Better boundaries.
Built together.
A shared laboratory for prompt injection policies. Explore the rules, inspect the evidence, and propose the next improvement.
A shared core. Different contexts.
Profiles for distinct operations, built from explicit rules. Review the exact policy before using it.
Loading policy library…
Explore the shared policy components ↘
These describe the existing workbench rules. Component selection must be approved by the application owner; inspected content cannot grant itself an exception.
Show the work.
Download the full bundle: policy, cases, expected values, observed outputs and run settings.
Shared results come from saved hosted runs. The service preserves the frozen policy, every planned case and every recorded response. This is evidence of that run, not a safety certification. New runs include independently verifiable runner signatures; older runs remain unsigned.
A contribution starts
with a question.
Change a policy through a GitHub pull request. Attach reproducible evidence, preserve failures, and let reviewers inspect the tradeoffs.
- Propose a changeFork a profile or improve a component in GitHub.
- Test a frozen suiteRun in the workspace or locally with your own Jev key. Include baseline and candidate evidence.
- Share the whole pictureShare a saved hosted run after inspecting it. External results belong in the PR for review; they cannot be uploaded as public benchmark evidence. Public reading needs no ChatGPT account.
Evidence the runner recorded.
Select your hosted policy evaluation. The service assembles its complete saved evidence; this form cannot supply or replace model answers. Nothing is rerun or charged.
Inspect the saved contribution before explicitly making it public. External files cannot receive hosted provenance or a verified badge.