liveBook · Manning

Target lifecycle

As explained in Designing AI Systems

A target lifecycle is the progression of an experiment target through draft, testing, active, and deprecated states based on evaluation outcomes.

A target lifecycle governs changes to prompts, model configurations, and retrieval configurations from draft through testing to active, then deprecates the replaced version. It keeps untested changes out of production while preserving an audit trail. section 8.2

A production-oriented lifecycle separates experiment metadata from artifact ownership:

  • Register a target by referencing its name and version in the owning Model Service or Data Service rather than duplicating artifact content. section 8.2
  • Run the target against an evaluation dataset and attach an EvaluationSummary containing overall, per-metric, and test-case results. section 8.1

Prompt-version promotion

Register a new prompt version in the Model Service, register the same name and version as a PROMPT experiment target, describe the intended change, and run evaluation against the candidate and current version. Compare quality scores before promoting the candidate. section 8.2

How Designing AI Systems covers it

How deep it goes
Defines states, metadata, evaluation results, and transitions, but defers artifact storage to owning services.
Examples
Uses a prompt-version change with a recorded motivation, evaluation run, comparison, and possible promotion.
What it assumes
Assumes owning artifact services and evaluation datasets, while defining the lifecycle metadata layered across them.
What you can do after
Operate controlled target promotion while retaining scores, rationale, version history, and deprecation state.
How it connects
Connects prompt versions, model configurations, retrieval configurations, evaluation datasets, and offline evaluation.

This explainer was generated by Manning's knowledge graph from the text of Designing AI Systems. The section links point to the parts of the book it draws on.