Define the changed boundary
API drift is a change in an integration’s observed contract or behavior that can invalidate its client’s assumptions. A removed response field is shape drift. A numeric amount changing magnitude can be a statistical signal of semantic drift even while the response remains valid JSON. A change is breaking relative to the consuming application: adding an unused field may be harmless, while removing a field used to authorize a write may be consequential. Detection identifies evidence of a difference; it does not automatically supply a correct migration.
Compare declared shape with observed behavior
A schema comparison can identify a missing required property or a changed type when a schema and a complete response are available. Statistical detection instead compares explicitly selected fields in comparable baseline and current windows. Those windows must represent the same operation, phase and cohort. Mixing currencies, endpoints or tenants can produce apparent drift that is really a change in the sample population. AGINDOW does not learn a trustworthy production baseline just because a client was wrapped.
Example: an amount changes magnitude
The quickstart supplies a labelled baseline of numeric values around 1 and a current window around 100 to the packaged statistical guest. Its configured Unit detector produces a finding; unchanged and insufficient-window controls do not establish successful detection. This is a controlled engine example, not evidence that a real vendor changed from dollars to cents. A magnitude difference is evidence to investigate, rather than permission to divide customer amounts by 100. If values keep the same distribution while their business meaning changes, statistical evidence alone can miss the change.
{"baseline":{"amount":1},"current":{"amount":100},"interpretation":"Investigate magnitude; do not infer a conversion"}Keep findings separate from protection
Observe mode records available findings and preserves the original call. A healthy engine with zero findings may mean that no detector was selected, that a window was insufficient or that no actionable signal was present. Check readiness, selection, scope and bypass separately. Configuring repair mode does not create representative windows and does not enable selective repair. In an agent workflow, label the tool and operation explicitly so the intended boundary is the one inspected.
Investigate before choosing a migration
Retain the permitted evidence with your application’s redaction and retention rules. Compare the vendor’s declared contract with independent application facts. A field rename may be eligible for a reviewed mechanical rule, while an ambiguous payment-method mapping needs a decision about meaning. Never fabricate a missing field or infer business correctness merely from a schema accepting the payload. Detection, a proposed patch and an applied verified repair are three distinct outcomes.
Evaluate with your own API reads
The browser playground starts direct HTTPS GET collection only when you choose to run it. Compare bounded baseline/current captures, actual HTTP status and supplied JSON Schema checks. Provider authorization and quotas remain your responsibility, and browser limits are not provider quota enforcement. The playground does not apply repairs or prove write protection. For SDK integration, use independently collected representative windows and preserve the original client’s timeout, credential and cancellation behavior.
Publication and evaluation
Published 28 October 2026; last updated 28 October 2026. AGINDOW documentation team. These explanations describe pre-alpha SDK behavior for local evaluation, not production accuracy or reliability benchmarks. Start with the runnable quickstart and evaluate your own boundary before widening scope.