Skip to content
PhD Journal Logo Tu vuo far el contrattino

2026-07-10 The Goodhart Law

Coordinating with David Chavez and the other folks: https://github.com/kg-construct/rml-core/issues/266

arcangelo7
arcangelo7Jul 2, 2026 · arcangelo7/rml-io-registry

test: add RML-Core RDB test cases (RMLTC0000-RMLTC0021a)

Port the 59 PostgreSQL test cases removed from rml-core in kg-construct/rml-core@2a232cf ("test-cases: only keep JSON") as RMLTC*-RDB, following the guidance in kg-construct/rml-core#266.

Deviations from the removed snapshot:

  • RMLTC0007h: objectMap references FirstName and graphMap references ID instead of the non-existent Name column, so the only intended error is the literal graph term (mirrors kg-construct/rml-core#143)
  • RMLTC0019a, RMLTC0020a: expected outputs resolve relative IRIs against http://example.com/ instead of http://example.com/base/ (mirrors kg-construct/rml-core#223)
  • RMLTC0002e, RMLTC0002g, RMLTC0002h, RMLTC0003a: object maps reference ID instead of the undefined IDs, so each error test fails for one reason only (mirrors the fixes to the JSON variants)
  • RMLTC0002h: the SQL query duplicates the "ID" column via an alias, as in R2RMLTC0002h, instead of an unquoted query that fails on PostgreSQL before the duplicate-name check is reached
  • RMLTC0003a: the undefined SQL version identifier is expressed as the undefined reference formulation rml:SQL2000Query with a valid query, matching the intent of R2RMLTC0003a
  • RMLTC0002i, RMLTC0002j: the queries select the "ID" and "Name" columns the mapping references (qualified names in 0002j, per R2RMLTC0002j); the expected id literal is typed xsd:integer per the SQL natural mapping
  • RMLTC0005a, RMLTC0005b, RMLTC0012e, RMLTC0016b: xsd:double literals use decimal notation, matching the registry convention
  • RMLTC0015b: references use the lowercase names produced by the unquoted DDL, as in RMLTC0015a, and the second query filters Lan = 'ES'; the only intended errors are the invalid language tags
  • RMLTC0019b: the table keeps only the row with the data error, as in the current JSON variant
  • RMLTC0002f: PostgreSQL JDBC driver and user instead of MySQL
  • RMLTC0004a: dropped a stray student_sport.csv from the CSV variant

All resource.sql files load and all SQL2008Query iterators behave as expected on PostgreSQL 16; mappings and expected outputs parse as Turtle/N-Quads.

arcangelo7
arcangelo7Jul 3, 2026 · arcangelo7/rml-io-registry

test: add missing BlankNode termType to RMLTC0001b-RDB

The PostgreSQL variant of RMLTC0001b never had the rml:termType rml:BlankNode subject since the original import into rml-core (0ff3dc8), diverging from both R2RMLTC0001b and RMLTC0001b-JSON, which generate a blank node subject as the test title states.

arcangelo7
arcangelo7Jul 3, 2026 · arcangelo7/rml-io-registry

test: restore delimited identifier scenario in RMLTC0002f-RDB

The PostgreSQL variant inherited from rml-core inverted the scenario of R2RMLTC0002f: the DDL used regular identifiers while the mapping referenced everything as delimited.

Restore the original design: delimited identifiers in the DDL and in rml:iterator/rml:reference, regular identifiers only in the subject template, which is the non-conforming reference under test.

Drop the "Within rr:template ID is ok, but Name is not" sentence from the description: it only holds under standard SQL uppercase folding, while PostgreSQL folds regular identifiers to lowercase and rejects both references.

Note: RMLMapper v8.0.1 does not reject the non-conforming template and materializes output for this mapping, exactly as it does for R2RMLTC0002f.

arcangelo7
arcangelo7Jul 3, 2026 · arcangelo7/knowledge-graphs-inversion

feat(conformance): support rml io registry rdb tests

Switch the RML conformance suite to the rml-io-registry RDB fork and add handling for the new RML vocabulary used by those cases.

arcangelo7
arcangelo7Jul 7, 2026 · arcangelo7/knowledge-graphs-inversion

fix(benchmark): reduce redundant subject-map queries

KROWN Mappings scenarios repeat the same predicate-object maps across multiple subject templates. Reusing every equivalent subject group made the generated SPARQL query grow unnecessary

Select one reducible subject group per predicate-object signature before building the query, keep non-reducible groups, and publish the updated benchmark results.

Let’s take the case with 5 TM and 8 POM

TM1 subject {p1}, -> p1...p8
TM2 subject {p2}, -> p1...p8
...
TM5 subject {p5}, -> p1...p8

All columns used in the subjects, p1…p5, are already available as literal objects in every TM. We don’t need 5 groups to rebuild the row. We can take only one representative subject map.

This optimization made mappings_5_8 go from 36,35s to 6,30s

I only considered the cases where TM < POM, otherwise it’s impossible to associate a value to a specific TM and, therefore, to a specific column/row

TM1: subject /table/{p1}, -> p1 e p2
TM2: subject /table/{p2}, -> p1 e p2
TM3: subject /table/{p3}, -> p1 e p2

These three subject templates have the same structure. There is no way to understand from the graph that p3 comes from TM3.

arcangelo7
arcangelo7Jul 9, 2026 · arcangelo7/knowledge-graphs-inversion

feat(benchmarks): add gtfs benchmark