What this edition covers, how each kind of coverage is established, and where it stops. Every boundary recorded here is a fact about this corpus, never about the Vedas.
This page exists so that the rest of the product does not have to argue. Every other surface states its scope in a clause and links here; everything the reader needs in order to judge whether a figure can support their question is set out below, in one place, at length.
The corpus
20,210 canonical mantras across four Saṃhitās, one recension each. No Brāhmaṇa, Āraṇyaka or Upaniṣad is held, and three of the four Saṃhitās are themselves partial in ways their ordinary names do not reveal. Each row below states the recension held and enumerates what that edition does not address.
Śaunaka recension, Saṃhitā onlyHeld · 5,839 mantras, each with a canonical key
The Paippalāda recensionNot held · another recension
The Gopatha BrāhmaṇaNot held · a Brāhmaṇa
The UpaniṣadsNot held · a later layer
A solid mark is what this build holds. A hollow one is a corpus the registry records as named and outside this edition, so an absence measured here is an absence from the held recension only. The order is the registry’s; nothing on this page asserts that any recension descends from any other.
The two exclusions most likely to mislead
The Yajurveda. In ordinary use the name covers both the White and the Black recensions. This edition holds the White one only, so no Taittirīya, Kāṭhaka, Maitrāyaṇī or Kapiṣṭhala material is present at all, and the prose Brāhmaṇa passages interleaved with the mantras in the Black recensions have no counterpart here.
The Sāmavedic gāna. The song collections are a parallel and larger body than the ārcika verse corpus held here, and they are the reason the Samaveda is a distinct Veda rather than a Rigvedic excerpt. Any claim that this product “has the Samaveda” while holding the ārcika alone would be false, and the ārcika is what is held.
Khila additions are outside the canonical counts, and an absence measured in any corpus here is an absence from the recension held, never from the tradition.
Translation coverage
Coverage is typed by provenance rather than counted as one number, because the kinds are not interchangeable: a rendering borrowed from a parallel verse in another corpus is translation assistance and not evidence about the corpus it appears beside.
Dedicated English translations
18,145
Multi-verse range translations
128
Reused parallel renderings
194
Non-English scholarly notes
24
Verses with no English
1,719
Every translation here is aligned, not reviewed
The English in this corpus is public-domain scholarly translation matched to a canonical key. Coverage measures that alignment; it does not measure translation quality, and no rendering here has been checked against the Sanskrit by this project.
The Samaveda has no English of its own
Zero of its 1,844 verses carry a dedicated translation. 173 verses show a published Rigvedic rendering against text verified character-identical, disclosed as reuse and never counted as Samavedic English; 1,671 verses carry none. The consequence runs further than the count: every layer derived from a corpus’s English translation is absent for the Samaveda rather than empty in it, so a zero returned from such a layer is a statement about this build and not about the text.
Measured live on this request: 18,427 own-corpus English renderings and 194 reused renderings, reported apart and never summed.
Recitation
Audio is a product layer over the text rather than part of the knowledge model. No recording is republished here: each is streamed from its source at the moment it is played, and a recording whose source numbers its verses by a different recension is rejected rather than approximately aligned.
Catalogued recordings
17,780
Rigveda
10,552
Samaveda
none
Yajurveda
1,777
Atharvaveda
5,451
Verified by ear
20
A checked mapping is not an audible review
This is the distinction to carry away from this section. All 17,780 catalogued recordings are published because their mapping was checked — the source’s own verse coordinates resolve to our canonical key, the Sanskrit matches this corpus’s text, and the media resolves — and not because anyone has listened to them. The figure above states how many carry an audible review.
What the check does rule out is the failure that matters most: a recording whose source numbers its verses by a different recension is rejected rather than approximately aligned, because playing the wrong verse is worse than playing nothing.
What has been heard
An owner sample of 20 recordings was listened to in full, with 20 verified and 0 rejected. A separate queue of 1,021 recordings is held outside the catalogue, of which 1,001 have not been listened to individually; none has been promoted into the published catalogue.
