Questions to run after each knowledge check or exercise, with notes on what to listen for. The checks test recall. The discussion is where the ideas attach to the participants' own work.
How to use it. Run the check or exercise first, then pick two or three prompts from its section. Give the room a minute of silence before anyone answers. Open "What to listen for" only on the facilitator's screen.
Prompts are written for the participants. The notes are written for you.
Plato's forms, sets and hierarchy, classes and instances, and the difference between an open world and a closed one.
Prompts
Name one thing in your work that is clearly a class and one that is clearly an instance. Now name one you are not sure about. What makes it hard?
Where does your organisation treat missing data as "false"? Where does it treat it as "unknown"? Has that ever caused a wrong decision?
A subclass inherits everything from its parent. Name a case where that inheritance would be wrong for you, and what that says about the hierarchy.
If two teams disagree about whether something is a kind of Customer, who decides, and where is the decision written down?
What to listen for
Confusion between a class and a role. "Supplier" is often a role a Company plays, not a kind of Company. Ask: can the same thing stop being one without ceasing to exist?
Closed-world assumptions hidden in reports. A dashboard that shows "no incidents" when the feed was down is a closed-world reading of open-world data.
Hierarchies built for convenience rather than truth. If inheritance breaks, the parent is usually wrong, not the child.
Nobody can name the owner of a definition. That is the gap an ontology is meant to fill.
Competency questions: the questions a knowledge graph must be able to answer, and how a good one decomposes into classes, properties, controlled values and rules.
Prompts
Write down the last question you asked that took more than a day to answer because the data lived in several places. Is it a strong competency question as written? Fix it.
Which shape was that question: a lookup, a traversal, an aggregation or an inference? Which shape does your organisation ask most, and which does it most struggle to answer?
Take one strong question from the table. Which concepts and relations would have to exist for a graph to answer it? Which of those exist in your systems today?
A question nobody asks is not worth modelling for. How would you find out which questions people actually ask?
What to listen for
Questions that are really reports. "Show me all orders" is a lookup with no decision behind it. Push for the decision the answer would change.
Vague quantifiers: "important customers", "recent issues". Each one hides a controlled value or a rule that someone must define.
Inference questions disguised as lookups. "Which suppliers are at risk?" needs a rule for risk before it needs any data.
The group realising the same question is asked in three departments with three different meanings. Link this to Domain Detective.
Ten words that appear in every domain, and the domain-specific words that finally settle which world you are in. The label did not carry the meaning. The schema did.
Prompts
Which of the shared words does your organisation use in more than one sense? Claim, Settlement, Discharge, Capacity, Record. What has gone wrong because of it?
Pick one shared word. Write its definition for your team in one sentence. Now write it for a neighbouring team. Where do they differ?
How many words did the room need before someone guessed? What was the word that gave it away, and why did that word carry more information than the rest?
Who in your organisation owns the definition of a term when two departments disagree? If the answer is "nobody", what happens instead?
What to listen for
Stories about integration projects that failed on vocabulary rather than technology. Two systems with a "settlement date" that could not be compared.
The instinct to solve it by picking one definition for everyone. Sometimes right, often not. An ontology can hold both meanings as two concepts with two labels.
Participants noticing that the general words were true but useless. This is the case for specialisation: description lives in the specific tiers.
Anyone claiming their domain has no ambiguous terms. Ask them to define "customer".
A blueprint of concepts, hierarchy, relations and attributes, then data that the blueprint checks. Participants built their own, from a blank page.
Prompts
What did you try to state that the builder refused? Was the builder right, or was your blueprint missing something?
Look at your blueprint view and your data view side by side. Which one changes daily in your organisation, and which one should change rarely? Who should be allowed to change each?
How many concepts did you need before your world felt described? What would you cut? An ontology earns its keep by staying small.
Where did you put an attribute that someone else in the room made a concept, or the reverse? Argue both sides.
Your second instance connected to your first through a relation you had defined earlier. What would it take to make that fact true in your real systems?
What to listen for
Frustration at the refusal. Good. That is a schema doing its job, and the same friction is what keeps bad data out of a real graph.
Blueprint changes described as routine. If the T-Box changes as often as the A-Box, the model is not settled and every report built on it is at risk.
Attribute-versus-concept arguments. The useful test: does it need its own identity and relations? A colour is an attribute. A supplier is a concept.
People discovering that their ontology looks like their org chart. Ask whether the world is really shaped like the reporting lines.
One query language, SPARQL, pointed at the data, then at the blueprint, then at both in one query. Everything was a standard, and the graph came out of the builder unchanged.
Prompts
The query that listed your classes had the same shape as the query that listed your customers. What does it change, for an AI system or an analyst, that the rules are readable the same way as the facts?
"For every fact, which rule allowed it?" put data and blueprint in one row. Where in your organisation today would you have to open two systems, or ask two people, to get that row?
The page ran standard SPARQL on a file you exported from another tool, in a browser, with no vendor involved. What would you now ask a vendor before signing, and what answer would end the conversation?
The "Which kinds have no data yet?" question found classes nobody had used. Is that a modelling error, a data gap, or a plan? Who decides?
What to listen for
The moment someone says the ontology "is just more data". That is the lesson landing. Ask what it means for governance if the schema is versioned and queried like a table.
Lock-in stories: a platform that could not export its own model, a query language nobody else speaks. Contrast with the file that moved between the builder and this page untouched.
Discomfort with the empty classes. Modelling ahead of data is normal; the point is that the graph can tell you where it happened.
Anyone equating "graph database" with the standards. Be precise: property-graph products are a separate family with their own ISO query language, GQL; the W3C stack is the one with open vocabulary standards and vendor-neutral files.
Closing the session
Each participant writes three lines and reads them aloud. No discussion, just the reading.
One concept my organisation needs to define and does not.
One relation between two things we talk about but never record.
One question we would ask tomorrow if the graph could answer it.
Collect the lines. They are the seed of a real competency-question list and a real T-Box, written by the people who will use it.