Evidence page · machine-counted

Śūdra appears once in the Rigveda. It appears 1,053 times in the law books.

If caste were the ancient, eternal order of Indian society, its vocabulary would be distributed evenly through the literature. It is not. Counted across 480 Sanskrit texts, the words of exclusion are nearly absent from the oldest layer and become dense in the legal one — and the untouchable category enters later still. This page publishes the counts, the layers, and the limits of the method, so the result can be checked rather than believed.

The curve.

occurrences per 100,000 words, by textual layer

Counted across 758 texts, 18.8 million words, searching each term together with its vṛddhi and orthographic variants. Raw counts are given beneath.

The śūdra rate rises more than a hundredfold from the oldest layer to the legal one. Not because society was newly stratified in the first millennium CE, but because a body of literature appeared whose subject is rank — and which had to name the ranked in order to regulate them.

The first appearance of the outcaste.

caṇḍāla across the whole corpus

The strongest single result concerns caṇḍāla, the term for the outcaste. Of 480 texts scanned, 132 contain it. The Vedic section contains almost none.

The one Vedic occurrence, read.

ŚB 14.7.1.22 — where the category is dissolved

A count is not a reading. The single occurrence of the outcaste term anywhere in the Śatapatha was located and checked directly in the file. It falls in book 14 — which is the Bṛhadāraṇyaka Upaniṣad, the latest layer of the text — and it does not do what a reader might expect.

The context is the description of the liberated or deep-sleeping self, in which social and moral categories cease to apply. The word's only appearance in the entire Śatapatha is in a sentence stating that the category does not hold.

What the Atharvaveda adds.

the first shift

The Atharvaveda — here the Paippalāda recension, 126,459 words — is where the vocabulary of social ranking first becomes noticeable. Set against the Rigveda:

Two things are worth noticing together. Varṇa is flat — it is present from the start, because it is an ordinary word (and, as this platform documents, it means colour before it means rank). But śūdra, the name of a ranked class, and strī, the marked category of woman, both rise sharply in the same text. The Atharvaveda is where the literature begins to sort people. And jāti — the birth-group that eventually becomes the operative unit of caste — is absent from every Vedic Saṃhitā in this corpus, arriving only later. See jātization.

Jāti, with the rebirth sense removed.

a figure that was unusable, now measured

The raw count for jāti — "birth" — cannot be read as caste vocabulary, because in Buddhist Sanskrit it is also a core doctrinal term meaning birth as rebirth, a link in dependent origination. The raw Buddhist rate of 58 per 100,000 words says nothing about social rank.

Each of the 3,273 occurrences across three layers was therefore classified by the words around it: doctrinal where the context carries the vocabulary of dependent origination (jarāmaraṇa, bhava, upādāna, saṃsāra, cyuti), social where it carries the vocabulary of rank (brāhmaṇa, śūdra, varṇa, kula, gotra, saṃkara), and unclear where neither dominates.

Note the Dharmaśāstra row first. Only 12% of its jāti occurrences are doctrinal. The law books use the word almost entirely for social rank — while in the Buddhist corpus 55% of occurrences are the rebirth doctrine.

Limits.

how this could be wrong

• Two terms are excluded from the argument as contaminated. Niṣāda is also the name of the seventh note of the Indian musical scale, and its apparent Upaniṣadic peak (13.3 per 100k) is unsafe without checking each hit. Varṇa also means colour, written letter, and praise — its Buddhist figure of 138.6 per 100k almost certainly reflects the "syllable" sense. Both are shown in the underlying scan file but neither carries any claim here.
• Transliteration variants determine the result. Searching caṇḍāla without the vṛddhi form cāṇḍāla produces a false zero for the Śatapatha — a difference of one occurrence that changes the reading entirely. Any term here may carry variants not searched: the counts are a floor, not a ceiling, and a zero is the most fragile figure a scan can produce.
• String matching, not sense disambiguation. These are counts of matched forms. Some will be false positives, and in the Buddhist case the jāti count is known to be dominated by a different sense. Where the platform has disambiguated — as with the Taittirīya scan — it says so; here it has not.
• Commentaries inflate old texts. The Chāndogya file includes Śaṅkara's commentary, roughly a millennium later than the Upaniṣad. Its 4 caṇḍāla hits cannot be assigned to the Upaniṣad itself without checking each one.
• The Atharvaveda here is the Paippalāda recension. The Śaunaka recension — the more commonly cited one — is not in this corpus, and the two differ in content and arrangement. The result should be treated as evidence about the Paippalāda, not about "the Atharvaveda" in general.
• Layer boundaries are approximate. Sections follow the GRETIL directory structure, which is a filing convention rather than a chronology. "Vedāṅga / Sūtra" and "Upaniṣad" in particular span wide date ranges.
• Sample sizes are uneven. One Āraṇyaka file and one Arthaśāstra file carry whole rows. The Buddhist row rests on 118 files. Densities from small samples are unstable and should not be over-read.
• The corpus is what has been digitised. GRETIL is large but not complete, and its composition reflects the priorities of Sanskrit digitisation — which are themselves a subject this platform treats as evidence.
• Absence is detection limit, not proof. Zero occurrences of a term mean the term is not in these files. It does not prove the social reality was absent — only that the literature did not name it.

Reproduce this.

corpus, query, gap