feat(ontology): add source-grounded occupational taxonomy

ARCHITECTURE.md

@@ -79,7 +79,7 @@ flowchart LR
 | `commitment_extraction.py` | Pluggable LLM derivation of a customer commitment (promise + deadline) from a post; `Null` default, `ContextualOrchestrator` real impl |
 | `temporal_expressions.py` | Pure Korean relative-time resolver for Global Ask (ADR 0150) |
 | `ask_time_axis.py` | Event-time vs ingestion-time clock choice for that window (ADR 0202) |
-| `ontology.py` | Loads `docs/ontology/lineageweave-kg.ttl`, the formal OWL 2/RDFS/SKOS vocabulary for the Knowledge Graph's node/edge types (ADR 0004) |
+| `ontology.py` | Loads the governed Turtle source tree (`lineageweave-kg.ttl` plus generated fragments), the formal OWL 2/RDFS/SKOS vocabulary for the Knowledge Graph's node/edge types and source taxonomies (ADR 0004, ADR 0252, ADR 0255) |
 | `ontology_neighborhood.py` | Bounded typed ontology/provenance neighborhood (ADR 0184); PostgreSQL stays authoritative, OWL subclass is not an instance edge |
 | `ontology_source_cursor.py` | Opaque HMAC source-window continuation (ADR 0124); keyset pagination, never OFFSET |
 | `period_report.py` | Fit GRM/GPCM on persisted IRT rows, FIPC-select, EAP-score a period (ADR 0003 slice 3; Bock & Mislevy, 1982) |
@@ -654,7 +654,8 @@ vocabulary (`node_type`, `edge_type`, `entity_relationship_type`,
 `person_side`, `corporate_entity_level`) actually matches what the
 Ontology/Semantic-Layer claim implies.
 
-`docs/ontology/lineageweave-kg.ttl` is a real OWL 2 / RDFS / SKOS
+`docs/ontology/lineageweave-kg.ttl` and its deterministic governed fragments
+are a real OWL 2 / RDFS / SKOS
 ontology in Turtle syntax: classes for `Post`/`Person`/`CorporateEntity`
 (with `OurSidePerson`/`CounterpartyPerson` subclasses), object
 properties for each `edge_type_code` and `entity_relationship_type`
@@ -668,7 +669,7 @@ specification over it, in the same sense W3C's own stack uses "semantic
 layer" (RDFS/OWL as the governed conceptual layer over raw data), not a
 separate BI-metrics product and not a parallel triple store.
 
-`lineageweave/ontology.py` parses the Turtle file once with `rdflib`
+`lineageweave/ontology.py` parses the Turtle source tree once with `rdflib`
 (pure Python, no Rust toolchain, unlike `fast-mlsirm`) and exposes the
 vocabulary as importable IRI constants, so application code has one
 canonical name per class/property instead of re-typing lookup codes as

CHANGELOG.md

@@ -8,6 +8,32 @@ All notable changes to this project are documented here. Format follows
 
 ### Added
 
+- The complete O*NET 31.0 Content Model Reference is now a deterministic
+  3,006-concept SKOS fragment with exact source IDs, names, descriptions, and
+  documented outline parents (ADR 0255). Its read model fails closed on
+  malformed hierarchy data; no occupation rating, person trait, score,
+  weight, or inferred crosswalk is introduced.
+- The complete 2018 SOC hierarchy is now published as a deterministic
+  ontology fragment: 23 major groups, 98 minor groups, 459 broad occupations,
+  and 867 detailed occupations retain their official codes, titles, levels,
+  and source-declared parents (ADR 0252). The read model supports fail-closed
+  lookup across all 1,447 nodes; no employer job-family mapping, weight, or
+  person-level inference is introduced.
+- The occupational-classification and worker-characteristic taxonomy is now
+  published in the canonical ontology: all 23 major groups of the 2018
+  Standard Occupational Classification (the O*NET job families) carry
+  official titles and codes verbatim, the four O*NET 31.0 job-zone categories
+  carry their published names and source values, and source-native
+  worker-characteristic families are addressable -- Fleishman's four ability domains,
+  Holland's six RIASEC interest types with the published hexagonal adjacency,
+  the six legacy O*NET work-value clusters, and the seven higher-order
+  revised O*NET Work Styles dimensions (ADR 0245). Typed derivation properties from
+  classifications to characteristics are declared but assert no instance
+  binding; a deterministic application read model
+  (`lineageweave.io_taxonomy`) exposes fail-closed lookups, and no numeric
+  importance or level rating is imported. Each scheme links to versioned
+  PROV source entities with publisher/creator and rights/license metadata;
+  the stable O*NET 31.0 Job Zone JSON carries its verified SHA-256.
 - Evidence-bound occupational construct semantics now keep cognitive
   abilities, work styles, work activities, affective reactions, performance
   behaviors, and FJA worker functions distinct. Record-to-construct links

