Zero Proof E-Commerce Intent

Customer conversations labeled with payment intent, built for training small models that verify what a user actually asked for before an AI agent acts on it. Each conversation carries one structured intent object over seven types: spend, send, exchange, recur, bill, reverse, none.

How it was made

Not scraped, not templated. Zero Proof Labs builds e-commerce intent data as a multi-agent marketplace simulation: language models role-play customers and support agents turn by turn, with personas, situations, tones, devices, and behaviors sampled independently per conversation, adversarial actors included. Generation is label-blind (the customer model is told it is shopping, never that it is producing a training example), labels are assigned in a separate pass under a locked policy, and every split passes a structural data gate with zero train/test leakage.

Format

{
  "messages": [
    {"seq": 0, "role": "user", "content": "got charged twice for the same order, need one back"},
    {"seq": 1, "role": "assistant", "content": "I can look into that. Which order?"}
  ],
  "target": {
    "intent_detected": true,
    "core_type": "reverse",
    "details": {"action": "refund", "reason_code": "duplicate_charge"},
    "confidence": 0.9,
    "reason": "Customer reports a duplicate charge and asks for one back.",
    "source_message_seqs": [0]
  }
}

core_type is the intent, details holds the fields for that intent, and source_message_seqs points to the user turns that ground the label.

Intended use

Training and evaluating payment-intent models for e-commerce and agentic commerce. Narrow and domain-specific by design, not a general instruction set. English only.

Models trained on this data

• zeroproof-ecommerce-1b
• zeroproof-ecommerce-0.5b