How it works

The Latent Reversal Commitment.

Every agentic transaction programs a conditional reversal on-chain at the moment of settlement. The merchant receives full payment immediately. The consumer gets a 90-day protection window. Nobody has to choose between finality and recourse.

Settlement path (T1)

Always executes. Instant, unconditional finality.

  1. 1Agent submits payment request

    AI agent (ChatGPT, Gemini, Claude, Copilot) initiates purchase via x402 / ACP / MPP. Kite Agent Passport provides cryptographic proof of user authorization scope.

    Pre-authorization
  2. 2CORGI scores the transaction

    LightGBM model computes p_dispute in real-time across 40+ features: agent identity signals, behavioral patterns, merchant dispute history, SPT scope.

    ~50ms
  3. 3LRC commitment created on-chain

    Atomically with settlement: a DisputeCommitment is instantiated on Kite Chain. committedAmount = txnAmount × p_dispute is earmarked from the merchant's reserve pool.

    Atomic with T1
  4. 4Full payment to merchant — final

    USDC transfers to the merchant wallet. Settlement is unconditional and irreversible. Merchant can use or withdraw funds immediately.

    Instant finality

Dispute path (T2)

Latent. Conditional. On-chain.

  1. A90-day window opens

    The LRC commitment lives on-chain for block.timestamp + 7,776,000 seconds. Consumer can file a dispute at any point within this window via the Corgi Clear portal.

  2. BDispute adjudication

    Merchant has 20 days to submit evidence. Corgi Clear ops team reviews under the LRC reason code framework (LRC-01 through LRC-04). Kite identity proofs queried for LRC-03.

  3. CArbiter ruling

    Corgi Clear arbiter (2-of-3 multisig) issues ruling. If upheld: triggerReversal() executed on-chain with evidenceRoot merkle commitment. USDC flows from merchant reserve to consumer.

  4. DNo dispute? Reserve auto-releases

    At expiry, expireCommitment() releases the earmarked amount back to the merchant's available reserve. FIFO across all open commitments.

The CORGI scoring model

Dispute probability, computed at authorization. LightGBM ensemble trained on years of card dispute outcomes across hundreds of Stripe-native merchants.

Agent identity signals

Kite Agent Passport tier and age, SPT merchant scope and amount ceiling, delegation depth, session duration vs. authorization window.

Transaction signals

Amount vs. merchant P95 distribution, agent velocity, shipping address novelty, cart composition vs. merchant norms.

Merchant history

Rolling 90d dispute rate, dispute reason code distribution, days-to-dispute average for this MCC, prior LRC-03 rate.

Calibration & thresholds

Isotonic regression so p_dispute is a true probability. Merchant-configurable precision floor, monthly recalibration on 90d outcome window.

Reserve commitment per transaction
committedAmount = txnAmount × p_dispute(txn)
p_dispute ∈ [0, 1] · calibrated via isotonic regression

Capital efficiency vs. flat reserves

ML-calibrated reserves are 7–14× more efficient than 5% acquirer holdbacks for low-risk merchants.

Merchant profileVolume / moDispute rateFlat 5% reserveCorgi Clear reserveCapital freed
Low-risk e-commerce$500,0000.2%$25,000$1.8k – $3.6k~$22,000
Mid-risk e-commerce$500,0000.8%$25,000$7.2k – $14.4k~$15,000
High-risk e-commerce$500,0002.0%$25,000$18k – $36k~$3,000+
Floor (all merchants)AnyAnyVariable$10,000 minimum

90-day trailing window. Reserve released FIFO as commitments expire.

LRC reason codes

The first dispute taxonomy native to agentic commerce.

LRC-01
Item Not Received
Order confirmed but no delivery. Carrier confirmation + signature required from merchant.
LRC-02
Item Not As Described
Product materially different from the listing the agent used to initiate purchase.
LRC-03Agent-native
Unauthorized Agent Action
Agent exceeded its delegated scope. Determined cryptographically via Kite Agent Passport.
LRC-04
Duplicate Charge
Two settlement txnHashes for the same order. Distinct order IDs required from merchant.

Where Corgi Clear sits in the MPP stack

Stripe's Machine Payments Protocol gives agents preventive guardrails. Corgi Clear adds the latent recourse layer those guardrails can't provide.

MPP — preventive layer
Spending controls, enforced at protocol level
  • • Per-transaction & daily/weekly budget caps
  • • Merchant allowlist + MCC restrictions
  • • Velocity limits and time-window rules
  • • USDC settlement on Tempo in <1s

Stops payments that violate the envelope. Cannot reason about whether a permitted transaction will be disputed — or unwind one once Tempo finalizes.

Corgi Clear — intelligence + recourse
Risk scoring before, on-chain commitment after
  • • p_dispute scoring on every MPP-permitted txn
  • • LRC commitment created atomically with Tempo settlement
  • • 90-day on-chain consumer protection window
  • • LRC-03 disputes resolved via Kite Agent Passport scope proofs

Because Tempo finality is irreversible, consumer recourse must be programmed at settlement — not bolted on after.

Prompt-injection attacks that redirect agents to unauthorized merchants are blocked by MPP allowlists. Disputes from permitted merchants — wrong item, non-delivery, scope violation — fall to the LRC.