No Samavedic recording is catalogued at all
The Samaveda is the one Veda defined by its sung realisation, which makes this the most conspicuous gap in the layer — and it is a gap in what has been published here, not a statement about the tradition.
Sāmavedic notation
The ārcika carries svara marks that record how a verse was pitched. They are held as what the printed witness prints, verse by verse, and never extrapolated into sung pitches.
Verses with a validated notation witness
1,136
Verses awaiting textual alignment
708
Gāna works modelled
none
A mark is not a melody
Each released witness is a source-explicit reading: the marks as one named witness prints them, verified to map to exactly one verse. The Kauthuma decipherment authority — the tradition that would turn those marks into pitches — is not held in this build, so the tone marks are stored as codepoints and no melody is ever reconstructed from them.
The remaining 708 verses are held back because their printed run could not yet be aligned to a canonical key with certainty. The gāna song collections are outside this edition altogether.
Evidence layers, and where each one reaches
Every claim carries the layer it came from, and a weaker layer is never promoted by being displayed beside a stronger one. The four grades, and what each can carry, are described on the sources page.
Several layers reach some collections and not others
The deity ascription that resolves to a named god comes from the Anukramaṇī, and that apparatus covers the Rigveda. The Atharvaveda has its own index, which records descriptive phrases rather than registry names, so the two are not one layer. The metre layer reaches the Rigveda and the Atharvaveda and neither of the other two. Where a layer has no edge into a corpus, this product answers “not built” rather than zero.
Ascribed and named are different claims
What the traditional index assigns to a hymn is never shown as a statement the Sanskrit makes. Where a hymn’s label has been projected onto the verses inside it, the projection is marked on the row, so a per-verse claim is never built from a container’s label — most Atharvavedic seer attributions are of that kind, and the Yajurvedic ones are not.
Deity mentions are graded, and the ambiguous ones are withheld
Vedic Sanskrit has one word for the god Agni and for fire. Every mention is graded certain, probable or ambiguous; the ambiguous ones are excluded from default figures, all three counts are always reported, and a passage in which every mention is ambiguous returns insufficient evidence rather than an empty list.
Lexical counts are minima
A count of verses in which a registered alias occurs is a floor, not a census: an unregistered spelling is simply not counted. Where a cell is known to understate for this reason it is labelled rather than silently corrected. The mention layer is also not one instrument — Rigvedic mentions come from a manual scholarly lemma annotation and the other three corpora from surface matching — so comparing a Rigvedic rate with an Atharvavedic one compares two methods as well as two texts.
Non-lexical resemblance was never built
There is no embedding, no vector index and no asserted semantic similarity anywhere in this graph. Cross-corpus relatedness here is textual, and that is a fact about the method rather than a finding about the Vedas.
Nothing in the derived layers has been reviewed by a person
The assertion layer is derived by rule from scholarly annotation, with a minority extracted by a language model; none of it has been read by a human reviewer, and its coverage is uneven across the four corpora. It is usable for a single passage and not as a corpus-level comparison. The interpretive claims in the build are permanently marked as candidates and each records the observation that would falsify it.
Ask
Retrieval runs first, over a fixed catalogue of channels, and the language model sees only what retrieval returned. Every factual sentence carries a citation that resolves to a passage in this corpus.
Benchmark questions
60
Acceptable outcomes
60/60
Honestly refused
17
Misleading
0
Invented
0
What it will not do
A question this build cannot reach returns insufficient evidence and names the dimension it could not reach. It does not guess, and it does not turn a gap in the graph into a confident denial about the Vedas. Of the 60 benchmark questions, 17 were answered that way, which is the correct outcome rather than a shortfall.
The prose depends partly on the configured model
Retrieval and citation validation are deterministic and run the same way every time. The sentences between them are generated, and a different model will write different sentences over the same evidence.
