A NumberShould CarryIts Evidence.
Most organisations have two reports that disagree and no way to settle which is right. Ronexer engineers the layer between operational records and the people, products and systems depending on them — with definitions, lineage, quality and correction designed in rather than added after the first dispute.
Next Step / Technical Discovery
- Owned where the record is made
- Agreed, versioned, dated
- Traceable to the rows behind it
- Measured, and published with the data
Why two reports disagree
Rarely because anybody was careless. Four ordinary conditions, each producing a figure defensible on its own and irreconcilable with the one beside it.
- 01
The same word means two things
Finance counts on invoice, the product team on activation, and both columns are labelled revenue. Neither is wrong, and nothing states which question each answers. An owned, versioned definition makes that a decision rather than a recurring argument.
- 02
A source changed and the model did not
An upstream field gained a value nobody mentioned. The transformation kept running, row counts looked ordinary, and the split has been wrong since. A checked expectation at the boundary makes that a failure rather than a report.
- 03
Yesterday's number changed overnight
A correction improved this month and rewrote last quarter with it. The new answer is better; nobody can say what the old one was, or why the next deserves more confidence.
- 04
The job succeeded and the data is wrong
One source arrived late, so the dataset published without it. Every check passed, because the checks confirmed the pipeline ran rather than that the data was whole.
What Ronexer engineers
Six areas of ownership, and what each one leaves behind. A data platform whose sole output is a running job is one nobody can interrogate.
- 01
Source and ingestion boundaries
Which system owns each fact, and what arriving from it is taken to mean. Produces a source inventory with a named owner per feed.
- 02
Transformation and models
Derived tables, their grain, and the reasoning behind each — a transformation changes meaning, so it needs somebody accountable. Produces a documented model layer.
- 03
Definitions and semantic ownership
Metric meaning held with a version and an owner, not in the head of whoever wrote the first query. Produces a definition register consumers can cite.
- 04
Lineage and observability
Which rows and which transformations produced a figure, and which checks ran. Produces a lineage model that answers the question without a code read.
- 05
Quality and reconciliation
Completeness, freshness and agreement with the operational systems — tested before publication, not after somebody complains. Produces a quality policy and a reconciliation design.
- 06
Access, retention and governance
Who sees which columns, how long rows survive, and what a privacy request must do. Produces an access model that survives review.
What a definition has to settle
Two teams can run the same query against the same table and still disagree, because the disagreement was never technical. A field named revenue is a column, not a definition. What ends the argument is a set of decisions somebody made and wrote down.
Grain
A customer, an order, an order line, or a day of one of those. Most metric disputes are two people summing at different grains.
Boundaries
Which cancellations, which internal accounts, which test records, which currency at which rate. Exclusions are as much the definition as inclusions.
Calendar
Fiscal or Gregorian, which timezone, and whether an event belongs to when it happened or when it was recorded. Those answers differ at every month end.
Ownership
One named person — not a committee, not whichever analyst was nearest the ticket. Meaning without an owner drifts by accretion until nobody recognises it.
Effective date
A definition that changes carries the date it changed from, so last year's figure stays interpretable under the rules that produced it.
Lineage, and what it is for
Not pipeline monitoring. Publishing a dataset is a commitment, and a decision cannot be defended by a number whose path cannot be reconstructed — so the test is what a reader may ask about a figure before acting on it.
- The source records behind a figure can be listed, not inferred from a diagram somebody drew a year ago.
- Every transformation between those rows and the result is identifiable, with the version running at the time.
- What was tested and what passed is published beside the dataset, so a green job is not mistaken for a correct one.
- Whether every expected source arrived is a stated property, and a partial dataset says so rather than looking whole.
- The age of the data is published with it, against a threshold set by what the dataset is actually used for.
Correcting history without hiding it
Data arrives late, sources correct themselves, and a model that was right becomes wrong the day somebody discovers what a field meant. None of that is avoidable. What is avoidable is a correction that quietly replaces last quarter and leaves nobody able to explain it to a board that saw the earlier figure.
So a restatement is an event, with a date, a reason and a scope. The previous answer stays reconstructable, the outputs it touched are identified rather than guessed at, and anyone who acted on the old number can be told.
What changes operationally
- The definition carries an owner, a version and a date, so two reports agree — or differ by a stated decision rather than by accident.
- Freshness and completeness are published with the data, so a consumer knows what they are reading before acting.
- A source that changes shape fails a check at the boundary, rather than reaching a report and being noticed a quarter later.
- The earlier answer stays reconstructable and the affected outputs can be listed, so restating a figure is an announcement, not a discovery.
Related pages
AI Engineering
Production AI engineered around a company's own data, permissions and workflows, with the evaluation and integration that make it usable.
Cloud Infrastructure
The operating environment beneath critical software: deployment, isolation, observability and recovery, engineered so failure stays bounded and the service returns to a known state.
Bring two reports that disagree, a metric nobody will put their name to, or a correction you cannot explain to the people who saw the earlier number.