Platform Blueprint Technical assurance Insights Company Start a technical conversation
Pillar 03 · Governed Lakehouse

Preserve evidence before interpretation.

The lakehouse separates arrival, preservation, reconciliation and business meaning so a transport choice, pipeline convenience or dashboard cannot silently become truth.

EvidenceLanded + Bronze
Current stateDeclared Silver
MeaningGold + semantics

Ingestion never interprets.

Non-negotiable

Ingestion preserves and reconciles declared source state. Joins, aggregates, entity resolution and measures begin only after the governed consume boundary.

The rule prevents transport logic from acquiring business authority and keeps replay possible from evidence whose meaning has not been rewritten.

Different layers answer different questions.

The medallion is useful only when each layer owns a distinct class of change and authority.

LayerQuestionAllowed changeNot allowed
LandedWhat arrived?Transport evidence and object identityRewriting the source object
BronzeWhat did the source state?Route-independent envelope and source payloadBusiness interpretation
SilverWhat is the declared current and historical entity?Typing, keys, sequence, CDC and SCD behaviourJoins and measures
GoldWhat does the business entity mean?Conformed dimensions, facts and grainPresentation-specific definitions
SemanticHow is the shared measure evaluated?Named measures and governed dimensionsA second calculation inside a dashboard or agent

Controls attach to the contract, not the presentation.

Classification, masking, tokenisation, lineage and freshness are reviewed at the governed object where all consumers inherit them.

Route equivalence

The same source fixture produces the same Bronze meaning through events, files, APIs or CDC.

History

Keys, sequence and SCD type are declarations with fixtures for late, duplicate and out-of-order change.

Semantic ownership

A measure is declared once. Dashboards and agents query it rather than carrying parallel formulas.

PII boundary

Classification and policy travel with data. Tokenised or masked surfaces are deliberate products, not ad-hoc query behaviour.

Freshness

An objective names the source, expected cadence, evidence, owner and response—not only a green tile.

Replay

Derived targets can be rebuilt from preserved evidence in isolation, compared, and promoted through the same acceptance path.

Limitation

A structurally valid model can still encode the wrong business meaning. Named data owners must accept grain, measures and exception policy.

Next pillar

Governed data is available. The ML Factory controls what becomes a model decision.

Explore ML Factory