docs/adr/0159-published-ontology-pages.md

@@ -40,12 +40,17 @@ and consumer migration plan.
 3. Publish equivalent machine-readable artifacts beside the HTML:
    `ontology.ttl`, `ontology.jsonld`, `ontology.nt`, the PROV-O support profile,
    and a source-digest manifest.
-4. Preserve `lineageweave-kg.ttl` byte-for-byte as the published Turtle
-   artifact. JSON-LD and N-Triples are generated from a canonicalized RDF graph
-   and are tested for semantic isomorphism with the source.
+4. For a single governed Turtle source, preserve `lineageweave-kg.ttl`
+   byte-for-byte as the published Turtle artifact. When a later accepted ADR
+   adds governed fragments, publish the merged graph as deterministic canonical
+   RDF that parses as Turtle; test every machine format for semantic
+   isomorphism with the complete source graph. Never concatenate independent
+   Turtle documents because their prefix and base declarations have
+   document-local scope.
 5. Do not add a build timestamp. The same source tree must produce the same
-   artifact bytes. The manifest records the source SHA-256 and the complete
-   published ontology-directory inventory instead.
+   artifact bytes. The manifest records every governed source path and SHA-256,
+   the ordered source-tree SHA-256, and the complete published
+   ontology-directory inventory instead.
 6. Run publication through `scripts/publish_ontology_site.py`, a fail-closed
    boundary that rejects duplicate HTML fragments, non-HTTP(S) linked IRIs,
    symlink outputs, source-overlapping outputs, and replacement of directories