Questions this build cannot answer
21 of the 21 questions the benchmark grades unanswerable are catalogued below, each with a probe that runs on this request and a question that can be answered instead.
Partly answerable
How are reused verses transformed?
Partial coverage
Why
A transformation typology exists and covers part of the layer: 6,271 of 6,907 reuse edges carry one, over a vocabulary of 15 values, and a separate formula-level typology reaches 4,368 edges. It is partial in two ways that matter. The typology describes the observable surface difference between two stored strings -- substitution, word division, accent notation -- and not the philological operation a scholar would name. And the uncovered remainder is not a residue: an edge without a type was not typed, which is not the same as an edge that was typed as unchanged.
What this is not
This is NOT the finding the frozen benchmark recorded. It graded this question NOT_ANSWERABLE on a measurement of zero typed edges, and the layer has since been built. Reading the frozen grade today would tell you the product cannot answer something it partly can. Equally, a surface-difference class is NOT a philological account of how or why a verse changed.
Measurements behind this limit
reuse edges
6,907
Every textual-reuse edge in the graph, across the five classes. The denominator for the typology's coverage.
edges with a transformation type
6,271
Reuse edges carrying a named transformation. This is the figure the frozen benchmark recorded as zero; it is not zero now.
formula edges with a transformation type
4,368
The formula layer's own, coarser typology. A second instrument over overlapping edges, never summed with the first.
transformation vocabulary size
15
Distinct transformation values. A typology with one value is a label, not a typology.
Ask this insteadThe cross-Veda relationship matrix, which reports each reuse class with its measured population and the classes that were never built at all.
Cannot be answered
Which passages are lexically different but conceptually similar?
Not yet modelled
Why
No non-lexical similarity measure exists in this graph: no embedding, no vector index, no asserted semantic resemblance. Every cross-corpus relatedness edge here is literal reuse or shared entity vocabulary, so 'conceptually similar' can only be operationalised as 'shares a word', which is what the question excludes.
What this is not
The literal-to-semantic split of cross-corpus relatedness in this graph is total, and read as a finding that would say Vedic relatedness is purely textual. It says no semantic measure was built. Entity-vocabulary overlap is offered under its own name and is not a substitute: three shared entities are weak evidence when they are heaven, sacrifice and soma.
Measurements behind this limit
semantic resemblance edges
NOT_BUILT, and therefore null rather than 0: the population does not exist. The census reports it as NOT_BUILT.
built relatedness edges
8,723
Cross-corpus edges that do exist, all of them literal reuse or entity-vocabulary overlap.
Ask this insteadEntity-vocabulary overlap ranked on distinctiveness rather than on a raw shared-entity count, with every row typing its own conceptual-similarity column as unestablished.
Cannot be answered
What deity communities emerge from the corpus?
Not yet modelled
Why
A community partition has been computed and 12 partition artifacts are stored, and not one of them has a member: 0 subjects carry a community assignment. So there is a partition and there is nothing in it, which cannot be returned as a community table without inventing the membership. Pairwise co-occurrence exists and is a different object: PAIRWISE_CO_OCCURRENCE_IS_NOT_A_COMMUNITY_PARTITION.
What this is not
This is NOT a finding that Vedic deities form no communities, and it is NOT an empty result. The membership was never built. Substituting the deity-pair table for it is how this question was graded MISLEADING: that table's top-ranked members were human patrons carried in the deity registry, presented in a deity column with nothing saying so.
