What API drift means
API drift is a difference between the behavior your application expects and the behavior an API now exhibits. Shape drift changes fields or types. Semantic drift changes meaning or value patterns even when the payload still satisfies a schema. Unknown means the available evidence does not support a more specific classification. These classes are separate from detector names such as Unit.
What the statistical guest detects
The embedded guest evaluates configured fields for numeric magnitude (Unit), enumeration, completeness, type and charset/format changes. Not every detector is eligible for every window. A finding says these configured windows differ under the detector’s rules; it does not establish a vendor migration’s intent or authorize repair. The quickstart’s Unit case uses 128 repeated synthetic documents per window to make the expected difference reproducible.
Choose comparable windows
A baseline is the earlier reference window; live is the current comparison window. Supply arrays of complete raw JSON documents, not individual extracted numbers. Select independent calls for the same tenant, connection, vendor, operation and segment. Match request/response side and field path. Mixing currencies, customers or API operations can create differences unrelated to API drift. The fixed quickstart windows demonstrate configuration, not production calibration.
Configure the selection
At SDK construction, selection supplies scope and fields. Each field has side (request or response), path, baseline, live and enumeration. Use a JSON Pointer such as /value for the selected field. The selection operation must match the guarded tool name. This selection is frozen for the engine instance; configure changes call policy, not the engine’s windows. Construct a new instance for a new reviewed selection. These SDKs do not automatically fetch baselines or refresh them from a dashboard.
Request and response inspection
Request inspection occurs before the callable; response inspection follows its result or error. A request finding uses the configured request-side windows, and response findings use response-side evidence. Keep these datasets separate. A response halt can follow a completed vendor write and cannot undo it. Whole malformed bodies, truncated inspection copies and transport failures are different evidence conditions.
When no finding is inconclusive
A default instance with no selection has no findings. Missing fields, ineligible detectors, insufficient samples, scope mismatch or exhausted inspection work can prevent a useful verdict. Different detector families have different eligibility requirements; 128 is the fixture window size, not a promise that every production dataset of that size is sufficient. The quickstart verifies that four samples do not produce its Unit finding. Check bypass and readiness separately before interpreting silence.