docs/adr/0245-io-occupational-taxonomy-in-the-published-ontology.md

@@ -0,0 +1,160 @@
+# ADR 0245: Occupational classification and worker-characteristic taxonomy in the published ontology
+
+**Status:** Accepted
+**Date:** 2026-08-26
+**Extends:** [ADR 0004](0004-knowledge-graph-ontology.md), [ADR 0145](0145-psychometric-channel-weight-estimation.md), [ADR 0207](0207-repository-case-ontology-namespace-canonical.md), [ADR 0232](0232-worker-function-taxonomy-in-the-published-ontology.md)
+**Superseded in part by:** [ADR 0252](0252-complete-2018-soc-hierarchy.md), which expands the major-group-only scheme into the complete 2018 SOC hierarchy.
+
+## Context
+
+Industrial and organizational psychology classifies work twice: once by
+*what* the occupation is -- the 2018 Standard Occupational Classification
+major groups, which the O*NET program publishes as its job families --
+and once by *which human characteristics* doing it exercises. The O*NET
+content model organizes those characteristics under every occupation as
+abilities, occupational interests, work values, and work styles
+(Peterson et al., 1999; Peterson et al., 2001). Holland's RIASEC hexagon
+supplies the interest vocabulary with a published structural claim about
+adjacency (Holland, 1997); Hogan and Holland (2003) established the
+relationship between personality and job performance but is not the source
+of the O*NET vocabulary; the revised O*NET Work Styles report supplies its
+seven higher-order dimensions. Fleishman and Quaintance (1984) supply the
+four ability domains. O*NET 31.0 publishes four Job Zone categories under
+source values 2 through 5 after combining the former first two zones.
+
+ADR 0232 already publishes the DOT/FJA worker functions, so stored
+evidence can name *how* a worker functioned on data, people, or things.
+It cannot yet say *which occupational family* the evidence belongs to,
+nor resolve its cognitive, affective, and behavioral content into the
+published characteristic families. The ontology is again the correct
+home, under the same two constraints that bound ADR 0232:
+
+1. The lookup-code round trip (`tests/test_ontology.py`) must stay
+   untouched: these concepts are not `common_lookup_value` rows.
+2. Measurement stays governed by ADR 0145: no numeric importance or
+   level rating from any occupational profile may be imported, fitted,
+   or renormalized.
+
+## Decision
+
+1. Publish all 23 major groups of the 2018 Standard Occupational
+   Classification verbatim -- official titles and ``NN-0000`` codes --
+   as a `skos:ConceptScheme` (`:socMajorGroupScheme`) of
+   `:OccupationalMajorGroup` concepts, matching the O*NET job-family
+   grouping.
+2. Publish the four O*NET 31.0 Job Zone categories (`:jobZoneScheme`) with
+   their published names and source values (`:jobZoneLevel` 2-5). These
+   values are source identifiers, not fitted or ordinal weights.
+3. Publish distinct source-native worker-characteristic families, without
+   collapsing them into cognition, affect, or behavior, as subclasses of
+   `:WorkerCharacteristic` inside one scheme
+   (`:workerCharacteristicScheme`):
+   - Fleishman's four ability domains (`:AbilityDomain`);
+   - Holland's six RIASEC interest types (`:InterestType`), each with
+     the standard Interest Profiler family description verbatim;
+   - the six historical O*NET work-value clusters (`:WorkValueCluster`),
+     explicitly labeled legacy because O*NET 31.0 no longer publishes the
+     Work Values branch;
+   - the seven higher-order dimensions in the revised O*NET Work Styles
+     structure (`:WorkStyleFamily`). The 21 lower-order dimensions and the
+     separate four-component occupation-level analysis remain an explicit
+     import gap; neither may be inferred from these family nodes.
+4. Assert only the published structural relation between interest
+   types: `:riasecAdjacentTo`, a symmetric property whose six asserted
+   pairs are exactly the hexagonal ring edges Realistic-Investigative-
+   Artistic-Social-Enterprising-Conventional-Realistic (Holland, 1997).
+   Adjacency is a similarity ordering, not a score.
+5. Declare `:OccupationalClassification` as the future common hierarchy
+   class and four domain/range-typed derivation properties --
+   `:occupationalAbilityDemand`, `:occupationalInterestProfile`,
+   `:occupationalValueOrientation`, and `:occupationalWorkStyleNorm` --
+   but assert **no instance binding**. Binding a major group to a
+   characteristic requires importing a versioned released source
