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.”
Mission
How it works
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.
Generate
Synthetic cases covering the failure modes in your domain
Evaluate
A held-out set, balanced and labeled on the decision
Post-train
SFT on labeled behavior, then RL against the eval set
Serve
The judge runs inline at $0.18 per million tokens
Judge
Every action checked against what was asked
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
Request
A person states what they want
Plan
The agent selects an action
Act
The action reaches the tool
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
Capabilities
Evaluation sets
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
We built it for refunds and disputes, where a wrong decision costs a chargeback, and published every artifact. Rerun the evaluation against our numbers.
Accuracy against cost to serve
Near-frontier accuracy at about one hundredth of the cost.
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.