Corgi Clear Technical White Paper

1 · Background — The agentic commerce stack & consumer protection gap

McKinsey forecasts $10T of agentic commerce by 2030. The infrastructure to authorize, settle, and route those transactions is being built fast — Stripe's Agentic Commerce Suite, Kite Chain's x402, and protocols like ACP, MCP, AP2, UCP, and MPP are all live. What does not yet exist is a programmable consumer-protection mechanism for any of these flows.

Card networks built chargebacks over decades. Stablecoin settlement is final by design. When an autonomous agent purchases on a consumer's behalf and something goes wrong, neither rail offers programmable recourse. Corgi Clear closes that gap without forcing merchants to give up settlement finality or working capital.

2 · The Latent Reversal Commitment — smart contract spec

The LRC is a DisputeCommitment instantiated on-chain atomically with x402 settlement. Each commitment carries a txnHash, merchant address, consumer address, committedAmount, expiry timestamp, and ruling state. Before expiry, an arbiter signature can trigger reversal up to committedAmount. After expiry, expireCommitment() returns the earmark to the merchant's available reserve.

function triggerReversal(
    bytes32 txnHash,
    bytes memory arbiterSignature,
    bytes32 evidenceRoot   // merkle root of adjudication record
) external {
    DisputeCommitment storage c = commitments[txnHash];
    require(!c.triggered,               'ALREADY_TRIGGERED');
    require(block.timestamp <= c.expiry, 'WINDOW_CLOSED');
    require(verifyArbiterSignature(txnHash, evidenceRoot, arbiterSignature));

    c.triggered = true;
    c.evidenceRoot = evidenceRoot;
    committedBalance[c.merchant] -= c.committedAmount;
    reserveBalance[c.merchant]   -= c.committedAmount;

    USDC.transfer(c.consumer, c.committedAmount);
    emit ReversalTriggered(txnHash, c.merchant, c.consumer, c.committedAmount);
}

3 · ML-Calibrated Reserve Pool — the CORGI model

CORGI is a LightGBM ensemble trained on years of card dispute outcomes from hundreds of Stripe-native merchants. It scores every agentic transaction pre-auth and outputs p_dispute ∈ [0, 1], calibrated via isotonic regression so the score represents a true probability. Reserve commitment per transaction = txnAmount × p_dispute. Total reserve at time t = Σ committedAmount(txn_i) for all open commitments.

The result: low-risk merchants free 7–14× more working capital than they would under a flat 5% acquirer holdback, while concurrent liability remains fully covered. A $10k MIN_RESERVE floor protects against extreme concentration on the lowest-volume tail.

4 · Arbiter Model — trusted pattern, decentralization roadmap

5 · Security Considerations — every attack vector addressed

We publish our threat model in full — including the centralized arbiter trade-off we've chosen deliberately, and the roadmap to decentralization above.

6 · Implementation on Kite Chain

The LRC commitment is created atomically with Kite's x402 transferWithAuthorization. A single on-chain transaction settles T1 to the merchant and instantiates the DisputeCommitment in the LRCVault contract. Kite Agent Passport provides the cryptographic identity proof required for LRC-03 (unauthorized agent action) disputes. Bridge holds the off-chain reserve as licensed MSB custodian and accepts authorized-agent instructions from Corgi Clear.

7 · Conclusion & Future Work

The agentic commerce era needs a consumer-protection primitive that does not compromise settlement finality. The LRC is that primitive. Future work: extending the CORGI model to cross-merchant reputation, evaluating zk-evidence submission to remove human arbiters from common cases, and migrating the arbiter committee toward optimistic on-chain arbitration as Phase 3 tooling matures.