9. The metric store | Semantic Layer course

Section 1

A question with no wrong answer

Next section

Section 2

A registry instead of an improvisation

Next section

Section 3

Meridian's first five metrics

Next section

Section 4

The same question, five times

Next section

Section 5

How a metric gets resolved at query time

Next section

Section 6

Build the NPL ratio yourself

Next section

Section 7

When the definition changes

Hands on

Govern a metric

25 min

Define the non-performing loan ratio and total exposure as MetricDefinition triples, write a SHACL shape that validates metric definitions are well formed, then run a query that resolves a metric by name and returns the number plus its definition provenance.

make lab-up            # once, starts Postgres + Fuseki + Ontop
make lab lab09-metric-store

Full instructions, starter files and the solution are in labs/modules/lab09-metric-store/README.md. The lab checks its own work, so you will know if it passed.

Check yourself

Quiz

7 questions, pass mark 70%. Answers are explained either way, so a wrong guess is not wasted.

1 An agent with no metric store answers "average processing time" as 11.4 days on Tuesday and 9.2 days on Thursday, with no data change. Which ACE property failed?

Accuracy, because one of the two numbers must be arithmetically wrong Consistency, because the same question resolved to two different defensible formulas Explainability, because the agent gave no citation None, because both numbers are within a reasonable range

2 Why does deterministic metric resolution serve ACE better than asking a capable model to compose the aggregation each time?

The stored expression executes faster than model-generated SQL A pre-compiled expression gives the same answer on every call, and citing its name and version tells the reader exactly which definition produced the number It removes the language model from the pipeline entirely Stored expressions are validated against the data at publish time, so they cannot return a wrong value

3 Which two properties does the guidance require of a metric definition, and why? (pick two)

Select all that apply.

Ontology-grounded, so the expression names classes and properties and resolves across federated sources through the virtual knowledge graph Materialized, so the metric value is precomputed nightly and stored as a triple Governed, so definitions move through the same draft, review and publish lifecycle as the ontology and are validated by SHACL Model-generated, so the definition stays current as the schema changes

4 Meridian's NPL ratio definition sets its numerator filter to "?loan a mrb:NonPerformingLoan" rather than to a days-past-due comparison. What does that buy?

It runs faster, because class membership is indexed The metric follows rule R1, so changing the arrears rule changes the metric without editing the metric It lets the metric work on loans with a null days past due It avoids needing a reasoner at query time

5 Total Group Exposure uses SUM(DISTINCT ?outstandingBalance) over a UNION of borrowed and guaranteed loans. Why DISTINCT?

To exclude loans that appear in more than one related-party group Because the UNION can bind the same loan twice, once as borrowed and once as guaranteed, which would double-count Mei Lin Tan's 88,000 Because SPARQL requires DISTINCT inside any aggregate To drop loans with a null outstanding balance

6 The regulator changes the NPL ratio definition in April 2026. What does the metric store do?

Edit the existing definition in place, so every dashboard immediately reflects current policy Publish version 2 alongside version 1 with a supersedes link, effective dates and a change reason, so old answers stay reproducible and pinned dashboards keep using v1 Delete the metric and require each team to define its own Keep both versions but stop citing versions, to avoid confusing readers

7 What does the SHACL shape for metric definitions enforce that prevents the original ambiguity returning?

That the expression is valid SPARQL syntax That exactly one mrb:metricExpression exists, along with a unit, a version and an approver That the metric value falls within a plausible range That the metric references at least two ontology classes