Case 02 · OPC line telemetry · ERP · conforming layer
Production data arrives continuously, per machine, on the plant's clock. Financial data arrives in periods, per product, on the ledger's clock. Putting them on one chart naively does not merge two truths — it manufactures a third one that is wrong.
Operations had line telemetry: events per machine, effectively per second, describing what the equipment was doing. Finance had the ledger: postings per product, per accounting period, describing what the business had earned.
Both groups wanted the same view — produced volume against financial margin — and both had already tried to build it. Each attempt produced a chart where the two series disagreed by an amount nobody could explain, so each attempt was abandoned, and the two functions carried on running separate versions of the same month.
A chart where two trusted series disagree by an unexplained margin does not cost you one number. It costs you both of them.
The disagreements were not bugs. They were the honest consequence of two systems measuring different things and calling them the same word.
A unit on the line is a thing the machine completed. A unit in the ledger is a thing that was sold, or transferred, or written down — and a unit produced on the last night shift of the month may post in either period depending on when it cleared inspection. A plant day runs shift to shift and does not align to a calendar day, which does not align to an accounting day in a 4-4-5 calendar.
None of that is resolvable by a better join. It is resolvable only by deciding, explicitly and in writing, what each term means and what happens at every boundary where the two systems cannot be made to agree.
I designed a conforming layer that sits between the two sources and does three things before any reporting touches the data: it applies an explicit definition of a unit per system, it maps plant days to accounting days through a rule rather than a timestamp comparison, and it carries a declared tolerance for the residue.
Rows that fall outside tolerance are not dropped and not silently adjusted. They are routed to a quarantine that is itself reported, with the reason attached. The exception queue became one of the more useful outputs of the project — it surfaced a recurring inspection-timing issue that neither function had been able to see from its own side.
Every one of these had a defensible alternative. The trade-off is stated because the trade-off is the point.
Every previous attempt had reconciled inside the visualisation, which meant the reconciliation was invisible, unversioned, and different in each chart. Moving it into a modelled layer made it one artefact that could be reviewed and tested.
Cost: slower to change than editing a chart expression, which was unpopular at first.
The single highest-value hour of the project was the one spent getting operations and finance to state, in one sentence each, what they counted as a unit. They had never been in the same room for that question.
Cost: it surfaced a disagreement that had to be escalated rather than modelled around.
Mapping a plant day to an accounting day by comparing timestamps works until a shift crosses midnight at a period boundary. Making it an explicit, inspectable rule meant the edge case was handled once and identically everywhere.
Cost: the rule needs maintaining when the plant changes shift patterns.
Silently discarding rows that will not reconcile produces a clean chart and an undetectable error. Reporting them produced an uglier first view and a much faster path to the underlying process problem.
Cost: an exception queue is only useful if somebody works it, which required naming an owner.
I built the conforming layer before I had a written owner for the exception queue, and for the first few weeks the quarantine simply accumulated. A reported exception nobody is accountable for is just a slower way of dropping the row. I would now treat naming that owner as a precondition for shipping, not a follow-up.
I would also push harder, earlier, on the unit-definition disagreement rather than modelling around it while it was escalated. Carrying both definitions through the layer for a period added complexity that a decision two weeks sooner would have removed entirely.