Operational architecture
What an operational ontology connects
A shift report and a pump controller can describe the same hours at the same pump and still disagree. How Ancilor ties both to one pump, one shift and one site, and keeps the difference in view until a person resolves it.
An operational ontology is a working description of what a business tracks and how those things relate: sites, tanks, pumps, shifts, deliveries, people. It also defines which actions are allowed on each of them, and who may take them. In the formal sense, the W3C's OWL 2 is a language for ontologies with formally defined meaning; the operational version is plainer, but keeps the same discipline.
The test is practical. When the figures at a site disagree at the end of a shift, can the reviewer see which pump, which readings and which report are involved, and who has to act next, without rebuilding the story from three screens? Ancilor, Harrimont's platform for enterprise data and operations, is designed around one model of this kind. This article follows one constructed example through it. The site, figures and times are illustrative.
Fifty litres at pump 3
At the end of the morning shift, the site manager records the shift in Ancilor Field: tank levels, the opening and closing reading on each pump, who worked, card and cash takings, and vouchers issued to account customers. Three records now describe pump 3 for the same hours.
| Record | What it says | Recorded by |
|---|---|---|
| Pump 3 readings in the shift report | 1,240 litres sold | The site manager |
| Controller total for pump 3 | 1,190 litres | One person entered it, a second confirmed it |
| Card and cash takings | Consistent with 1,240 litres | The site manager |
Several explanations fit the 50 litres: a misread meter, a controller channel mapped to the wrong pump, fuel drawn for a calibration check and poured back, a total typed against the wrong shift. Each has a different owner. The model's job is to make the difference precise enough that the right person can take it up.
Name the things before the screens
The objects are ordinary: the site, the tank, the pump, the shift, the shift report, each reading and each controller total. Each has one home and an identifier that screens and connections refer to. The relationships carry the meaning. A pump draws from a tank. A reading belongs to a pump and a shift. A report belongs to a site and a day, and once it is accepted, its closing tank figures become the next shift's opening stock.
With those links in place, the difference has an address: pump 3, morning shift, this report, this site. A later question, such as whether the same pump drifted last week, becomes a query rather than a project.
The mapping is a record too
Sources rarely agree on names. The controller numbers its channels one way, staff number pumps the way they appear on the forecourt, and the accounting package knows the site by a cost-centre code. Joining on a label that only looks right is how a reading lands on the wrong pump, and the result looks coherent while being wrong.
The approach that holds up is to keep each source's own identifier beside the shared one, and to treat the mapping between them as a record in its own right, with who set it and the date from which it applies. When a pump is replaced or renumbered, the mapping changes from that date and earlier readings keep their earlier mapping. The trade-off is upkeep: someone has to own the mapping. The alternative is a quiet error that surfaces months later in a stock count.
Keep what happened apart from what the record says now
A reading is an observation at a moment. Submitted, returned and accepted are states of the report, and each is reached through a named action. A run-out date for the tank is an estimate. One editable status field for all three hides which is which.
Each entry should also carry two times: when it happened and when it was recorded. A controller total typed in the next morning still belongs to yesterday's morning shift. Ancilor keeps each submitted version of a shift report, so a correction does not erase what the reviewer saw when they made their decision. If the 50 litres turn out to be a misread meter, the manager corrects the reading and submits again, and both versions remain to compare.
Show where each figure came from
A figure is only as useful as the account of where it came from. For pump 3 the reviewer needs the manager's readings, the controller total, who entered and who confirmed that total, and how old each one is. The W3C's PROV Ontology, for one, is a published vocabulary for representing and exchanging provenance information generated in different systems.
Missing, stale and conflicting figures need different responses:
- No controller total yet is not a zero. It is a gap with an owner, and the reviewer should see what the report is waiting for.
- A connection that returned nothing is not proof that nothing was sold. The age of the last good read belongs on the screen.
- Two figures that disagree stay side by side, with the difference assigned to a person, until someone explains it.
