C. Relationship Catalog

Built from the actual corpus, not from speculation. Relationships the corpus does not encode are listed in C.4 as absent, not invented.

C.1 Canonical relationships (encoded in the research itself)

OBSERVED total: 878 canonical relationship instances resolve across the corpus (40 + 6 + 4 + 520 + 312 minus overlap). The current publisher derives 6.

C.2 Value kinds inside related_documents — a parser constraint

OBSERVED, resolving all 63 values across both spellings:

INFERENCE: related_documents is a heterogeneous union, not a list of references. A vNext parser must classify each value — path (two bases), ID, or prose — and must model "this value is not a link" as a legitimate outcome, not an error. Treating the field as a list of IDs (what the current engine does) discards 40 of 63 values.

C.3 references is bibliography, not linkage

OBSERVED, 16 files:

references:
- Shannon, C. E. (1948), A Mathematical Theory of Communication
- Legge et al. (1997-2007), visual span and reading research

INFERENCE: any parser that treats references as internal references — a natural reading of the name — would generate 100% false relationships. This is recorded explicitly so the vNext contract does not repeat the mistake.

C.4 Relationships the corpus does NOT encode

Per §9 and §28, these are not introduced into the canonical model:

• finding → evidence (no finding is a distinct object; see B.6)
• evidence → finding
• hypothesis → experiment → result
• contradicts / refutes
• theory → supporting findings

OBSERVED: RP-VE-2026-0001 carries evidenceIds, hypothesisIds, theoryIds — but only in the 12-file research-publisher seed corpus, not in the 117-file VE research corpus, where those keys appear zero times.

OPEN QUESTION: whether ROS intends evidenceIds/hypothesisIds/ theoryIds to become the canonical typed-link vocabulary. They are in the engine and in the seed corpus but absent from real research. This should be decided by the ROS owner, not inferred by the publisher.

C.5 DERIVED PUBLICATION RELATIONSHIPS

Deterministically computable, useful for navigation, not canonical research. Each must be labelled as derived in the published output.

C.6 Cardinality, cycles, dangling

• Cardinality matters for originates (1 document → many frontier records; observed max 104 documents producing 520 records) and for source_rep (many artifacts → 1 REP).
• Cycles are legal and present: web-component-framework-architecture-recommendation-v1.md and web-component-framework-adrs-v1.md reference each other through related_documents. A vNext graph walker must be cycle-safe.
• Dangling targets are legal in authored research (9 unresolved of 63) and must produce a warning, never a build failure — these are real research documents referencing work that does not exist yet.

C.7 Integrity findings (§10)

Classified per §10's required distinction.

Invalid ROS — none found

./ros validate passes on the research-publisher corpus. No duplicate IDs anywhere (49/49 unique authored, 520 unique frontier).

Valid but suspicious — warn, do not block

Legitimate research states — never flag

• supersedes: null / superseded_by: null — nothing has been superseded yet.
• status: computational-pilot-complete-human-study-pending — an honest statement of an incomplete research programme.
• 520 frontier records with status: Open — open questions are the point.
• A REP with no experiment report yet.

Cross-corpus integrity defect found during this discovery

OBSERVED: research/packages/RP-VE-2026-0001--research-publisher.md in this repository declares relatedDocuments: [DF-2026-001, HY-VE-2026-0001]. DF-2026-001 does not exist — the corpus migration renamed CN-2026-001 to DF-VE-2026-0001. The reference was not updated, and neither the publisher (which silently drops unresolvable edges) nor ros validate reported it.

This is a concrete instance of weakness W6 and is exactly the class of defect the vNext integrity view should surface.