Plan: Applied Insurance Reference — Phase 3, Eligibility Readings and Crosswalks

Unit ID: applied-insurance-reference-phase-3 Epic: applied-insurance-reference Status: Rolling-wave outline. AIR-3.1 to AIR-3.3 are substrate slices, detailed when their ADRs are accepted. AIR-3.4 and AIR-3.5 need the cause scheme and are detailed at AIR-2.4. Decisions: ADR-A100 (D11), ADR-A103 (D12) Sketches: mork-bridge.md, peril-vocabulary.md §6.8

Scope

Make Eligibility decide correctly when a peril list is flat (A-100) and when a path reaches several values (A-103), then let reviewed crosswalks carry a flat or simply taxonomic list to the reference. Milestone M2 is a check across cause, agency and mechanism, run against the reference, a flat list and a crosswalked list.

AIR-3.1 to AIR-3.3 change substrate. Each follows ADR-A-C2: the ADR’s non-insurance examples are authored as fixtures before the README prose, and no insurance term enters either layer.

Slices

Tests at L1 and L2:

• 3.1: on a flat scheme, a required code matches, an excluded code is excluded, every other member is Undetermined with exe:NoHierarchy, and a hierarchical scheme evaluates as before. SPARQL, SHACL and SWRL agree. The OWL backend refuses
• 3.2: each reading over zero, one and several values, including an exclusion read EveryValue, and a binding without a reading behaving as today. A negated condition swaps Permitted and Denied and keeps Undetermined with its diagnostic. SPARQL and SHACL agree
• 3.3: SWRL derives only the sound subset. OWL classes stand in the expected subsumptions
• 3.4: a crosswalk is stale when either edition changes. An inexact mapping never decides
• 3.5 (M2): the cyber write-back profile of peril vocabulary §6.8 decides on the reference, is Undetermined with a reason on a flat list, and decides exactly mapped codes of a crosswalked list. A risk whose perils are all of sudden onset is Permitted under EveryValue

Order and lanes: machines, branches and merges. Substrate item S2 also changes Eligibility, and rebases onto AIR-3.2.

AIR-3.1 in detail

Machine: R (Claude Code). Branch: air/3.1-flat-hierarchy. Validation Pack: applied-insurance-reference-3.1. Decision: ADR-A100, with its implementation note.

Invariant: under elg:HierarchicalMatch, a resolved scheme in which no member has a skos:broader link to another member decides only the concepts a condition names. Every other member is Undetermined with exe:NoHierarchy, in every backend, and hierarchical schemes evaluate exactly as before.

Authored in ADR-A-C2 order: the two example fixtures first, then the law, then the code.

1. Examples in ontology/eligibility/examples/, no owl:Ontology, no insurance terms:
   • flat-scheme-lending.ttl: a contract whose unscoped voc:boundScheme is a hierarchical industry classification (manufacturing above food processing and textiles, plus retail and wholesale), and a voc:SchemeBinding scoped to one lender that binds the lender’s flat sector list. The list reuses only concepts whose broader concept lies outside it (food processing, textiles, retail, wholesale), because skos:broader belongs to concepts and a reused parent and child would carry their link into the list. One HierarchicalMatch condition requires manufacturing or retail and excludes textiles. Recorded decisions: under the classification, food processing and retail Permitted, textiles and wholesale Denied. Under the lender’s list, retail Permitted, textiles Denied, food processing and wholesale Undetermined with exe:NoHierarchy.
   • flat-scheme-employment.ttl: an exclusion-only condition (“every role except contractors”) over a flat job-title list. Contractors are Denied, and every other title is Undetermined, not Permitted, since L12’s default inclusion needs the hierarchy to know a title is not a kind of contractor.
2. The law elg:L14 in ontology/eligibility/vocab/eligibility-vocab.ttl and its README mirror (§4 law list and the decision table, next to L11), with elg:lawRegister elg:SemanticLaw: “Hierarchy precondition. Under hierarchical match, when no member of the resolved scheme has a broader concept within the scheme, a candidate that is a member and is neither a required nor an excluded concept leaves the condition undetermined.” The README’s “HierarchicalMatch is valid when…” sentence cites L14. Eligibility vocab 0.6.0 → 0.7.0 (MINOR, a new individual).
3. The diagnostic exe:NoHierarchy in ontology/mork/spec/Executable.ttl, beside exe:AboveExclusion, citing elg:L14. Executable 0.5.0 → 0.6.0 (MINOR). No ontology imports either document, so the cascade is the catalog and the release register only.
4. The IR (tools/mork_compilers/src/mork_compilers/eligibility_ir.py): _expand takes whether the resolved ordering has any skos:broader link. When the match is hierarchical and it has none, a member is Denied if it is an excluded concept, Permitted if it is a required concept, and Undetermined otherwise. ConceptPlan gains a derived no_hierarchy property the backends read.
5. SPARQL (sparql_backend.concept_select): a plan with no_hierarchy uses the equality decision with Undetermined as its last branch, and reports exe:NoHierarchy as the diagnostic. Scheme membership still wraps it, so a non-member stays exe:OutsideScheme.
6. SHACL (shacl_backend): reads the expansion already. Its Undetermined message names exe:NoHierarchy for such a plan.
7. SWRL: no change. concept_facts reads the expansion, so Undetermined members derive nothing.
8. OWL (owl_backend): refuse a plan with no_hierarchy with an IRCompileError citing ADR-A100, since a design-time class cannot express Undetermined.
9. Tests: a new test_flat_schemes.py, reusing the SHACL and SWRL helpers of test_concept_backends.py (which already imports test_hierarchical_conditions.py, so new cases there would import in a cycle), and the two examples added to tools/test_eligibility_examples.py’s EXAMPLES. The one existing change registers exe:NoHierarchy in test_concept_backends.py’s closed set of diagnostics.
10. Docs: tools/mork_compilers/README.md (L14 in each backend), and the Eligibility row of docs/architecture/ontology-architecture.md §3. Then mise run build:ontology-catalog and mise run build:ontology-releases.

