Ægir brief: holdout-target / train-reference sample pair

Atelier classifies a holdout target sample. NHSVM (and MaxSim enrichment) must train on a disjoint reference whose BFO/CCO genus IRIs cover that target. No shared collection slugs. When the reference SKOS is a superset of the target’s SKOS (skos_missing empty), perfect_possible is true and a DST run on the target can converge.

Accept gate (do not invent a second checker): Atelier atelier.sdg.pair.assert_pair(target_id, reference_id) in /home/rch/local/src/zndx/atelier.

Identifiers

Atelier can already select a genus-covering disjoint slice from the same pin (just sdg-sample-reference <target-id>). That is not enough for perfect NHSVM. Ægir must emit collections from finepdfs + SchemaPile that realize the emit list below without using the two target slugs.

Required genus IRIs (#AT.00000021.PAIRIRI)

Reference terms[].genus union must cover:

bfo:0000002
bfo:0000003
bfo:0000004
bfo:0000015
bfo:0000016
bfo:0000023
bfo:0000040
cco:ont00000958
cco:ont00000995

(bfo:0000003 is unmapped to a SKOS upper anchor on the position collection; still required on the reference.)

Same-pin disjoint collections that already cover these IRIs (do not reuse as the Ægir product; they are the genus-layer fallback only): drug-application-0614fd04, webinar-session-927f0888, faculty-researcher-ad473424, inspection-event-9fdc6d58.

Emit list (skos_missing / vocabulary gaps)

Until a reference sample exists, treat the target vocabulary_gaps as the emit list. After assert_pair, replace this list with live PairReport.skos_missing (target SKOS codes absent from reference annotations.csv).

Entity gaps (need SKOS terms + disjoint realizing collections):

• computational-material-record-044e3d9a: ComputationalMaterialRecord, ComputationalCode, MaterialSystem, MaterialProperty, ParsingPipeline, HighPerformanceComputerCenter, BigDataAnalyticsTool, ParsedDataStore, VisualizationService
• position-ac7bde24: Position, Employee, Department, CollectiveBargainingAgreement, PerformanceReview

Column gaps (need SKOS SDG.DOM.* or entity properties, not new target tables):

active_status, algorithm_type, approval_date, center_name, checksum, chemical_formula, classification, code, code_name, comments, confidence_interval, created_at, creation_date, crystal_system, data_integrity_status, developer_group, employment_status, end_date, end_date_time, endpoint_url, established_date, expiration_date, file_format, file_size_bytes, first_name, formation_energy, hire_date, last_name, last_updated, lattice_parameter_a, lattice_parameter_b, lattice_parameter_c, license_type, max_concurrent_users, measurement_method, name, novel_materials_discovered, numeric_value, operational_status, peak_performance_flops, pipeline_name, property_name, rating, records_processed, review_date, salary_range, service_name, space_group_number, start_date, start_date_time, status, storage_size_g_b, store_name, termination_date, title, tool_name, total_compute_nodes, total_records, union_name, unit_of_measure, updated_at, uptime_percentage, version

Target SKOS that a reference must include (superset; 24 codes):

SDG.ARTIFACT, SDG.DISPOSITION, SDG.DOM.BIRTHPLACE, SDG.DOM.COUNTRY, SDG.DOM.CREATED_DATE, SDG.DOM.DATE_PUBLISHED, SDG.DOM.EFFECTIVE_DATE, SDG.DOM.GEOGRAPHIC_PROVENANCE, SDG.DOM.HOME, SDG.DOM.ISSUED_DATE, SDG.DOM.LOCATION, SDG.DOM.PUBLISHED_ON_DATE, SDG.DOM.REGIONAL, SDG.DOM.SINCE, SDG.DOM.STATE, SDG.DOM.TEMPORAL_DESCRIPTOR, SDG.GDC, SDG.GENERIC, SDG.ICE, SDG.ICE.DESCRIPTIVE, SDG.INDEPENDENT_CONTINUANT, SDG.MATERIAL_ENTITY, SDG.PROCESS, SDG.ROLE

Disjointness (#AT.00000020.PAIROVERLAP)

Reference collections must not include computational-material-record-044e3d9a or position-ac7bde24. No shared manifest.collections[].slug. Do not copy target CSVs into the reference.

How Atelier will assert_pair

from atelier.sdg.pair import assert_pair
report = assert_pair(
    "sdg-corpora/b24ef9f60660_macbook",
    "sdg-corpora/b24ef9f60660_reference",
)
# report.perfect_possible  → DST on the target is allowed
# report.skos_missing      → emit list (must be empty for perfect NHSVM)

Classification invoke after accept:

just classify-flow --host \
  --source-id=sdg-corpora/b24ef9f60660_macbook \
  --reference-id=sdg-corpora/b24ef9f60660_reference \
  --target=phase:nhsvm

NHSVM trains on the reference; the target is the holdout. Do not train on the target. Pin must stay b24ef9f6….

Deliverable shape

A sample directory Atelier can load as sdg-corpora/b24ef9f60660_reference: manifest.json (collections with slug + genera), annotations.csv, tables/*.csv, RI-closed bundles. Land under Atelier build/sdg_sample/b24ef9f60660_reference or emit in sdg-corpora on this pin and let just sdg-sample-reference pack it.

Sources: finepdfs (scripts/emit_finepdfs_postings.py) and SchemaPile (scripts/emit_schemapile_postings.py / build_collections.py). Do not use GitTables fixture lanes.

Status (2026-09-16)

Reference sample is on disk. Accept gate:

perfect_possible True
skos_missing ()

Live PairReport.skos_missing is empty. The emit list above is the vocabulary that was driven (entity + column gaps authored as SDG.DOM.* and realized on the two collections). Do not train on the target.