MLDLC-KIT

The life cycle

A model's life runs from an idea to production and back again whenever something changes. mldlc models it on ADLC (author, discover, scaffold, deploy, test, observe) and makes explicit the two artifacts ADLC leaves implicit: intent.md and spec.md.

The gates, who decides each one, and how a change re-enters

Who does what

Part Role
Skills The procedure and the judgement for each stage: what to ask, what to write, what quality bar to apply. They guide an agent or a person.
mldlc The deterministic gates: document checks, evidence from MLflow, the approval ledger, promotion and rollback. A skill cannot declare a gate passed.
People Approve G0 (intent), G1 (spec) and G5 (release), and set the acceptance thresholds.

The stages

Gate Stage Artifact Skill
G0 Intent lifecycle/intent.md mldlc-intent
G1 Spec lifecycle/spec.md with model-inputs, acceptance and triggers blocks mldlc-spec
G2 Data a data version and a validation run recorded against it mldlc-data
G3 Build a training run and a registered model, from committed code mldlc-build
G4 Evaluate the run's metrics against the spec's thresholds mldlc-evaluate
G5 Release lifecycle/releases/v<N>.md and the production alias mldlc-release
- Operate the spec's triggers mldlc-operate
- Iterate lifecycle/changes/*.md mldlc-iterate

How a change is noticed

Every approval is stored in lifecycle/gates.yaml with the evidence it covered: the hash of intent.md and spec.md, the data version, the MLflow run id, the production version. If any of that changes, the gate reports stale and so does every gate after it. Nothing has to remember to invalidate it.

A change to... re-enters because
the business goal, users, constraints G0 the intent is what the spec was written for
target, label timing, inputs, thresholds, triggers G1 the contract changed
a new data version or source G2 the validation must be redone on it
the model, features or code G3 a new candidate
new evaluation data, such as final labels G4 the same model, judged again
promote or roll back G5 only the release changes

mldlc reentry <kind> prints the gate and the gates to pass again. The production model keeps serving until its replacement passes G5, so an iteration never leaves a gap, and a rollback only returns to a version that was released through G5.

What is deliberately left to people

  • The intent, the thresholds and the release decision. The tool refuses --by gate-check for them, and the skills tell an agent to ask. The tool cannot verify identity, so review gates.yaml like code.
  • Choosing validation and training tools. The kit asks only for the evidence in the evidence contract.
  • Drift and live-quality monitoring. Triggers cover what the latest training run's metrics can show.