New/zeroproof-ecommerce-1b, an intent judge for agentic commerce. Read the release →

Proof of intent

The trust layer
for AI agents

Let agents transact on your customers’ behalf, with a verifiable record of every action. Runs in your transaction path. Works with the agents and tools you already use.

Customer

“There’s an $81.40 charge from Amazon I never made, I want it refunded.”

Intent

zeroproof-ecommerce-1b

Typereverse · refund
Amount$81.40
Confidence0.85

Proof

Verified

Signed by the agent and the tool server.

sha256:7f3a…1b2

A refund request is read as an intent, and the resulting action is attested.

Everyone can prove what an agent did. Nobody can prove it was what you asked for.

Intent

Read what the customer actually asked for, before the agent acts.

Proof

Attest what the agent did, signed by the agent and the tool server.

The loop

Every action we catch trains the next judge. Models get replaced; the loop is what we keep.

01

Research

Language models are already capable of transacting on a person’s behalf; however, financial errors are where tolerance for AI mistakes ends.

The barrier is not capability but verification.

Delegation becomes safe when three questions can be answered about any agent action:

  1. 01What is the agent about to do?
  2. 02Is it aligned with what the user intended?
  3. 03Did the agent actually execute it?

Our research answers the first two before execution, by reading intent from the conversation, and the third after it, with a cryptographic proof. The result is a trust layer in which every answer is a signed, independently checkable claim.

02

Our Work

Intent Detection

We train small models that recover a customer’s true intent from conversation, before an agent acts on it.

One transaction through the trust layer: intent read before execution, a signed proof issued after, and a mismatched action stopped before it reaches the payment rails.

Behavioral Attestation

We build the cryptographic record of what an AI system did and produced. Each action carries a zkTLS proof signed by both the agent and the tool server.

The same primitive applied to content: each hop signs what it produced, the publisher verifies before publishing, and the disclosure the AI Act requires is carried as verifiable evidence.

03

Models and results

Two models fine-tuned for e-commerce payment intent: rank-16 QLoRA adapters over gemma-3-1b-it and qwen2.5-0.5b-instruct, served from a single L4 behind an OpenAI-compatible vLLM endpoint. Weights, training data, and the evaluation set are on Hugging Face.

Accuracy vs cost to serve

Near-frontier accuracy at about 1/100th the cost.

1007550250$0.10$1$10$ per 1M output tokens, log scalegemma-1bGPT-5Sonnet 5Opus 4.8Zzeroproof-ecommerce-1b$0.18 per 1M

Macro intent-type accuracy (%), identical balanced held-out rows for all models. Frontier cost at published list prices; ZeroProof cost from measured throughput on a single L4.

The same data trains zeroproof-ecommerce-0.5b to within five points at half the size.