A team can build every guardrail in this course — tiered defences, least privilege, approval gates, audit logging — and still ship a regression to production, because none of those controls answer the question a release process exists to answer: is this specific new version, today, actually safe and good enough to replace the version currently serving real users? Guardrails protect a running system from bad inputs and bad actions at runtime; a release gate protects users from a bad version ever starting to run in the first place, and the two are complementary controls that fail differently when either is missing.
Teams that skip a real release gate tend to discover regressions the same way: a spot-check by an engineer who happened to try the new version, or worse, a support ticket from a real user after the bad version was already serving traffic. A model card and an eval-gated release process exist to replace that spot-check with a documented, repeatable decision — what this version is known to do well, what it is known to do poorly, and what score it has to clear on which evaluation before it is allowed to reach a real user.
This lesson covers what a model card actually is and is not, how to build a release gate that blocks a regression before it ships rather than reporting one after, and where the two connect to the guardrails and evaluation harness built earlier in this course.
Analogy🏏Cricket
🏏 Think of it like cricket: a fast bowler is not cleared to play a Test match on a physio's gut feeling that he 'looks fine in the nets.' A documented fitness protocol exists precisely because 'looks fine' has broken down before — a bowler passes a casual look, breaks down on day two of a Test, and the team is left a bowler short for the rest of the match. The actual protocol used by national boards is specific and measured: a graded fitness test with defined thresholds for bowling workload, a documented injury history the medical staff reviews before selection, and a formal sign-off that has to be given before a name goes on the team sheet — not an informal impression from watching one net session. Jasprit Bumrah's returns from back and other injuries have specifically gone through staged, documented fitness clearances rather than a single practice-session judgment, precisely because the cost of getting it wrong mid-match is so much higher than the cost of a slightly slower, more careful clearance process. Just as a bowler needs a documented fitness protocol with defined thresholds before he is cleared to play, not just a coach's impression from the nets, a new model version needs a documented evaluation with defined score thresholds before it is cleared to serve real traffic, not just an engineer's impression that it 'seems to work.' The insight is that the entire cost of a bad release, like the entire cost of a bowling breakdown, lands after the moment nobody checked carefully enough, so the checking has to happen before, not after.
🏏 Showing the Cricket analogy — a Cricket version isn’t available for this concept yet.