Adversarial probes for the gate: disabling the L14 branch in _expand fails AIR31-06, 07 and 08. Disabling it in sparql_backend.concept_select fails AIR31-03, 04, 06 and 07.

AIR-3.2 in detail

Machine: R (Claude Code). Branch: air/3.2-set-readings. Validation Pack: applied-insurance-reference-3.2. Decision: ADR-A103.

Invariant: a bound condition reads several values by the reading its binding declares, a negated condition swaps Permitted and Denied and keeps Undetermined with its diagnostic, a binding without a reading and a condition without negation evaluate exactly as before, and the SWRL and OWL backends refuse what they cannot yet compile.

Authored in ADR-A-C2 order: examples, then laws and README, then code.

1. Examples in ontology/eligibility/examples/, no insurance terms:
   • set-reading-admissions.ttl: applicants bound through ex:holdsQualification. One condition, read elg:SomeValue, requires mathematics or physics. A second, read elg:SomeValue and elg:negated true, requires medicine (“holds no qualification in medicine”). Recorded: history and physics admitted, history alone refused, no qualification Undetermined, a medic refused by the negated condition.
   • set-reading-trial.ttl: patients bound through ex:hasDiagnosis, read elg:EveryValue, over a permitted list that excludes one diagnosis. Recorded: all permitted admitted, one excluded refused.
2. Eligibility spec (spec/eligibility.ttl and its README mirror, 0.6.0 → 0.7.0, MINOR): elg:ValueReading (a class, added to the disjoint classes axiom), elg:valueReading (functional object property, domain elg:EvidenceBinding, range elg:ValueReading), elg:negated (functional datatype property, domain elg:Condition, range xsd:boolean).
3. Eligibility vocabulary (0.7.0 → 0.8.0, MINOR): elg:SingleValue, elg:SomeValue, elg:EveryValue, and the laws elg:L15 (set readings, as ADR-A103 decision 1 states them) and elg:L16 (negation, decision 4). README: the evidence bindings paragraph and the law list.
4. Eligibility shapes (.version 0.1.0 → 0.2.0, MINOR): the structural shapes gain sh:maxCount 1 for elg:valueReading and elg:negated, mirroring the functional axioms.
5. Cascade (ADR-A86 addendum item 1, each MINOR): every importer of eligibility/0.6.0 re-pins, and so on through their importers: eligibility-vocab, Instrument 0.6.0 → 0.7.0, Behaviour 0.6.0 → 0.7.0 (it imports Eligibility and Instrument), and the Capacity execution profile 0.6.0 → 0.7.0, with their README mirrors. The list is re-derived with grep before and after editing.
6. IR: EvidencePath.reading (default elg:SingleValue), and negated on concept and interval plans. Readings apply to bound conditions only. A question offering several candidates stays Undetermined (exe:SeveralCandidates).
7. SPARQL: for SomeValue and EveryValue, decide each value as a single candidate is decided, then group by subject and aggregate by strong Kleene logic. No value is Undetermined with exe:MissingCandidate. A set that is Undetermined reports a diagnostic of one of its Undetermined values. Negation swaps the final decision. Profiles read the condition’s final decision.
8. SHACL: the same readings and negation, through the SPARQL-based constraints.
9. SWRL and OWL: refuse a plan with a reading other than SingleValue, or with negation, with an IRCompileError naming AIR-3.3, until that slice lands.
10. Tests: tools/mork_compilers/src/mork_compilers/test_set_readings.py, and both examples in tools/test_eligibility_examples.py.
11. Docs: Eligibility README, tools/mork_compilers/README.md, the Eligibility row of docs/architecture/ontology-architecture.md §3, and solution-design-specification.md (set readings as an evaluation feature). Then mise run build:ontology-catalog and mise run build:ontology-releases.

Documentation deltas

Exit gate

A-100 and A-103 Accepted and implemented. M2 demonstrated. MB-Q2 answered.