane. No trusted setup. No ceremony dependency.

The registry choice
ObsqraFactRegistry or Herodotus

Builders choose their verification path. Use ObsqraFactRegistry on OBSQRA_PROOF_CHAIN for primary L3 settlement with the full StarkForge context. Or verify directly through the Herodotus Integrity verifier for STARK proofs that need to inherit Starknet's native storage proof infrastructure. Both paths are live. Neither requires trusting a backend.

Recursive means compounding security. A proof on OBSQRA_PROOF_CHAIN (L3) propagates its validity upward — to Starknet L2, then to Ethereum L1. Every layer above adds its own security model without requiring re-proving. The L1 Ethereum bridge is confirmed live on Sepolia. The L3→L2 recursive path (SNOS) is in integration. The full tri-chain stack is the roadmap already in progress — not a whitepaper promise.
Signature execution mode

SNARK-in-STARK.

Two independent proofs. Two different trust models. Both must clear before capital moves. This is the execution architecture that makes the composition more than the sum of its parts.

The SNARK lane
Proves: the AI output was correct.

EZKL generates a Halo2 proof that a specific model produced a specific output from specific inputs. ModelBridge re-proves that in Groth16. Garaga verifies it on-chain via BN254 pairing in native Cairo. The Cairo policy gate only opens when this proof passes.

//Model identity committed via KZG — different model, proof fails
//Policy bounds hardcoded in circuit — cannot be overridden
//Requires trusted setup — the Powers of Tau ceremony
The STARK lane
Proves: the execution was faithful.

Stone prover wraps the entire execution trace — policy gates, constraint checks, settlement routing. The same prover that generates Starknet block validity proofs. Independent of the SNARK lane. No trusted setup. Security relies on collision-resistant hashing and DEEP-FRI.

//No trusted setup — cannot be compromised by a ceremony
//Proves execution integrity, not just output correctness
//Independent — SNARK compromise cannot forge a STARK
Why both? — Groth16 is fast and efficient but requires a trusted setup. If the Powers of Tau ceremony were compromised, a Groth16 proof could be forged. The STARK lane eliminates that as a single point of failure. In FULL_DUAL_PROVER mode, both proofs must clear independently. A compromised trusted setup cannot forge a STARK. A STARK prover failure cannot fake a Groth16 commitment. Two systems, two trust models, one receipt that holds across both.
Three execution modes — select based on value and trust requirements
Mode
What runs
Gas
When to use
EZKL_ONLY
EZKL inference. Off-chain verification only. No on-chain artifact.
0 L2 gas
Advisory checks, dev flows
EZKL_BRIDGE
EZKL → ModelBridge → Groth16 → Garaga on-chain. SNARK lane only.
~34M L2
Standard execution gating
FULL_DUAL_PROVER
SNARK lane + STARK integrity lane. Both run independently. Both must pass.
~77M L2
High-value execution
What we ship

Two primitives.
The stack composes from them.

// Primitive 01

The Prover

Five independent proving lanes behind one API. Call it, get a receipt you can verify on Voyager. Groth16, STARK, Noir HONK, native KZG, and hash fallback — each with a live transaction hash, each with a different trust model. No local prover. No configuration. One endpoint, the full composition.

Hosted Stone prover — 8 CPU / 32GB dedicated infra
Open Starknet RPC — self-hosted Juno at /rpc
Madara L3 settlement — OBSQRA_PROOF_CHAIN
L1 Ethereum bridge — confirmed on Sepolia
Proof sequencer status at /proofs/sequencer-status
/cloud →
// Primitive 02

ModelBridge

The open-source Circom circuit that converts EZKL/Halo2 ML inference proofs into Groth16 commitments Cairo policy contracts enforce as execution gates. The composition that doesn't exist elsewhere in the ecosystem.

Open-source — github.com/Obsqra-Labs/zkdefi/circuits
Binds model identity (KZG weight commitment) on-chain
Policy bounds hardcoded in circuit — can't be overridden
Replay protection — timestamp + ReceiptRegistry nullifier
ModelBridge ~2MB zkey · ModelBridgeHeavy ~8MB
// ModelBridge.circom — the binding signal input model_hash; // KZG weight commit signal input input_hash; // private inputs signal input bounds[N]; // policy constraints signal input timestamp; // freshness window signal output commitment; // → Garaga → Cairo gate
github.com/Obsqra-Labs/zkdefi/circuits →
Five proving lanes

Every lane. Every claim. Transaction hash.

