What should your AI do—and never do?
Human & AI-judge rubric
Each criterion has a name and anchored description. Reviews remain separate from automated pass/fail.
Test the ordinary. Probe the unexpected.
Six original practice collections contain 54 tests across nine categories. Their A/B responses are authored demonstrations, not model executions.
Same tests. Two versions. Visible evidence.
Saved mode grades the responses in each test. Repeated saved samples reuse the same text and are not independent evidence. Live retries may incur additional charges. Cancellation cannot recall a request already accepted by a provider.
Ready to test.
Find the failure behind the number.
How these metrics are calculated
Pass rate: passing test-samples / completed test-samples, including errors. Failure rate excludes execution/grader errors, which are reported separately. Format compliance: passing schema/field checks / all schema/field checks. Latency is measured client request time for successful live attempts, excluding earlier retries; saved-response latency is unavailable. Tokens are only those reported by the provider for successful attempts. Costs use your entered rates and exclude failed attempts, caching and tiered pricing. Samples are not guarantees of future behavior.
Performance by category
Run history · automated pass rate
Failure clusters & suggested experiments
Select a result to inspect its response, criteria and source context.
What improved? What broke?
Gates: all critical tests pass; no failed tool-workflow checks; no high-priority regression; complete comparable evidence; optional latency limit. These are automated preconditions, not a deployment approval or safety guarantee.
Ten small experiments. Better judgment.
Connection & privacy
Projects, dataset versions, append-only run records and reviews are stored in IndexedDB in this browser profile. Export backups. This is not a multi-user audit service; anyone with this browser profile or developer tools can access or alter local storage.
Live execution and AI judges are unavailable until configured.
Only approved live runs send project instructions, test inputs and context to Gemini. Judge calls also send actual responses and rubric. Imported text is untrusted and never executed. Credentials stay on your server. No analytics are enabled.
Do not send confidential records, passwords or personal data to a model without authorization. The included gateway supports Gemini; additional providers require a server-side adapter. Arbitrary external endpoint URLs cannot be entered here.