In the data · For architects and data teams
You cannot automate a metric your organization has not defined.
You have built a warehouse before. The pipelines are not the hard part and you know it. This page is about the layer above them: what a definition has to contain, where it sits, and how it stays true once the systems underneath start moving.
The failure mode
Every query is correct. The answers still disagree.
This is not a data quality problem. The values are accurate. The definitions underneath them were never reconciled.
The same name, resolved twice
On-time means ship-by-promise in the ERP and ship-by-request in the reporting layer. Both queries are correct. Both dashboards are wrong about each other, and the difference is invisible until someone reconciles by hand.
The exception that was never encoded
Partial shipments, reworks, intercompany transfers. The formula handles the clean case; the floor is mostly the unclean case, and each team has quietly chosen its own treatment.
A model trained on an unsettled label
An agent or a forecast inherits every one of those disagreements and reports them back as fact, with more confidence and at higher volume than the spreadsheet it replaced.
Cleaning the data does not fix any of these, because none of them are dirty. That is why the work starts above the pipeline rather than inside it.
The semantic layer
A name and a formula is not a definition
It is a formula with a name on it. Five fields are the difference, and the fourth is where the arguments actually live.
Meaning
What the metric asserts about the business, in a sentence an operator would recognize. If two people read it and picture different events, it is not finished.
Owner
One named person who can settle a dispute about it. Not a committee and not a team: definitions fail on authority, not on analysis.
Calculation
The arithmetic, against named source fields in named systems. Written where it can be diffed, not embedded in a report nobody can open.
Exceptions
The cases the formula does not cover and what happens to them. This is where most of the real disagreement has been hiding.
Scope
The value stream it belongs to. Two definitions can coexist if they serve different decisions. What cannot coexist is two definitions under one name.
These are versioned alongside the value stream map they belong to, so a change to a definition is a reviewable diff rather than a conversation somebody half-remembers.
Where it sits
Above the platform, below everything anyone opens
Two bands carry the argument: the namespace that makes the plant floor addressable, and the semantic layer that reconciles ERP and shop floor definitions. Everything else is a box you have seen before.
Plant floor·PLCs, SCADA, historian
Business systems·ERP, PLM, MES, QMS
MQTT broker·HiveMQ or EMQX
Edge gateway·filter, buffer, forward
Unified namespace·enterprise / site / area / line / cell / tag
Managed streaming ingest
Kafka, only if replay is needed
Snowflake
Databricks
Semantic layer·reconciles ERP and shop floor definitions
OEE dashboards
Predictive models
Board reporting
We are reporting-tool agnostic and platform agnostic. Snowflake or Databricks, Kafka only where replay is actually needed, whichever BI tool is already installed. The definitions and the semantic model are the product; the rest is wiring, and it should be chosen against the model rather than against a vendor reference architecture.
Keeping it true
A definition that is not monitored is a definition that has already drifted
Source systems change, someone adds a status value, a team quietly adopts a workaround. The comparison that found the divergence in the first place is re-run against live systems.
This is not a governance program bolted on afterwards. It is the last beat of the same loop, and it happens to be what ISO/IEC 42001 asks for.
Increasingly your customers ask about this in supplier questionnaires before you have decided to care about it.
Who runs it afterwards
Your team. That is the exit criterion.
We document, train, and leave. A model your team cannot extend is a dependency you bought rather than a capability you built, and it fails the moment the engagement ends.
Extending it to a second value stream is not automatically cheaper than the first. What is certain is that the second one starts from ratified definitions, and that the two together answer questions neither could alone.
Where do your definitions stand?
Twelve questions, and a note on what each bad answer costs you downstream. Score it yourself before anyone sells you a platform.
Get the Definition Audit→