Task definition and baseline evaluation
Describe correct, incomplete and unacceptable results. Test the current approach on permitted examples so later changes have a useful comparison.
Model improvement services
Choose the experiment around the behavior you need.
A scope should explain why a proposed change may help and how that will be checked.
Instructions and application workflow
Review prompts, examples, validation and human review. A clearer task or ordinary rule may resolve an issue without changing model parameters.
Relevant-document retrieval
Assess how approved information reaches the model when the answer depends on it. Include source freshness and the user's permission to view the material.
Dataset preparation and adaptation
Where adaptation is appropriate, review the examples, expected outputs, use permissions and quality checks. Keep separate evaluation material and record the experiment configuration.
Specialized training feasibility
Assess the task, available data, compute and specialist requirements before promising a training program. A feasibility review should have an agreed deliverable and stopping point.
Deployment and ongoing checks
Plan the serving environment, application interface, operating cost and release recovery. Repeat quality and capacity checks before changing a deployed model.
Need private infrastructure?
Review private AI deployment for the data and operating boundary. Use the adaptation planning guide to prepare the task.