Measurements behind this limit
stored community partitions
12
Community artifact nodes in the graph. Non-zero, and this is the figure that makes the older wording of this card false: a partition was computed. It carries no membership.
deities with a community assignment
0
Subjects carrying a community, louvain or partition membership. This zero is the finding: it is about the graph, not the pantheon.
eligible deities
157
A denominator, not a census. The one documented eligibility predicate, which excludes 22 human patrons, 7 danastuti gift-praise labels and the 28 ABSTRACT labels ruled NOT_DEITY per label, and keeps the 5 ruled UNDECIDED. This figure read 192 while the query excluded structure='HUMAN' alone. Read it together with devatas_without_an_eligibility_ruling below.
devatas without an eligibility ruling
0
Devata nodes carrying no is_deity ruling. While this is non-zero the eligibility layer has not landed and eligible_deities falls back to curated structure, which drops the 29 curated non-members but keeps the 28 abstractions. The transitional state is a row rather than a caveat because a reader who takes a count from a row is the reader who did not read the caveat.
pairwise co occurrence edges
306
Deity pairs that share passages. Real, and not a partition: pairs cannot be read as communities without a clustering nobody ran.
Ask this insteadDeity pairs far above chance, with lift and per-corpus counts over all four Vedas. It answers a narrower question honestly and is still pairs.
Partly answerable
Which ritual objects recur most?
Partial coverage
Why
There is a real ranking and it covers part of the question. Ritual implement means an object a modelled rite is curated to use, which makes the population small: 15 of 41 objects in the registry are linked to a rite, and 15 of those are reached by the mention layer at all. What is partial is per-implement recall rather than class membership: every count is a lexical minimum over registered whole-word aliases, and the caveat sourced 'ritual_objects_recurring' measures the shortfall for the sacrificial post and the altar. Procedure in the Samhita's own words is thinner still, at 3 step edges across 103 rites; a second layer holds 3,121 sutra-attested steps from 11 works, which are located points in independently numbered sources rather than one procedure.
What this is not
The ranking is NOT a census of Vedic ritual apparatus and a low count is NOT absence from a Veda. Two ceilings sit above it: the curation, which links 15 of 41 objects to a rite at all, and the corpus, which does not include the Brahmana and Srautasutra prose where the apparatus is actually described. The earlier version of this answer ranked the chariot and the thunderbolt first, which is why the class is now defined by use in a rite rather than by being an object.
Measurements behind this limit
rituals modelled
103
Rites with a node. A rank in the table is a rank within these.
step edges
3
Steps a Samhita text numbers in its own words, whole graph. On this evidence no rite has a recoverable full sequence.
procedure step edges
3,121
Steps a Srautasutra or Grhyasutra prints -- a different claim about a different source, so a separate measurement and never summed with step_edges. Each work numbers its own sequence, so these are located points and still not a recoverable procedure.
curated implements
15
Objects a modelled rite is curated to use. The population the ranking is drawn from.
objects in the registry
41
Every curated object, ritual or not. The difference from curated_implements is the curation ceiling: an object with no link to a rite is absent from the ranking however often the corpus names it.
implements reached by the mention layer
15
Of those, the ones any passage is measured to name. The gap is alias recall, not textual absence.
Ask this insteadThe full ritual layer view, which returns the implement ranking together with the figures that bound it and an explicit list of what it does not cover.
Cannot be answered
Where does the graph record scholarly disagreement about the Vedas?
Insufficient evidence
Why
The graph records no pair of rival scholarly readings of the Vedic text. It holds a small number of interpretive claims, only some of which are about the text at all, and no two of those contradict each other. The one contradiction it does record is cross-category -- an interpretation of a tradition against a measurement of this dataset -- and those two are in dialogue rather than in conflict.
What this is not
This is NOT a claim that the Vedas are uncontroversial. No attributed commentarial position is held by this product at all, so the absence is an absence of sources.
Measurements behind this limit
recorded contradiction pairs
1
Pairs the graph marks as contradicting. Read each pair's disagreement_kind: a cross-category pair is not scholarly dissent.
Ask this insteadThe recorded contradictions with their kind stated per row, alongside the explicit statement that none of them is a rival reading of the text.
Cannot be answered
Which mantras are central bridges between concept communities?
