How much evaluation does this feature need?

A score would need weights across the six questions, and those weights would be arbitrary: possible harm to a person matters more than volume, and an aggregate number hides that ordering. The page returns one of four tiers, each carrying a cumulative list of practices with numbers that size them. Spot review is the floor; each higher tier adds its own items. The tier names what changes in the setup, and the answers that would change it show how sensitive the verdict is.

Seven of the nine rules that escalate the tier read the harm answer, more than any other question. When the worst plausible outcome is harm to a person, the verdict is at least measured rates against a bar, because a single bad output can have consequences that volume and reversibility cannot offset. Asking harm first means you see the most influential question first.

For a required question, the page shows no verdict until all four are answered. For an optional question (how often it runs and how often the setup changes), the page gives the best verdict it can and names what each unanswered question would have decided: volume decides whether a rare failure is frequent enough to matter at scale, and change cadence decides whether a fixed set needs to track each change.

The profiler maps eleven dimensions of your feature into six categories of written items (failure families, case classes and evidence expectations among them), all qualitative. This page asks six questions, four shared with the profiler, and returns a tier with a sized setup list. Two of its six questions are new to this page: worst plausible harm and how often the setup changes.

That page decides whether the feature should be an agent, a fixed workflow, or no model at all, and for an agent feature it comes first. Its four shape questions (whether the steps can be written down in advance, whether a person could write the rule, whether the systems it acts on expose callable interfaces, and how many steps the model chooses) do not appear here. The depth question on this page applies to either shape.

Each tier drops the items that only the heavier tier carries. Going from gated to measured rates drops the frozen gate rule, the judge validated against human labels, the sized zero-failure claim, the named signer and monitoring after launch. Going down to a fixed set drops the bar set in advance, the intervals, the paired check, the separate severe count and the grader check. Going down to a spot review drops the rerun on every change; outputs are still read by hand before launch and after each change.