On this page
Five labels, not one score
Verification milestones
Detected
A dated source points to a candidate capability. It does not establish navigation or completion.
Reached
An executor reached a defined workflow milestone, but did not establish a correct final submission.
Pre-submit verified
A reviewed adapter captured and blocked the exact final payload. This is not merchant acceptance or live completion.
Sandbox completed
An authorized sandbox returned a success state. It is a separate cohort from a live-site result.
Live completed
A valid completion was observed with appropriate authority and provenance. It is not universal reliability.
Text equivalent: detected → reached → pre-submit verified → sandbox completed → live completed. Each arrow describes a stronger, different observation; it is not a promise that one state implies the next.
How rates are counted
Denominators stay visible
The objective denominator includes started valid attempts with success or legitimate task failure outcomes. Blocked attempts, policy stops, inconclusive outcomes and infrastructure errors are excluded from that denominator and shown separately. All-started counts include these operational outcomes. Under ten qualifying attempts, the map shows successes over attempts without a headline percentage; at ten or more it may show a rate and 95% Wilson interval.
Age matters
Freshness and comparability
Fresh is under seven days, aging is seven through thirty days, stale is beyond thirty days, and unknown means never actively tested. An older fallback never masquerades as current evidence.
Results are comparable only when the objective, task version, executor or agent, interface, authentication context, region, provenance, and time window align. Sandbox, live, deterministic, imported, and real-agent observations remain separate cohorts.
Where evidence comes from
Provenance and corrections
Independent tests, external benchmarks, merchant declarations, and customer-reported observations are labeled separately. Customer-reported observations are quarantined: they do not enter independent public denominators, default route ranking, or conversion analytics. Public evidence is a reviewed derivative with a content hash, expiry, and correction linkage; raw traces, personal data, credentials, and signed URLs remain private.
Payment cannot change rankings or suppress independent evidence. If a published field is wrong, use support@webterra.ai; corrections append a superseding record rather than rewriting history.
Consent and feedback
Customer reports are bounded and reviewable
An authenticated developer may opt in to exchange an exact route recommendation for a short-lived receipt and report one terminal outcome. A receipt is not permission to execute, and a report does not prove independent reliability. Report details are bounded, URLs are rejected, and private traces are never published.
You can withdraw feedback consent from the developer integration; it revokes unexpired receipts. Eligible detailed receipt and report records are deleted after 30 days while a minimized audit record is retained. For access, deletion, correction, or support requests, contact support@webterra.ai.
Use the evidence carefully
Search the task map, inspect the site map, or download the public dataset. Developers can use the free read-only API. None of these surfaces executes a third-party task.