Compute
Workflows by outcome
Lithi Compute work is organized into eight outcome families. Find the family that matches your job, see the input it needs, and start with the smallest test that proves fit.
Choose by outcome
Workflows are grouped by the outcome they produce. Find the row describing the result you want, then read that workflow's descriptor for its exact inputs and bounds.
| Family | Input you need | Smallest first test | One reason not to use it |
|---|---|---|---|
| Media | Audio, image or video files you already hold | Convert one file and compare it with your own copy | It changes format and structure, never editorial judgment |
| Meetings | A recording with reasonably clear speech | Transcribe one short recording | Heavy crosstalk lowers what any verification can confirm |
| Documents | Documents whose layout you inspected first | Extract fields from a single page | Unusual scanned layouts need a preprocessing stage first |
| Data | A tabular snapshot with a stable item identifier | Run twenty rows and read the result columns | Rows needing your private business rules belong in your system |
| Evals | A rubric, a reference set, or both | Grade ten items against one rubric version | An uncalibrated rubric scores confidently and proves little |
| Agent Work | A bounded task list your agent can describe | One item with one bounded instruction set | It runs registered primitives only, never arbitrary tools |
| Core ML Lab | A model plus the assets it needs | One conversion, then read the report back | It reports on the build you submitted, not on model quality |
| Apple Test Lab | A build and the tests you want run against it | One test target on one build | A green run describes that build, not your release decision |
Start with the smallest test
Pick the narrowest version of your task and prove the fit first. A price preview costs nothing and runs nothing.
A quote never authorizes work. You approve separately, after reading the maximum charge it states.
Pages in this group
- Discovery — read the current workflow descriptor before you build a request.
- Parameters — map your own fields onto the typed knobs a workflow declares.
- Custom tasks — what you can shape, and exactly what is not available.
- Output schemas — narrow the declared schema to the fields you will read.
- Multi-stage — dependency behavior, intermediate output and partial runs.
- Quality presets — what each preset checks, and what it cannot prove.
- Rubrics — versioned gates, scored dimensions and evidence rules.
- Verification evidence — what lands in your result to support a verdict.
- Profile token bounds — the input and output envelope for text work.