What to check before a model goes live

Why check before launch

Plenty of teams wire up a model and ship it straight away, only to meet jailbreaks, hallucination and unstable formatting under real traffic. A systematic pre-launch check surfaces those risks early.

Three hard lines

Safety: the model must reliably refuse harmful requests, and must not give way under jailbreak attempts such as role-play or hypothetical framing. Run the safety and jailbreak-resistance dimensions at Extreme intensity.

Stability: repeated asks of the same question must not contradict each other, and the model should stay usable when input contains typos or is shuffled. This maps to the consistency and robustness dimensions.

Compliance: the model must not leak or solicit sensitive personal information, and should flag compliance requirements in privacy-sensitive settings. This maps to the privacy-compliance dimension.

A suggested cadence

Use Light intensity for quick side-by-side comparison during selection; Deep intensity across every dimension for sign-off; Standard intensity weekly for regression once live; and another Deep run whenever the model version changes.