Platform Blueprint Technical assurance Insights Company Start a technical conversation
Pillar 04 · ML Factory

Promotion is a decision, not a file copy.

The ML Factory turns objectives, features, evaluation, promotion and serving into governed contracts. A portable engine supplies the mechanism; client model projects retain purpose and policy.

InputGoverned features
VerdictChampion gate
OutputServing bundle

A candidate never promotes itself.

Non-negotiable

A candidate becomes champion only when an explicit gate accepts named evaluation evidence against the current objective and incumbent.

The registry records the verdict and lineage. It does not become the policy simply because it can move an alias.

Specifications travel through stable ports.

A model specification names objective, target, feature contract, evaluation set, metrics, thresholds, serving shape and owner. Engine ports bind that intent to client infrastructure without owning the business decision.

Specify

Objective, target, contracts, metrics, thresholds and accountable owner.

Materialise

Point-in-time-correct feature sets with lineage to governed sources.

Train

Immutable run evidence with code, environment and input identities.

Evaluate

Named evaluation set, threshold failures and challenger comparison.

Promote

An auditable verdict moves the serving reference, never the candidate.

Serve

Versioned bundle includes preprocessing, signature and compatibility.

Correctness before performance

Feature time is part of the contract.

Training data must not contain information unavailable at the inference timestamp. Offline evaluation that violates that rule is evidence of leakage, not model quality.

Serving consumes the accepted bundle as one versioned unit. Reconstructing preprocessing separately at runtime creates an untracked model change.

  • Can point-in-time feature fixtures prove future information is excluded?
  • Can a candidate below threshold be shown failing promotion?
  • Can every champion resolve to evaluation, data, code and specification?
  • Can a serving consumer refuse an incompatible signature?
  • Can the prior accepted bundle be restored without retraining?
  • Do drift signals name a response owner and decision path?
Limitation

Process integrity does not prove that a model is appropriate for a specific purpose. Objective, impact, fairness, monitoring and use acceptance remain accountable client decisions.

Next pillar

A model can recommend. The AI Factory controls context, tools and effects.

Explore AI Factory