+   profile (for example an O*NET database release) with provenance in
+   its own future decision; inventing per-family profiles here would
+   fabricate evidence.
+6. Like ADR 0232, none of these concepts carries `:lookupCode`; ranks
+   and levels are scale positions from published tables and are never
+   used as weights; every IRI is minted in the canonical
+   repository-case namespace (ADR 0207).
+7. The application read model lives in `lineageweave/io_taxonomy.py`:
+   cached, deterministically sorted records for each scheme;
+   well-formed-key lookups that return ``None`` for genuinely
+   undeclared codes or levels (honest unknown); fail-closed
+   ``ValueError`` for malformed keys, malformed TTL declarations,
+   neighbors outside the closed RIASEC vocabulary, or a type without
+   exactly two neighbors.
+8. Each concept scheme names its source entities through
+   `prov:wasDerivedFrom`. Source entities retain title, publisher or creator,
+   explicit release/version, source URL, and applicable rights or license.
+   `:sourceArtifactSha256` is present only when an exact stable artifact was
+   downloaded and hashed; the O*NET 31.0 Job Zone JSON is pinned to SHA-256
+   `f66d665a2e507c825a71aedb2c13ba22765e8259bc6c7fe5b3cdfd8105475a66`.
+   A dynamic page without a reproducible artifact carries no invented digest.
+
+## Consequences
+
+- Job families, job zones, and the full published
+  source-native characteristic-family vocabulary become addressable and
+  citable inside the semantic layer before any persistence decision exists.
+- A classification never supports inferring an individual's cognition,
+  affect, personality, behavior, competence, suitability, or job
+  performance. Those uses require their own intended-use validity and
+  fairness evidence and are outside this decision.
+- Per-major-group characteristic profiles remain deliberately absent:
+  the derivation properties make their future shape typed and
+  addressable without asserting anything the sources do not state at
+  this granularity.
+- Adding DB-backed profile bindings later means one migration plus
+  provenance-bearing import of a released source database -- the
+  extension path is additive by construction, mirroring ADR 0232.
+- `tests/test_io_taxonomy.py` pins the published titles, counts,
+  adjacency structure, closed vocabularies, and fail-closed lookups, so
+  drift toward invented constructs fails CI.
+
+## Verification
+
+- `tests/test_io_taxonomy.py`: completeness (23 groups, 4 zones, 6
+  types, 6 legacy clusters, 7 style dimensions, 4 ability domains), verbatim
+  official titles, code-shape validation, published hexagon adjacency,
+  deterministic ordering, canonical namespace, lookup round-trip
+  isolation, and fail-closed lookups.
+- `tests/test_ontology.py` continues to pass unchanged: the round trip
+  sees no new lookup codes.
+- Source-provenance tests require every new scheme to resolve to the declared
+  PROV entity and verify O*NET version, publisher, CC BY 4.0 license, artifact
+  digest, and SOC version/publisher/rights metadata.
+
+## References
+
+Fleishman, E. A., & Quaintance, M. K. (1984). *Taxonomies of human
+performance: The description of human tasks*. Academic Press.
+
+Hogan, J., & Holland, B. (2003). Using theory to evaluate personality
+and job-performance relations: A socioanalytic interpretation.
+*Journal of Applied Psychology, 88*(1), 100-112.
+https://doi.org/10.1037/0021-9010.88.1.100
+
+Holland, J. L. (1997). *Making vocational choices: A theory of
+vocational personalities and work environments* (3rd ed.). Psychological
+Assessment Resources.
+
+Peterson, N. G., Mumford, M. D., Borman, W. C., Jeanneret, P. R.,
+Fleishman, E. A., Levin, K. Y., Campion, M. A., Mayfield, M. S.,
+Morgeson, F. P., Pearlman, K., Gowing, M. K., Lancaster, A. R., Silver,
+M. B., & Dye, D. M. (2001). Understanding work using the Occupational
+Information Network (O*NET): Implications for practice and research.
+*Personnel Psychology, 54*(2), 451-492.
+https://doi.org/10.1111/j.1744-6570.2001.tb00098.x
+
+National Center for O*NET Development. (2024). *Revisiting the work
+styles domain of the O*NET content model* (updated May 2026).
+https://www.onetcenter.org/reports/Work_Styles_New.html
+
+National Center for O*NET Development. (2026). *Job zone reference:
+O*NET 31.0 database*.
+https://www.onetcenter.org/dictionary/31.0/json/job_zone_reference.html
+
+U.S. Department of Labor. (2018). *2018 Standard Occupational
+Classification System*. Bureau of Labor Statistics.
+https://www.bls.gov/soc/

