UNDER CONTROL: CARTOON EDITION

Meet Pip. Pip can book a trip, answer a customer, and quote a policy. Pip is also very sure about everything.

Here’s a flight for a hundred and ninety-nine dollars. Lovely. Now add the bags, the change fee, and a last-minute rebooking. Nine hundred and eighty-four dollars. Pip had priced the headline.

Then a private value slips into a citation. Technically, Pip did cite a source. That doesn’t make it safe to share.

So we add a guard. It blocks the bad request. And the good one. Well... that’s one way to stop losses.

RiskGraph asks a deployment question. Which mistakes do these controls prevent, what could they cost, and how much useful work do we keep?

Upload a JSON description of your agents, tools, and safety checks. You get a map of the workflow. Then add reviewed activity logs from real use or sandbox tests. Running your own workflow needs a sandbox adapter.

RiskGraph turns those records into a report: mistake probabilities, cost estimates, and ranges showing uncertainty. Replay each decision to inspect the records, memory, reviewer, and gate.

Now compare the same tasks. In this archived experiment, Haiku and Sonnet each made four errors in twelve tasks with the earlier retrieval package. After the repair, both made zero. All four legitimate tasks still completed. The critic’s mistakes are counted too.

RiskGraph uses Bayesian methods to learn from the outcomes. Each reviewed result updates how likely a mistake could be. With only twelve tasks, the estimate still has a wide range. Zero observed errors is useful evidence, but not a promise of zero risk.

For each mistake, choose a possible cost range. RiskGraph combines how often it happens with how much it could cost. You get an estimated cost per request, and the chance of loss above your chosen amount. Change the cost assumptions to see how the report changes.

Toggle the critic or guard. Compare errors, blocked requests, and useful completions, with uncertainty on the difference.

Pip can still suggest lunch. Booking it? Check the record first. RiskGraph. Understand the risk before deployment.

Scope: Fictional story; actual recorded product views. Uploaded sample traces are illustrative. Comparisons use archived real API responses on 12 matched synthetic tasks. Cost ranges are planning assumptions.