How Drift verdicts work
Every pull request gets scored against your behavioral load map — a deterministic model, built from real sessions, of which elements your users actually depend on: how much interaction each carries, whether it sits in the path of goal completions, and how ritualized returning users' behavior around it is.
No AI guesses a verdict. Every number in a check comes from sessions, manifests, or replay runs.
The verdicts
| Verdict | Meaning |
|---|---|
| CLEAN | The PR touches nothing carrying meaningful user load. Most PRs should land here — a quiet check is a real result, not a miss. |
| ADVISORY | Worth a glance. Examples: a change touches a surface users rely on but the PR description doesn't mention it (unclaimed_surface), or returning users interact with the element ritually (stereotypy_risk — muscle memory breaks quietly). |
| WARN | A high-load element — one that appears in the path of goal completions — was removed, renamed, or plausibly disturbed (high_load_modified). |
| BLOCK | The deterministic tier: a recorded real-user journey failed when replayed against your preview deployment (path_broken). This is the only failing verdict, and it's overridable — you stay in charge of the merge. |
The two tiers of evidence
- Deterministic — path replay runs your users' actual recorded journeys against the PR's preview deploy. Requires a preview deployment; only this tier can produce BLOCK.
- Probabilistic — load-map signals computed from what the PR changes. These warn and advise; they never block.
Backtests (historical releases) use the probabilistic tier only — previews for shipped releases no longer exist — so backtest verdicts cap at WARN, and the report says so.
Coverage
Every check carries a coverage figure: the share of observed interactions that resolved to stable element identities. Un-stamped markup resolves by semantic anchors (role + accessible name + route), which covers a lot but not everything — icon-only buttons and heavily dynamic text can fall through. Installing the UXSense build stamp takes coverage to near-total; low coverage makes checks quieter, never noisier.
Graded against reality
Predictions aren't fire-and-forget. Every check is joined to the Release Impact Report that ships it, and signals that fail to predict measured behavior change get demoted. Checks are unmetered on every tier — no verdict is ever withheld for billing reasons.
← All help articles