Not yet modelled
Why
Two independent absences, and only one of them is what the benchmark recorded. A community computation now exists -- 12 partition artifacts are stored -- but 0 subjects carry a membership in one, so there are communities without members and nothing to be a bridge between. Separately, 0 mantras carry a betweenness or bridging score and 0 metrics of that kind are stored, so even a populated partition would have no centrality to rank on.
What this is not
This is NOT a finding that no mantra bridges anything. It is also NOT the benchmark's reason any more: the benchmark said no community computation existed, and one does. The membership edges are what is missing.
Measurements behind this limit
stored community partitions
12
Community artifact nodes in the graph. Non-zero, and it is not what the question needs: a partition with no members assigns nothing.
deities assigned to a community
0
Subjects carrying a membership in one of those partitions. This zero is the first of two reasons the question cannot be answered.
mantras with a bridge score
0
Mantras carrying any betweenness or bridging score. The second reason, and independent of the first.
stored bridging metrics
0
Stored metrics of a bridging kind, under any name. Searched by name rather than assumed absent.
Ask this insteadDeity co-occurrence pairs with lift and per-corpus counts, which is a real measure over the same evidence and is pairs rather than communities.
Cannot be answered
Which cross-Veda passages express similar ideas without textual reuse?
Not yet modelled
Why
All 6,907 cross-corpus relationships in this graph are grounded in shared wording, and the question excludes exactly that. There are 0 asserted resemblance edges and 0 nodes carrying an embedding, so a non-lexical similarity can be neither retrieved nor computed. Conceptual similarity without textual reuse therefore cannot be separated from chance.
What this is not
This is NOT a finding that cross-corpus relatedness in the Vedas is purely textual. It is a statement that this graph measures only the textual kind.
Measurements behind this limit
lexical relatedness edges
6,907
Every cross-corpus relationship this graph holds. All of them rest on shared wording, which is why they cannot answer a question that excludes it.
asserted resemblance edges
0
Non-lexical resemblance asserted by any relationship. A typed zero: the measure was never built, so this is not a finding that no two verses resemble.
nodes carrying an embedding
0
Nodes with a vector representation. Zero, and there is no vector index either, so a similarity could not be computed at query time.
Ask this insteadThe cross-Veda matrix, which reports the built classes with their populations and prints SEMANTIC_RESEMBLANCE as NOT_BUILT rather than as a zero row.
Cannot be answered
What semantic roles does each corpus assign, and how do they differ?
Not yet modelled
Why
The layer reaches all four corpora -- AV 6,167, RV 27,057, SV 364, YV 1,543 -- and the question is still not answerable, for a different reason than a missing corpus. Each corpus is annotated by a different instrument (AV MORPHOLOGY_RULE_PREDICATE_ONLY, TREEBANK_DEPREL; RV MODEL_EXTRACTION, MORPHOLOGY_RULE, MORPHOLOGY_RULE_PREDICATE_ONLY; SV CROSS_VEDA_TEXT_IDENTITY; YV MORPHOLOGY_RULE_PREDICATE_ONLY, TREEBANK_DEPREL), so a measured difference between two corpora here would be a difference between annotation projects rather than between the texts. The role reading the question asks for is narrower again: 2,502 assertions carry an explicit agent and they span 1 corpus, so there is no second corpus of roles to compare. No divergence statistic exists anywhere in the graph.
What this is not
This is NOT a finding that the layer is absent outside the Rigveda -- it reaches all four corpora. Nor is a smaller figure for a corpus an absence of semantic structure in that text: it is a thinner annotation. The layer must never be aggregated across its derivations, which differ in strength, in reach and in what they were derived from.
Measurements behind this limit
assertions
35,131
The whole layer, across four corpora (AV 6,167, RV 27,057, SV 364, YV 1,543). Not one figure to be compared against another: the derivations differ per corpus, so this total is a size and not a measurement of anything.
assertions carrying an explicit agent
2,502
The subset that states who acts, which is the reading the question asks for. Rigvedic only, because it comes from a morphological annotation covering the Rigveda alone.
