| Current position | Next useful action | Evidence needed to move on |
|---|---|---|
| We cannot state the failure or current cost | Observe the workflow and build a baseline ledger | Reproducible examples, accepted-task definition and cost boundary |
| We lack permission to use the data | Resolve rights and permitted processing | Written decision for the intended training, evaluation and hosting uses |
| We have examples but no trustworthy test | Build labels, rubric and protected holdout | Separation from training and independent acceptance ownership |
| Current prompting, retrieval or parsing is clearly weak | Improve and measure the relevant baseline | A fair comparison against the practical alternative |
| A stable gap remains and an owner can maintain the service | Commission a bounded feasibility test | Pre-agreed gates, deliverables and stop conditions |
| A candidate passes offline tests | Run staged operational validation | Load, security, fallback, rollback and human-review evidence |
| Economics or quality fails | Stop, rescope or improve the existing path | A documented failure cause and reconsideration trigger |
Do not interpret the rows as a requirement to train a model. The cheapest useful result can be evidence that a document service, retrieval repair, routing layer or current API already meets the requirement. RouteLLM gives a research basis for considering routing, while Extend’s extraction documentation illustrates a specialist product category worth testing for document workloads.