p-064 — Clear the Boundary-3 residue p-062 left behind¶

Status: active. Phase A (the two FRBR-axis families) has shipped. Phase B (title shapes) and Phase C (the expressionOf domain family) are open.

Problem¶

p-062 rescoped config/shapes/bffi.shape.ttl to what lkd.rdf actually declares and turned Boundary 3 on for every record. It did not clear the violations that remained. Over the 319 field-coverage probes, 10 records came out flagged:

Shape	Records	Cause
`bffi-prov:AxisLinkRangeShape`	130, 1130, 1240	A `bf:Hub` routed to `bffi:Expression` by its marcKey while something points at it with `bffi:expressionOf`, whose `rdfs:range` is `bffi:Work`.
`bffi-prov:ExpressionOfDomainShape`	1100, 1110, 1111	A `bf:Instance` asserting `bf:expressionOf`; the predicate's `rdfs:domain` is `bffi:Expression`.
`bffi-prov:WorkDomainShape`	051, 1051	`bffi:classification` on a `bffi:Item` (MARC 051) — `rdfs:domain` is `bffi:Work`.
`bffi-prov:MainTitleShape` + `bffi-prov:TitleRangeShape`	210, 1210	`bffi:mainTitle` asserted by a node that is not a `bffi:Title`.

Every one of these records converts and round-trips; the shapes are reporting genuine lkd.rdf mismatches in the emit, not broken records. Boundary 3 is non-blocking, so the cost of leaving them is noise in _validation.jsonl that hides a real regression when one lands.

What measurement showed before writing any code¶

bffi:workManifested is the Manifestation → Work link the emit actually produces: 346 of 346 Manifestations, always exactly one object. bffi:manifestationOfWork is never emitted, because route_work_split migrates bf:hasInstance to the BNode Expression first.
Owners of the four Work-domain predicates: everything sits on a Work except bffi:classification on 2 bf:Item nodes (field-coverage probes 051 / 1051, MARC 051) and bffi:genreForm on 1 bf:Instance (real fixture record 2394080). Nothing else is off-axis.
The reverse converter reads a Hub by URI fragment plus bffi:marcKey, never by rdf:type — so retyping a Hub cannot change the reconstructed MARC. All 3 affected records reconstruct byte-identical MARCXML after the retype.
The reverse converter reads Work-domain predicates through ?m bffi:workManifested ?work, so an off-axis term is not just a shape violation: it is silently dropped from the reconstructed MARC.

Phase A — the two FRBR-axis families (shipped)¶

Hub retype. route_hubs forces bffi:Work on any Hub that is the target of bffi:expressionOf, overriding the marcKey discriminator. Clears all 3 AxisLinkRangeShape violations; round-trip neutral (measured). The Expression signal stays on the Hub's bffi:marcKey.

Manifestation → Work lift. route_manifestation_work_domain_props moves a Work-domain predicate off a bffi:Manifestation onto the Work it manifests, resolving through bffi:workManifested → bffi:manifestationOfWork → the Expression detour, first non-empty shape winning, and lifting only when exactly one Work resolves. Recovers the one real off-axis genre term (record 2394080) into the reconstructed MARC.

Item classification. Not lifted. MARC 051 is the LC class number of one specific copy; asserting it as the Work's classification would make the reverse direction emit a MARC 050 the source never had. lkd.rdf declares bffi:classification with rdfs:domain bffi:Work and offers no item-level alternative, so bffi:classification moves to its own bffi-prov:ClassificationDomainShape, which accepts bffi:Work or bffi:Item. This is an ontology gap to raise with NLF — see the gap note in docs/bf_to_bffi_mapping.md.

Result: 10 flagged records → 5.

Phase B — the title shapes (open)¶

Records 210 / 1210 assert bffi:mainTitle on a node that is not a bffi:Title. Not yet diagnosed: establish which marc2bibframe2 shape produces it (MARC 210 abbreviated title is the likely source) before deciding between a routing and a shape rescope.

Phase C — `bf:Instance` asserting `bf:expressionOf` (open)¶

Records 1100 / 1110 / 1111. The mirror image of the Hub case: here the subject is wrong for the predicate's domain, and the fix is a triple swap or an axis-default pick rather than a retype. Needs the same round-trip-neutrality measurement the Hub retype got before it ships.

Out of scope¶

The emit is not byte-deterministic — two runs of identical code differ for 280 of 319 records, because route_work_split mints BNodes whose labels vary per run and the Turtle serializer orders anonymous blocks by label. That contradicts the idempotency rule in CLAUDE.md and is worth its own plan; it is not a Boundary-3 problem.