Docs · Features

Reports

Reports aggregates the cost, token, volume, QA and post-edit numbers already recorded on your runs and hands them back grouped the way you ask, on screen or as a CSV. It only reads: a report never launches a run, never re-translates, and never moves money.

What it does#

You pick exactly one report kind. Cost always shows AI model spend as its own column, and needs the manager role or higher on a remote or web surface; grouped by workflow, and only for an admin or operator, it adds vendor human spend as a second column, so the AI and human figures are never combined into one number. Tokens covers input, output, and cache read and write counts. Volume covers jobs, segments, words, and characters. QA is a dimension-by-severity cross-tab of MQM LQA findings with per-card scores, including both machine-scored and reviewer-authored findings. Post-edit covers edit distance and post-editing effort: segments edited, episodes, character and word distance, percent changed, and edit time.

Rows are grouped by customer, project, provider, engine, date, workflow, locale, or vendor for the operational kinds; QA adds a dimension axis for the cross-tab, and QA and post-edit both add an actor axis, which needs the manager role or higher on a remote or web surface because it ranks named people, and on post-edit specifically buckets by an editor's email. An axis a kind does not support degrades to a single total with a caveat naming it. Every report is scoped to what the caller may see first, then narrowed by any filters, so a report can never widen visibility past the caller's own access.

Getting started#

List the report kinds to see the columns each one renders.
Run one, grouped, adding filters (a customer, a project, a provider, a date window) to narrow the scope.
Export the same table as CSV when a spreadsheet or billing pipeline needs it.

report kinds
report run cost --group-by customer
report run cost --group-by project -o costs.csv

On the web, the same report builder lives on the Reports page, with a Download CSV link on any run.

Works with#

Engines: model spend on every run is what the cost and tokens reports aggregate, reconciled with the same pricing the engine ledger uses.
Vendors: on the workflow axis, for an admin or operator, vendor post-edit spend is reported alongside model spend as its own column, the AI-versus-human split, since Reports reads vendor spend as the consumer of that data.
Documents: volume and token counts roll up from the jobs that translated documents and keys, so a report is a grouped view over the runs the document pipeline produced.

Current limits#

Current limits

The QA card count only counts cards that have findings; a clean, zero-finding card writes no rows. That means the average card score is not an average across everything scored, and because a re-score can add or remove findings, the count can shift over time with no new work having happened.

A reviewer-authored finding is counted in the total findings and severity columns, but excluded from the card count and average score, since it is a hand-logged note rather than a scored card, and it never changes a workflow's automated gate. A withdrawn finding is likewise excluded from every QA and vendor report bucket, though it is retained in the underlying data stream, so a report and a data-warehouse export can legitimately differ by exactly the withdrawn rows.

Post-edit capture is not uniform across every surface: a full document save is captured everywhere, but a save from the segmented editor is currently a web-only signal, so a segmented document edited only through a headless install produces no post-edit data for that path.

Reports reads other features' data purely as a consumer. Turning Reports off never touches the underlying data, and turning off one of those other features only removes its own rows from a report, not the report itself.