These are not examples or simulations. Five live verifier lanes running on Starknet Sepolia. Every one has a transaction hash on Voyager. If you don't believe the claim, verify the proof.

Groth16 / Garaga
groth16_garaga
BN254 pairing in native Cairo. Primary zkML gate lane. Used in EZKL_BRIDGE and FULL_DUAL_PROVER. ~34M L2 gas.
STARK / Stone
stark_integrity
Stone execution integrity lane. No trusted setup. DEEP-FRI. Independent from SNARK. The second proof in FULL_DUAL_PROVER.
Noir HONK
noir_honk
Aztec's Noir proving language, HONK backend. Alternative path for Noir DSL applications.
Native KZG
native_kzg
kzg_mpcheck_v1 strict path. strict_binding=true enforced. Direct polynomial commitment verification.
Hash Fallback
hash_only
Poseidon hash verification. 0 L2 gas. Advisory checks and development flows where full proof weight isn't required.
0 gas · EZKL_ONLY mode
Built on StarkForge

zkde.fi — the proof it works.

The live application built entirely on StarkForge's two primitives. zkde.fi is what the composition produces when it's applied to private DeFi. 316 on-chain receipts. 14/14 core claims validated. Nothing here requires trusting us — verify it at zkde.fi/test.

Consumer Product — Built on StarkForge

zkde.fi

Private execution. Provable behavior.

Private swaps, lending, governance, prediction markets — the full DeFi surface, proof-gated. Your reputation is a ZK credential that gates what you can access without revealing the underlying data. The SNARK proves the AI output was correct. The STARK proves the execution was faithful. The receipt exists before the transaction does.

316
On-chain receipts
$75M+
TVL tracked
97.5%
Verification rate
14/14
Claims validated
SNARK-in-STARK execution loop
SNARK — Groth16 / GaragaAI output proven
Reputation — FICO-pack ZK credentialAccess gated
STARK — Stone / IntegrityExecution faithful
Receipt — ReceiptRegistryLoop closes
Composed from

What StarkForge composes.

StarkForge composes these. None of them alone is enough. Together they are. Credit where it's due — and the composition is open source.

Forward

What the composition unlocks.

The composable ZK execution layer is live on Starknet Sepolia. These are the surfaces it opens — now and as the stack matures.

Live today
Proof-gated DeFi
zkde.fi

Private execution. Reputation-gated access. The SNARK-in-STARK loop applied to capital movement. 316 on-chain receipts. Live.

ZK Reputation
Portable FICO-pack

Six Groth16 verifiers that produce portable ZK credentials. Solvency, risk tier, trader performance, strategy integrity — provable without revealing the underlying data.

Private governance
ZK voting + DAO

The private_vote circuit enables on-chain governance where votes are cryptographically valid but voter identity is never revealed. DAOConstraintManager live on Sepolia.

Coming next
Q2 2026
near-term
StarkForge SDK

One package. Drop proof-gating into any AI agent. ModelBridge + Garaga verification + L3 settlement, wrapped in a single SDK. The goal: proof-gated execution in under 10 lines of code.

Verified MCP Gateway

ModelBridge extended to the MCP layer. Every AI agent tool call produces a StarkForge receipt. Any agent — any model, any framework — gets cryptographic accountability for its actions.

Q3 2026
scaling
Reputation API

Protocols pay to consume ZK credentials at the gate. Credit eligibility, risk passport, solvency proof — cross-protocol reputation without raw data exposure. The FICO-pack as a product surface.

Mainnet deployment

Current stack is live on Starknet Sepolia + Ethereum Sepolia. Mainnet deployment follows audit and verifier network launch. Every claim on Sepolia carries forward.

Q4 2026+
long arc
GATE standard

AEGIS-1 — a shared execution language for constrained delegation, policy-aware execution, and proof artifact lifecycle across protocols. The composition as a cross-ecosystem interoperability standard.

Chain-agnostic expansion

StarkForge is built on Starknet but ModelBridge is chain-agnostic by design. Proof-gated execution for any AI agent on any chain that can verify a Groth16 commitment.

The composition is live.
Build on it.

API key. Five proving lanes. ModelBridge open source.
Every claim has a transaction hash. Testnet live. Mainnet coming.