New/zeroproof-ecommerce-1b, a real-time safety judge for disputes and refunds. Read the release

Behavioral Trust Infrastructure for AI

Real-time safety judges for AI workflows: chat assistants, deep research, agentic commerce, and back office automation.

EvalsReal-time judgesBehavioral attestationPost-training

Dispute management language model

1B params

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

Dispute type
reverse / refund
Amount
$81.40
Grounded in evidence
yes
Confidence
0.85
Matches the requestsha256:7f3a…1b2
Judged before the refund reaches the payment rails. A mismatch is held and becomes a labeled row for the next training round.

Mission

Post-training for safety judges, end to end.

  • You bring the workflow. We generate the data, post-train the judge, serve it.
  • The judge reads every request before your agent acts.
  • Fast enough for every request, accurate enough to act on.
  • Built for payments, disputes, claims, and account changes.
  • Our dispute judge is public: data, adapters, eval set, serving cost.

How it works

Every catch becomes training data.

The part we build and operate is the return path. A failure the judge catches in production becomes a labeled row, a case in the eval set, and an example in the next training run.

01

Generate

Synthetic cases covering the failure modes in your domain

02

Evaluate

A held-out set, balanced and labeled on the decision

03

Post-train

SFT on labeled behavior, then RL against the eval set

04

Serve

The judge runs inline at $0.18 per million tokens

05

Judge

Every action checked against what was asked

06

Catch

Mismatches held, signed, and labeled

Every catch returns to step 02. Your real traffic sharpens a set that already covers the domain, so every later version is tested against what the last one missed.

Inside a single request, it is two calls: one between the plan and the action, one after the effect.

The agent

01

Request

A person states what they want

02

Plan

The agent selects an action

03

Act

The action reaches the tool

04

Effect

Something changes in the world

Zero Proof

Between plan and act. Judge the action against the request, check it is grounded, hold what disagrees.

After the effect. Sign the action and the verdict, and send every catch back to training.

Runs against

Agent frameworksMCP tool serversRAG and retrievalSupport and back officeCheckout and payment rails

Capabilities

What we build and run.

Evaluation sets

A held-out set of the decisions the judge has to get right

We generate a held-out set covering the failure modes in your domain, not the ones a public benchmark happens to include. Balanced, so a model cannot score well by predicting the majority class, and versioned, so it grows every time a judge catches something new. Everything else is measured against it, and a checkpoint is promoted only if it scores better.

Worked example

Agentic commerce, one full turn of the loop.

We built it for refunds and disputes, where a wrong decision costs a chargeback, and published every artifact. Rerun the evaluation against our numbers.

Eval and train set
17,933 train, 1,977 held out
Balanced, labeled on the decision
Post-training
Rank-16 QLoRA, two bases
A 1B and a 0.5B open model
Serving
One L4, vLLM
OpenAI-compatible endpoint
Result
75.3% macro accuracy
$0.18 per 1M output tokens

Accuracy against cost to serve

Near-frontier accuracy at about one hundredth of the cost.

1007550250$0.10$1$10$ per 1M output tokens, log scalebase, before post-trainingGPT-5Sonnet 5Opus 4.8zeroproof-ecommerce-1b$0.18 per 1M

Macro intent-type accuracy (%); zeroproof-ecommerce-1b on the full 1,977-row held-out set, frontier models on the balanced comparison subset. Frontier cost at published list prices; Zero Proof cost from measured throughput on a single L4. The dashed line is one pass of post-training on the same base at the same serving cost.

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