Ask this insteadThe deity mention layer, which reaches all four corpora, though by two different instruments that must not be averaged. For the assertion layer, read one corpus at a time and read the derivation with it.
Cannot be answered
Which concepts connect Agni and Soma by a route that is not simple lexical co-occurrence?
Not yet modelled
Why
The question requires a measure that excludes same-verse co-occurrence, and the two candidates both fail. The asserted concept-to-concept layer holds 180 relations in total, of which 66 are typed as interpretive claims: a path found through a layer of that size says more about which relations were curated than about the corpus. And there are 0 embeddings, so a second-order distributional similarity with a chance baseline cannot be computed at all.
What this is not
This is NOT a finding that Agni and Soma are unconnected. Same-verse co-occurrence connects them densely; the question asks for a route that is not that, and that route is what is missing.
Measurements behind this limit
concept to concept relations
180
Relations between concepts of any kind. Small, and the size is the point: a route through a layer this sparse is an artefact of what was curated.
curated interpretive relations
66
Of those, the ones typed as an interpretive claim rather than read from a source. A route through these is a route through an editor's judgement.
nodes carrying an embedding
0
The second-order distributional route the question allows as an alternative. Zero, so that route does not exist either.
Ask this insteadDeity co-occurrence with lift, which states plainly that it is same-verse co-occurrence and does not dress it as a conceptual route.
Cannot be answered
Which entities act as bridges between the ritual sub-graph and the cosmological sub-graph?
Not yet modelled
Why
The question presupposes two declared domains with enough members to constitute sub-graphs. There is one: 219 of 384 concepts carry a domain and 1 distinct domain is declared. A cosmological domain is not declared anywhere in the ontology, so there is one side and no second side to bridge to. Partitioning by label instead would substitute a class hierarchy for a domain, which is a different claim.
What this is not
This is NOT the benchmark's reason. It recorded that neither domain was declared; one now is. It is also NOT a finding that no entity spans ritual and cosmology -- the second domain is simply not written down.
Measurements behind this limit
concepts
384
Every concept node. The population a domain partition would have to cover.
concepts assigned to a domain
219
Concepts carrying a declared ontology domain. Non-zero, which is a change from the benchmark's reading that no domain is declared anywhere.
declared domains
1
How many distinct domains are declared. One. A bridge needs two sides, and this is the measurement that says there is one.
Ask this insteadThe ritual layer view, which reports the declared ritual domain with its membership and the ceilings on it, without inventing a counterpart.
Cannot be answered
Which lexical forms are reused across Vedas while their semantic function changes?
Not yet modelled
Why
Detecting a function change needs a function assignment that does not depend on the wording. This graph has none: all 35,131 assertions are derived from the verse's own tokens by 5 paths, so a verbatim-reused verse is assigned the same function in both corpora by construction and a change can never be observed. The layer also carries 8074 assertions outside the Rigveda, so there is no second corpus to compare against.
What this is not
This is NOT a finding that reused verses keep their function. The instrument is incapable of registering a change, which is a different statement and the one this card makes.
Measurements behind this limit
semantic assertions
35,131
Every stored semantic assertion. All of them are computed from the verse's own tokens, which is why they cannot detect a function change.
derivation paths
5
How the assertions were derived. Two paths, one rule-based and one model extraction, and neither is independent of the wording.
assertions outside the rigveda
8,074
The layer's reach. A cross-corpus comparison needs both sides, and this zero means only one side exists.
Ask this insteadThe reuse layer itself with its per-pair evidence quotes, which shows what changed in the wording without claiming anything about function.
Cannot be answered
Which passage pairs have the strongest evidence of directional textual reuse?
