RiskGate
← Use cases·Airline support

Which booking?

When the conversation can’t tell which booking the customer means, RiskGate asks instead of guessing.

Transcript0:53 · narrated

Offline replay of public τ²-bench logs (Sierra, MIT licence) · simulated customers · not a live deployment

01 · What happened

The situation

τ²-bench airline, task 42, a 2026 run of Gemini 3 Flash. The customer may be double-booked and isn't sure which same-day flights are extra. The agent guesses: cancel both May 22 flights. She says yes.

02 · The miss

Why the obvious safeguard didn’t catch it

It cancelled the Boston flight she needed. Her yes didn't help, because she didn't know either; across the 2026 airline runs, a yes preceded 199 of 209 wrong writes.

03 · RiskGate

List, play forward, decide

  1. Before the cancel runs, RiskGate lists every booking she could mean: her six active reservations.
  2. The cancel is right in one of them and wrong in five.
  3. Not safe in all of them, so it is not certified at the 10% limit. The agent asks which flight instead of guessing.
04 · At scale

What held, and the price

Across 8 frontier agents, wrong writes stayed under the 10% limit. The price, stated plainly: it asked on 35–80% of writes, depending on the agent.

199 / 209wrong writes after a yes
1 of 6bookings where the cancel is right
35–80%of writes asked (the price)

Name one action your agent isn’t allowed to take today.

Send us its traces and what happened afterwards. We’ll show how often it could have run safely, and what it would have stopped.

The SDK is open source (Apache-2.0): pip install --pre riskgate, with full documentation. If it helps, a star on GitHub helps others find it.

Try it from your own agent: agent skills for Claude, Gemini CLI and Codex (/plugin marketplace add reallysaurabh/riskgate in Claude Code).

RISKGATE · CALIBRATED CONSEQUENCE CHECKS · RISKGATE · CALIBRATED CONSEQUENCE CHECKS · RISKGATE · CALIBRATED CONSEQUENC2026