The material world

Lexical mentions only.

Lexical mentions on Sanskrit evidence. A mention is not a claim that the passage is about agriculture.

Every metal is returned against every Veda, so a cell that found nothing says NO_LEXICAL_MATCH in its own row rather than going missing: absence here is a fact about the matcher, never about the text. That distinction is load-bearing and there is a known false cell. The Yajurveda names ayas at VSM 18.13, where the Devanagari source writes it with avagraha elision (`me `yas ca me`) and it folds to the token `yasca`, which is the relative pronoun in 11 of its 12 corpus occurrences; registering it would land 11 wrong-sense mentions, so ayas reads NO_LEXICAL_MATCH for YV although that verse names it. `ayo` was rejected on the same precision test -- 6 token hits, of which only about half are the metal. Aliases are whole-word inflections only: the bare stem `ayas` substring-matches 645 times inside payasa/madayasva, and the bare `syama` token-matches a verb form at RV 6.5.7 and sits inside asyama 22 times, so only the attested `syamam`/`syamam` forms are admitted. Counts are lower bounds. `per_1000_mantras` is the figure to compare Vedas on, and it reorders the raw counts: gold is RV 38 against AV 34 raw but 3.601 against 5.823 normalised.

There is deliberately no Sarasvati RIVER node: its forms match 172 mantras of which the great majority are the goddess, and a river node claiming them would assert 'this passage is about a river' of passages about a deity. Ancient names are not mapped to modern identifications.

Five tribes, 42 mentions. Published separately from rivers_and_tribes because the join of the two is empty: the tribes are attested and the rivers are attested, and they never share a mantra in this corpus.

The lexical mention layer reaches all four corpora, so a corpus with no figure had no registered alias match rather than no coverage. A null is still not a zero: alias recall is partial and measured per entity, not assumed.

Every metal is returned against every Veda, so a cell that found nothing says NO_LEXICAL_MATCH in its own row rather than going missing: absence here is a fact about the matcher, never about the text. That distinction is load-bearing and there is a known false cell. The Yajurveda names ayas at VSM 18.13, where the Devanagari source writes it with avagraha elision (`me `yas ca me`) and it folds to the token `yasca`, which is the relative pronoun in 11 of its 12 corpus occurrences; registering it would land 11 wrong-sense mentions, so ayas reads NO_LEXICAL_MATCH for YV although that verse names it. `ayo` was rejected on the same precision test -- 6 token hits, of which only about half are the metal. Aliases are whole-word inflections only: the bare stem `ayas` substring-matches 645 times inside payasa/madayasva, and the bare `syama` token-matches a verb form at RV 6.5.7 and sits inside asyama 22 times, so only the attested `syamam`/`syamam` forms are admitted. Counts are lower bounds. `per_1000_mantras` is the figure to compare Vedas on, and it reorders the raw counts: gold is RV 38 against AV 34 raw but 3.601 against 5.823 normalised.

Every registered metal is returned against every corpus, so no cell is omitted. A cell with no lexical match carries a null count and NO_LEXICAL_MATCH rather than 0: the registry admits attested whole-word inflections only, so the miss is a fact about the matcher.

Known-wrong cells, typed in their own rows and not only here: metal (ayas) in YV, attested at VSM 18.13. Read each cell's source_witness before concluding an absence.

Normalising reorders gold (hiraṇya), lead (sīsa): each of those rows ranks corpora differently raw than per 1,000 mantras, and the row states both orderings.