Not yet modelled
Why
Direction exists for 1,684 of 6,907 reuse edges, and it is one prior applied wholesale rather than per-pair evidence: the Kauthuma arcika is an arrangement of Rigvedic verses and its own tradition names the Rigveda as the source, so every edge in that pair inherits the same direction from the same sentence. The criterion states that a single global prior applied to one Veda pair is not per-pair evidence, which is exactly what this is. The remaining edges are explicitly typed symmetric, which is the honest form and not a weaker direction.
What this is not
The directed edges are NOT wrong and NOT to be discarded: the Samavedic borrowing is well established. They are simply not per-pair evidence, and ranking pairs by 'strength of directional evidence' over them would rank one sentence about a tradition six hundred times.
Measurements behind this limit
reuse edges
6,907
Every reuse edge. The denominator for how far direction reaches.
edges stored as directed
1,684
Edges stored with a direction. Real, and all of them inherit one tradition-level prior about one corpus pair rather than carrying per-pair evidence.
distinct direction bases
3
How many distinct grounds for direction are recorded. Two: one prior, and an explicit statement that the rest are symmetric and must not be read as a claim about which text came first.
Ask this insteadThe reuse layer as stored, where a symmetric pair is typed symmetric and the one directed corpus pair states the prior it rests on.
Cannot be answered
Which Devatas are associated with which specific offerings, on per-verse source-stated evidence?
Not yet modelled
Why
There is no deity-to-offering relation in the sense the question needs. The 14 edges that exist are identity links between a devata slot and the substance it names, and 0 of them carry a per-verse textual span -- their evidence is a lexicon locator. The tempting substitute, the 2,154 verses that name a deity and an offering together, is co-occurrence: it would rank Indra with soma because both appear in the same hymns, not because a verse says soma is offered to Indra.
What this is not
This is NOT a finding that the Vedas do not assign offerings to deities -- they do so constantly. It is a statement that this graph has no edge that records it.
Measurements behind this limit
deity to offering edges
14
Edges from a deity to an offering. Non-zero, and every one of them is an identity link -- Soma to soma, the cow to cattle -- where the devata slot and the substance are the same thing, not a record of who receives what.
edges with per verse evidence
0
Of those, the ones whose evidence is a span in a particular verse. Their evidence is a concept-lexicon locator instead, which is a dictionary entry and not a passage.
passages naming both
2,154
Verses that name a deity and an offering. Large, and it is co-occurrence: naming both in one verse is not a statement that one is offered to the other, and treating it as one is how this question was answered wrongly.
Ask this insteadThe ritual layer view, which reports the offerings a modelled rite is curated to use and states the curation ceiling above them.
Cannot be answered
Which deity epithets cluster with which actions?
Not yet modelled
Why
Epithets have no occurrence layer. The graph holds 13 epithet nodes attached to deities by 13 edges, and 0 verses are annotated as containing one. Clustering epithets with actions requires knowing which verse each epithet occurs in, so that it can be joined to the verb-argument layer in that verse. The first requirement is absent outright.
What this is not
This is NOT a finding that Vedic epithets carry no action profile. Nothing was measured, because the layer the measurement would run over does not exist.
Measurements behind this limit
epithets
13
Epithet nodes in the graph. A handful, and they are a glossary rather than an annotation of the corpus.
passage level epithet occurrences
0
Verses annotated as containing an epithet. This zero is the whole answer: with no occurrences there is nothing to join to the action layer.
deity level epithet edges
13
Epithets attached to a deity rather than to a verse. This is what exists, and it says which epithets belong to whom, not where they occur.
Ask this insteadA deity's action distribution from the agentive layer, which is real, is Rigveda-only, and says nothing about epithets.
Cannot be answered
Which interpretive claims in the graph have conflicting evidence, and what exactly is the conflict?
Not yet modelled
Why
The layer holds 6 claims in total, of which 2 are about the Vedic text rather than about this dataset's construction, and 0 are attributed to a named scholarly position. The criterion needs at least a dozen attributed claims about Vedic content with typed supporting and contradicting evidence on both sides. 2 carry a contradiction at all, and those contradict a reading of this dataset.
