MerchantGuard

MerchantGuard is an Autonomous Dispute and Margin Recovery Copilot built for Track 3: Reinvent Commerce. It addresses the multi-billion-dollar problem of e-commerce chargebacks and friendly fraud, where raw LLMs fail due to hallucinated claims, prompt injection, and ignorance of unit dispute economics. Our architecture enforces strict Rule-First Governance ("Zero LLM Decisions"): 1. Deterministic Policy Gate (Rules R-001 to R-006): Hard-coded business logic verifies return windows, objective delivery proof (GPS geo-match & signature), and case completeness. Procedural disputes (e.g., late filing) are resolved deterministically with 0 LLM tokens spent. 2. Calibrated Margin Engine: A Logistic Regression model predicts Win Probability P(Win) and evaluates Expected Value: EV = P(Win)*(Order+Margin) - (1-P(Win))*Dispute_Fee. This prevents Economic Fee Traps where dispute fees exceed recoverable margin. 3. Isolated Evidence Synthesis: Google DeepMind's open-weights Gemma-4-31B-IT (self-dploy/Gemma-4-31B-IT) running on AMD Cloud Infrastructure via the AMD Radeon Cloud Token Factory synthesizes arbitration rebuttal briefs strictly grounded in verified evidence IDs ([F1], [F2]). 4. Cryptographic SHA-256 Ledger: Every decision is chained into a tamper-evident audit ledger verifiable live from the UI. 4 Core Scenarios: - Clean Win (CASE-1001): GPS + signature verified -> positive EV (+$226.57) -> CONTEST with cited brief. - Ambiguous (CASE-1002): Unattended delivery -> safe human escalation (NEEDS_HUMAN). - Window Bypass (CASE-1003): Filed after 21 days (>14d window) -> instant 0-token contest. - Economic Trap (CASE-1004): $15 fee on $3.80 margin -> auto-recommends ACCEPT_REFUND. Empirical Proof & Security: Passes 32/32 unit tests, zero leaked credentials, live telemetry with 3.1s-15.6s latency on AMD Token Factory, and runs in a hardened non-root container deployed live on Railway.