A single green tick would hide all three cases. So would an average.
Put decision rights in the model
Seeing a record is not the same as being allowed to change it. In Ancilor Field, the rules travel with the actions. The site manager saves and submits the report, and submitting needs a named worker on the shift and passing pump checks. A reviewer who is not the author accepts the report or returns it with a note. The controller total is entered by one person and confirmed by a second.
A save made on an out-of-date copy is refused as a conflict instead of overwriting newer work, and a month cannot be closed while any report in it is still unaccepted. Actions are logged with who took them and when.
For pump 3 that gives a clear path. The reviewer returns the report with a note naming the pump. The manager checks the meter and corrects the reading or explains the difference. If the mapping is at fault, its owner corrects it from a stated date.
AI works inside the same boundary. It can read the report and the controller totals, set out the difference and draft the note to the manager. It does not accept, return or edit the report. A person decides and acts in the application. The article on AI-assisted decisions sets out what an assistant has to show before a person decides. The technical reports RATIO and LIBRA set out research behind this separation; both propose architectures and report no product results.
One model or separate applications
Separate applications are quicker to start, and each team gets screens shaped around its own work. The cost arrives later: the same site in four lists, and every question that crosses them answered with a spreadsheet. One shared model gives common identifiers and one place for the rules, but it asks for agreement on what each word means, and discipline, because a change to a shared object reaches every screen that reads it.
Ancilor is designed around one model of sites, assets, people and events, while each module, from Ancilor Logistics to Ancilor Vision, keeps the screens its users need. The model is designed to be extended one workflow at a time, so the meaning of each word is agreed with the people who own that work rather than in advance for everyone.
Change the model without rewriting the past
The model will change. A site adds a tank, two product grades are merged, a pattern of two shifts becomes three. Each change raises the same question: what do earlier records mean under the new definitions? The pattern that holds up gives each definition a date from which it applies, retires old identifiers instead of reusing them, and moves existing records by a written rule that someone has checked. A report accepted under two shifts should still read as it did when it was accepted. Changing a shared definition is a decision in its own right, with an owner, because every screen that reads the object changes with it.
Where the model stops
Ancilor reads from the systems a business already runs, such as accounting, pump controllers, bank files and spreadsheets. In the workflow above, nothing flows back. Ancilor does not write to the pump controller or the accounting package: a correction is made in Ancilor and logged with its author and time, the controller keeps its own record and the accounting package remains the record of account.
The model should not become another silo either. Many businesses already keep a data warehouse or an enterprise data platform for reporting, and their analysts need to match these records with their own. That is easier when each object keeps a stable identifier, each source keeps its own identifier beside it, and the mapping between them carries its dates, as described above. Ancilor is designed to work beside the systems and data platforms a business already uses, and each connection is agreed with the people who own the system.
Prove it on the daily close
Start with the daily close at one site. Agree the identities, map the sources, keep the existing approval path and run Ancilor read-only beside it, comparing its view with differences the team has already resolved. Then test the awkward cases on purpose: a renumbered pump, a controller total entered a day late, a missing total, an author trying to accept their own report. Each should end in a clear state with a named owner. Only then extend the model to deliveries, maintenance and the next site.
References
- PROV-O: The PROV OntologyW3C Recommendation, 2013. www.w3.org/TR/prov-o
- OWL 2 Web Ontology Language Document Overview (Second Edition)W3C Recommendation, 2012. www.w3.org/TR/owl2-overview
- RATIO: An Architectural Study of Context-Aware Artificial Intelligence in Enterprise SystemsArgjend Bajraktari. Technical report, version 1.0. Zenodo, 2026. zenodo.org/records/23911976
- LIBRA: An Evidence-Centred Architecture for Professional Memory and Decision ContinuityArgjend Bajraktari. Technical report, version 1.0. Zenodo, 2026. zenodo.org/records/23986865
Start with one workflow.
Bring one process, the systems it already runs on and the decision your team needs to make.