Miatz playground

The Guardrail Builder

Write the check that would have caught it. Then test it.

Coming to the Build-LabBYOK AICross-Cuttinggen · dev
The Demystify signature

most AI-safety content tells learners guardrails matter; this makes them build one and watch it fail against a fluent-but-wrong sample — the gap between 'I wrote a check' and 'my check actually catches things' is the reveal

How it works

What goes in, what comes out

What it does

Learner picks a failure scenario from the war-room library and authors a guardrail — a schema check, a least-privilege tool scope, or a pass/fail rubric line — in a structured builder. The tool then fires a battery of good, bad, and adversarially plausible-wrong sample outputs at the learner's own guardrail and scores whether it actually catches what it claims to.

You bring

write or configure one guardrail/rubric line, then click 'stress-test it'

You get

a pass/fail readout per sample — caught the bad one, wrongly flagged the good one, or let fluent-but-wrong slip through — plus a robustness score

You control

guardrail type (schema/tool-scope/eval-rubric), difficulty scaling L1-L3, 'attack my own rubric' adversarial mode

Where it's used

FREE, top-of-funnel — /playgrounds gallery; the same mechanic is reused inside Agent Builder, MCP Playground, and Evals Runner labs at L1-L3

output-format guardrailtool-scope guardraileval-as-guardrailrubric designadversarial self-testing

Concepts click when you open the machinery.

Three labs are already live and free — the same hands-on style this playground brings to its module.