docs/adr/0252-complete-2018-soc-hierarchy.md

@@ -0,0 +1,50 @@
+# ADR 0252: Complete 2018 SOC hierarchy as a generated ontology fragment
+
+**Status:** Accepted  
+**Date:** 2026-08-27
+
+## Context
+
+ADR 0245 publishes only the 23 SOC major groups. That is insufficient for
+occupation-level evidence: the official 2018 SOC contains four aggregation
+levels and 1,447 classifications. A label-derived parent or a locally invented
+job-family crosswalk would violate the repository's evidence boundary.
+
+## Decision
+
+1. Import the complete official 2018 SOC structure: 23 major groups, 98 minor
+   groups, 459 broad occupations, and 867 detailed occupations.
+2. Preserve the source row's level and parent exactly. Publish `skos:broader`
+   only from that parent column; never derive hierarchy from code digits or
+   titles.
+3. Keep the normalized source snapshot at
+   `docs/ontology/data/soc-2018-structure.csv` and generate
+   `docs/ontology/soc-2018-structure.ttl` deterministically. The source XLSX
+   SHA-256 is
+   `ade08af40923266f3a854842e888ca3e93c15b26a147c20a2b12a61f4c4f4077`;
+   the normalized CSV SHA-256 is
+   `7de1c9d4da14d8eeb95197974d9dc1989752ebda235dd234b1693f336891f68e`.
+4. Treat SOC as a statistical occupational classification, not an employer's
+   job family, job series, position, person trait, or psychometric score. Those
+   bindings require separately authorized source assertions.
+5. Runtime and publication loaders merge the governed Turtle fragments into
+   one graph. The public artifact remains one canonical ontology namespace and
+   is serialized from that merged graph, rather than concatenating independent
+   Turtle documents. Its manifest identifies and hashes every governed input.
+
+## Consequences
+
+Occupation evidence can address every official 2018 SOC level without an
+invented mapping. The generated fragment is larger, but review remains bounded
+by the pinned source digests, deterministic renderer, exact counts, parent
+closure, and graph tests.
+
+## References
+
+U.S. Bureau of Labor Statistics. (2018). *2018 Standard Occupational
+Classification system*. U.S. Department of Labor.
+https://www.bls.gov/soc/2018/
+
+U.S. Bureau of Labor Statistics. (2018). *Standard Occupational
+Classification and coding structure, 2018 SOC*. U.S. Department of Labor.
+https://www.bls.gov/soc/2018/soc_2018_class_and_coding_structure.pdf

docs/adr/0255-onet-31-content-model-ontology.md

@@ -0,0 +1,51 @@
+# ADR 0255: O*NET 31.0 content-model semantic layer
+
+**Status:** Accepted
+**Date:** 2026-08-27
+**Extends:** ADR 0245, ADR 0248, ADR 0252
+**Supersedes:** ADR 0248 decision 2 only, solely where it prohibits publishing
+the complete O*NET vocabulary in LineageWeave IRIs
+
+## Context
+
+The high-level concepts in ADR 0245 cannot address most work-related
+cognition, affect, behavior, requirements, activities, or context. O*NET 31.0
+publishes 3,006 content-model elements with stable element IDs, names,
+descriptions, and a documented outline hierarchy encoded in those IDs.
+ADR 0248 preferred external O*NET RDF identifiers and prohibited a complete
+local vocabulary. The version-pinned Content Model Reference is instead the
+reproducible authority adopted here, so this later decision replaces only that
+publication choice. ADR 0248's evidence-bound assertions, construct
+separation, source-only cross-scheme links, and prohibitions on person-trait,
+score, weight, equivalence, and causal inference remain normative.
+
+## Decision
+
+1. Publish all 3,006 Content Model Reference rows as SKOS concepts in a
+   versioned `:onet31ContentModelScheme`. These repository-case IRIs identify
+   the pinned 31.0 snapshot; they do not assert identity or equivalence with an
+   external O*NET RDF term.
+2. Preserve `element_id`, name, and description verbatim. Derive
+   `skos:broader` only by O*NET's documented rule: remove the final
+   period-delimited outline segment. A missing parent fails generation.
+3. Type each node by the six published root domains and, where present, the 18
+   published second-level branches. These types are navigation boundaries, not
+   psychometric scales or assertions about a person.
+4. Keep Work Styles distinct from affective reactions and abilities distinct
+   from observed performance. Do not infer occupation ratings, person traits,
+   causal relations, weights, or crosswalks from this vocabulary.
+5. Pin the exact O*NET 31.0 JSON artifact SHA-256
+   `db59c30e4240931edce59310f2747f5476f058984b55f58f72c6f29faa30186f`
+   and verify deterministic Turtle reproduction.
+
+## Consequences
+
+The semantic layer can address the complete published O*NET conceptual
+vocabulary. Occupation-specific observations and the separately published
+ability/skill/style-to-activity/context linkages remain provenance-bearing
+data imports governed by later decisions.
+
+## Reference
+
+National Center for O*NET Development. (2026). *O*NET 31.0 database: Content
+Model Reference*. https://www.onetcenter.org/dictionary/31.0/json/content_model_reference.html