The Together listing explicitly combines data strategy, algorithm choice, controlled experiments, task-specific evaluation and compute sizing. Scale’s listing centres on benchmarks, failure modes and the connection between observed weaknesses and post-training interventions. These are useful procurement clues: require the supplier to show who actually performs each task.
Use the following proposed capability map for an internal team or an external partner. It is a staffing design, not a claim that every market employer uses these titles.
| Capability | Typical project responsibility | Evidence to request |
|---|---|---|
| Applied ML/post-training engineer | Data formatting, baseline comparison, training and ablations | A reproducible run with pinned inputs and a losing experiment explained |
| Evaluation engineer or applied scientist | Rubrics, holdouts, graders and uncertainty | An example where a misleading metric was corrected |
| Data/annotation lead | Provenance, labels, disagreement and coverage | Label guide, adjudication log and dataset lineage |
| Inference/platform engineer | Runtime, load, deployment, monitoring and rollback | Load-test traces and a documented incident rehearsal |
| Product/domain owner | Defines accepted work and error consequences | Signed metric dictionary and prioritised failure examples |
| Security/privacy owner | Access, retention, supplier processing and incident coordination | Data-flow diagram and responsibility schedule |
| Technical delivery lead | Scope, dependencies, milestones and handover | A plan that allows a no-build outcome |
Anthropic’s evaluation account and Hamel Husain’s field guide provide practical background for testing evaluation competence. Neither prescribes a staffing ratio.