Fixture and version

Graph version

2.13.0

Graph nodes

475

Cases in the fixture

60

Expectations met

54 met · 6 unmet

Graph content hash 9d0641791cb90d00ce812f89b29da690de324b255edf6db48e8ca9585a70a94c · dated 2026-08-28. Every case, expectation, negative reason and source URL lives in one file: src/testing/golden/cases.ts.

5 of 60 cases are not independently verified: their expectations do not yet carry an article-level source with a URL and a retrieval date for every value. They still run and are reported, and they are excluded from the sourced counts above.

Cases per category

Composition is part of the result. 26 of 60 cases assert that something does not apply, 18 of 60 are one half of a paired case where only one variable differs, and 8 of 60 turn on the status of an instrument in time rather than on its subject matter.

Metric matrix

Exclusion precision

100.0%

32 negative assertions — how often an instrument the fixture says does not apply was nevertheless presented as applicable

Inclusion recall

100.0%

61 containment assertions — how often a duty the fixture requires was actually derived

Flag recall

83.3%

12 flag assertions — contested, deferred, stayed, not-in-force, role-dependent, insufficient input, prohibited

Mutation kill rate, per operator

A green suite may be green because the graph is right or because the cases assert nothing that could go wrong. The harness settles that: it injects one plausible graph fault, re-runs the suite, and asks whether any previously-met case now fails. Faults that nothing notices are named below. Recorded 2026-08-28 against graph v2.13.0: 8 mutants per operator, 58 of 475 nodes actually asserted about by the fixture.

Surviving mutants — the blind spots

Each line is a specific graph error the suite does not currently detect. This list is the useful output of the harness; the percentages above are only its summary.

Failures accepted with an owner

A case whose expectation is not met may be accepted rather than fixed, but only in writing: with an owner, the date it was accepted, and the reason. The suite asserts that every case marked this way still fails, so an accepted failure cannot quietly become a passing one.

• s-legalresearch · Externally-authored scenarios

  tier lands on rc-limited; role-dependent reading not surfaced as a flag

  Owner kg-curators · accepted 2026-08-15

• s-admissions-utility · Externally-authored scenarios

  authored contested reading not surfaced as a flag

  Owner kg-curators · accepted 2026-08-15

• r-hr-provider · Role pairs

  provider/deployer duty split not applied to Art. 43

  Owner kg-curators · accepted 2026-08-15

• r-hr-deployer · Role pairs

  provider/deployer duty split not applied to Art. 43

  Owner kg-curators · accepted 2026-08-15

• r-rebrand-user · Role pairs

  Art. 25 role shift not applied

  Owner kg-curators · accepted 2026-08-15

• g-scope-drift-before · Externally-authored scenarios

  The reasoner tiers this as high risk on the employment keyword alone ('HR handbook'), although the described system evaluates no person and matches no Annex III point — the over-triggering failure the deep-search audit predicted. Recorded as a disclosed recall/precision gap on the tier, not as a claimed duty: the case asserts no negative expectation, so nothing wrong is published while the detector is narrowed.

  Owner RAIN editorial · accepted 2026-08-17

What this does not cover

• It is not legal advice and not a legal opinion. The expectations were authored by reading statutory text; agreement with them is agreement with a reading, not a determination by a competent authority or a court.
• No independent third party has checked this fixture. Where a case is marked not independently verified, no article-level source is attached to every expected value yet.
• 60 cases are a sample, not a population. The graph carries 475 nodes and the fixture asserts about 58 of them. A fault in a node no case asserts about cannot be detected here — by construction, not by oversight.
• The mutation figures describe a sample of faults (8 per operator), chosen from the asserted region. A 100% kill rate for an operator would be a statement about those mutants only.
• Crosswalk faults remain 0/8 killed, and the zero is real — a declared assertion-vocabulary gap, unchanged by the latest release. The crosswalk-swap operator turns an overlaps_with mapping into equivalent_to. The derivation path the fixture asserts over never reads the relation type — the crosswalk view is a separate surface — so no case can observe the change. This is a coverage gap in the assertion vocabulary, not a fixture-authoring gap, and it is reported rather than papered over with a case that cannot fail.
• Nothing here measures the law itself being current. Instrument status and dates are tracked separately, per node, with a retrieval date and a source; the suite only checks that the derivation respects the status the graph records.
• Confidence bands are ordinal. Cases may assert that one situation is banded strictly lower than another; no case asserts an absolute confidence number, because that would be a claim about the implementation rather than about the law.
• Coverage of markets is uneven. Most cases run against EU and US expectations; other jurisdictions in the graph carry far fewer, or none.

Related: the public track record logs every correction made after publication, and sources and references lists the instruments behind the graph.