What the four Vedas share

A METHOD CENSUS, not a finding about the corpus. The distinction is the whole point of this query: the literal/semantic partition of cross-Veda relatedness in this graph is 100/0, and read as a finding that would say Vedic cross-corpus relatedness is purely textual. It says nothing of the kind. It says no semantic-resemblance measure was ever built, which is why the semantic row is typed NOT_BUILT rather than returned as a zero beside the built populations. PARALLEL_TO is deliberately absent: all 69 of its edges are within a single Veda, so it does not enter a cross-Veda partition at all. SHARES_ENTITY_VOCABULARY_WITH is listed under its own kind and NOT as semantic resemblance -- see entity_vocabulary_overlap_candidates.

Cells are typed, not merely counted. 26 of 48 cells carry a measured count; the rest are null with a status, because a class that never reached a pair and a pair the texts do not share are different answers and this matrix refuses to render both as 0.

Do not sum a pair's cells into a relatedness score. An exact parallel, a directed reuse and a shared entity vocabulary assert different things, and the pairwise totals shown (AV-RV 2,812, RV-SV 3,789, RV-YV 953, AV-SV 551, AV-YV 346, SV-YV 272) are the sum of the measured cells only.

THIS QUERY DOES NOT ANSWER THE CONCEPTUAL-SIMILARITY QUESTION, and the `conceptual_similarity` column says so in every row rather than leaving it to this caveat. What it measures is entity-vocabulary overlap: both passages mention the same registry entities. No non-lexical resemblance measure exists anywhere in this graph -- no embedding, no vector index, no asserted resemblance -- so `semantic_resemblance_population` is NOT_BUILT, and that structural zero is typed rather than returned as a 0 a reader would read as 'no such resemblance in the corpus'. It was renamed from `conceptually_similar_not_reused` in V3.1. Under the old name it returned 25 confident-looking cross-Veda pairs ranked on a raw shared-entity count, and the V3.1 benchmark diagnosis graded Q22 and Q49 MISLEADING for it: both frozen criteria exclude lexical overlap in terms, and a caveat is not sufficient because the reader who runs the obvious query never sees the caveat. RANK ON `distinctiveness`, NOT ON `shared_entities`. Three shared entities are weak evidence when they are heaven, sacrifice and soma, which between them touch a large share of the corpus, and strong evidence when one is altar (vedi), which appears in 17 passages. `distinctiveness` is the inverse-document-frequency sum over the shared entities and `rarest_shared_passages` is the document frequency of the rarest one. The layer is materialised, which is also why it is fast: computed online it was an exact all-pairs self-join over a hub-skewed degree distribution (maximum 1,206) taking 5.3 s, and it exceeded the 1.4 GiB transaction memory limit and died outright as soon as the shared entities were collected. Pairs already joined by EXACT_PARALLEL_OF, NEAR_PARALLEL_OF or REUSES_TEXT_FROM are excluded, because the questions ask for resemblance WITHOUT shared text.

SemanticAssertion is one label over five derivations of unequal strength that must not be summed: 28,370 MORPHOLOGY_RULE_PREDICATE_ONLY and 2,406 MORPHOLOGY_RULE (derived by rule from the Sanskrit lemma annotation), 1,532 TREEBANK_DEPREL, 364 CROSS_VEDA_TEXT_IDENTITY (projected from a letter-identical Rigvedic verse rather than annotated in their own corpus) and 2,459 MODEL_EXTRACTION (TIER_D, unreviewed model output over a 19th-century English translation). The layer reaches all four corpora unevenly -- RV 27,057, AV 6,167, YV 1,543, SV 364 -- and none of the 35,131 has been reviewed by a human. Their reach differs as much as their strength, so a blended count would read as one instrument where there are five. This caveat previously ended 'All 4,865 are Rigvedic', which was true of an earlier state of the layer and became a blended total of exactly the kind it exists to refuse.

PARALLEL GRANULARITY: PADA 4,079 edges over RV, VERSE 6,907 edges over RV/AV/SV. The quarter-verse (PADA) layer reaches the Rigveda only. Pada tags come from the Rigvedic token annotation and no other corpus in this graph carries a quarter-verse division, so its single-corpus reach is a property of the annotation and not a finding about the other three Samhitas. It is typed PADA and the verse layer is typed VERSE, so the two are never summed by accident.

This matrix is the cross-corpus population only, and it is not the whole parallel layer: 325 parallel edges join two passages inside a single corpus and carry no corpus pair at all (EXACT_PARALLEL_OF 256, PARALLEL_TO 69). They are excluded from every cell here and reported per class in relationship_classes.within_one_veda_edges, so a cross-corpus share computed from this matrix alone would overstate itself by that much.