Authoring and review
Follow the Semogram authoring loop from intent and proposed changes to validation, asset approval, separate execution and evidence-based review.
AI-assisted authoring uses a reviewable artifact or change set. Describe the task, let the assistant clarify context, inspect the proposed revision, and review validation. Approve its save, create, or apply action when it is right. Manual forms are also available for supported workflows.
Approval saves or applies the asset described by the tool. It does not imply that a pipeline has run, a query is published, or a forecast has been produced. Those lifecycle actions are separate.
Pipeline dry-run checks configuration without executing data operations. Connection checks, query execution, forecast runs, and governed writes each have their own behavior; inspect what a control actually does before using it.
After execution, inspect evidence and diagnostics, review artifacts or assertions where appropriate, and record prediction outcomes when observable. Corrections remain inspectable records. Improving later results requires the relevant data, definitions, queries, guidance, or model behavior to use them; feedback alone is not a guarantee of automatic learning or reversible effects.
Continue with Pipelines, Forecasts, or Runs.
Keep each decision explicit
| Stage | Result to verify |
|---|---|
| Describe | The task, scope and intended output are clear |
| Review | The exact proposed revision matches that intent |
| Save or apply | The approved asset change is recorded |
| Execute | A separate run or query actually produced a result |
| Evaluate | Evidence and observed outcomes support the conclusion |
If a check fails, return to the layer responsible for it. Editing instructions, fixing a connection and recording an outcome solve different problems; keep their records separate so the next reviewer can follow what happened.
A proposal does not match the request
Correct the context in the authoring conversation and inspect the new artifact revision before approving. Keep the proposal/run IDs and exact error text for Getting help.