Choose a useful question
Use Insights → Analytics for throughput, quality, workload, usage and delivery reporting. Select an explicit time window and grouping, then check the population and observation time before comparing percentages.
Distinct documents, source versions, bulk work items and provider attempts are different units. Admission-date and outcome-date cohorts answer different questions. Missing telemetry is unknown rather than zero.
Use Assurance
Production brings together operational checks such as exceptions, QC outcomes and queue age using their stated populations. Personal monitoring targets change the view; they do not activate policies, processing or notifications.
Codebase shows existing schema versions, action-policy versions and permitted benchmark/readiness evidence. It links to the actual editing/publishing workspaces and is administrator-restricted. It does not scan a customer Git repository.
Pipeline shows recorded stage observations, failures and timings. Optional/unobserved stages are not automatically failures or successes. A redaction proposal is not a delivered redacted file, and the view does not represent external customer CI/CD checks.
Read the current processing and review counts alongside operational checks. The example flags an exception and a QC failure rate above its configured target. Production, Codebase and Pipeline show different kinds of assurance evidence.
Inspect usage and billing
Where authorised, open a job’s Usage & billing report for the whole-job snapshot. It is independent of document pagination and can include documents, known pages, attempts, retries, worker time, ledger charges/refunds and attributed provider tokens.
Known-page coverage must accompany page counts. A multipage TIFF file count is not its page count. Missing provider telemetry cannot be reconstructed as genuine usage; reserved quantities must not be added to settled charges.
Download the dated billing CSV for reconciliation. It is not a customer invoice. Customer tariffs and billing-period close are separate commercial configuration.
Avoid misleading conclusions
P95 stage timings cannot be added together to claim an end-to-end P95. Estimated completion is not a guaranteed SLA. A high validation pass rate does not establish that every document is accurate, approved or delivered. Compare metrics only when their scope and coverage match.
Compare like-for-like measurements
Set population and window before comparing values. Distinct documents, source versions and provider attempts are different units. Missing observations are unknown, not automatically zero. Loaded queue pages are not cohort-wide reporting. Use matching periods and coverage when comparing trends and inspect the underlying record before interpreting a change as an operational problem.