What this is not
This is NOT a finding that Vedic scholarship agrees. It is a statement that this graph does not hold scholarly disagreement, and that the handful of claims it does hold are its own, not anyone's.
Measurements behind this limit
interpretive claims
6
Every interpretive claim in the graph. Six, which is the size of the whole layer and not a filtered subset.
claims about the vedic text
2
Of those, the ones about the corpus rather than about this dataset's own construction. The question explicitly excludes the second kind.
claims attributed to a named position
0
Claims attributed to a named scholarly position. This zero is decisive: an unattributed claim cannot be one side of a scholarly disagreement.
claims with a recorded contradiction
2
Claims carrying a typed contradiction. Real, and they contradict a reading of this dataset rather than a named scholar.
Ask this insteadThe claims as stored, each with its falsifier, read as this dataset's own positions rather than as a survey of scholarship.
Cannot be answered
How reliable are the graph's assertions, measured against a gold set?
Not yet modelled
Why
There is no human-annotated evaluation set, so there is no reliability curve and no calibration. The confidence values several predicates carry are pipeline priors and not calibrated probabilities -- for a number of predicates the value is a single constant on every edge -- so a threshold on them selects nothing. The strongest review this graph holds is a per-passage re-read by a model; nothing anywhere is human-reviewed.
What this is not
A confidence of 0.9 here is not a 90% chance of being right. It is a provenance marker. Treating it as a probability, or filtering on it, produces a selection that looks principled and is not.
Measurements behind this limit
human annotated assertions
0
The size of the gold set. This zero is why no reliability figure is published anywhere in this API.
Ask this insteadThe quality tier and evidence basis on each edge, which record how an assertion was arrived at rather than how likely it is to be true.
Cannot be answered
Which major entities have coverage too thin to support a claim about them?
Not yet modelled
Why
The graph can enumerate its 384 entities and rank them by how little evidence each carries. What it cannot do is say which *major* entities are underrepresented, because that needs a manifest of entities that ought to be present, and 0 entities carry an external expectation. The 100 curated attestation counts are counts of what was found, so an entity missing from the graph entirely is missing from the ranking too -- which is precisely the population the question is about.
What this is not
A ranking of the thinnest entities the graph holds is NOT an answer to this question, and serving it as one would be the question's own trap sprung: the entity with no node at all ranks nowhere.
Measurements behind this limit
entities in the graph
384
Every entity the graph holds. It can rank these by thinness; that is not what the question asks.
entities with a curated attestation count
100
Entities carrying a curated attestation figure. Real, and it is a count of what was found, not a statement of what should have been found.
entities with an external expectation
0
Entities carrying an externally sourced expected count. This zero is the trap the question sets: without it, 'too thin' has no denominator.
Ask this insteadThe entity listing with each entity's measured evidence count, read as a statement about the graph's coverage rather than about the corpus.
Cannot be answered
What does the Samaveda's musical dimension contain?
Not yet modelled
Why
It contains NOTATION and not song, and those are different answers. no of 1,844 Kauthuma arcika verses now carry the Kauthuma numeric svara marks their source printed on them -- 0 combining Devanagari Extended cantillation codepoints, read off one pinned revision of a community transcription and aligned to this corpus's own coordinates. 0 verses carry no notation and say why in 0-valued typed reason on the verse itself, so the absence is readable rather than inferred; coverage is NOT uniform across the four collections. What is still entirely absent is the sung realisation: the gana collections are excluded by the work's own declaration (SAMAVEDA_GRAMAGEYA_GANA, SAMAVEDA_ARANYAKAGEYA_GANA, SAMAVEDA_UHAGANA, SAMAVEDA_UHYAGANA, SECOND_SAMAVEDIC_RECENSION, SAMAVEDIC_BRAHMANA, UPANISAD) and require their own work identifier, and there are 0 verse-to-saman edges because there is no canonical saman to point at.
What this is not
The marks are NOT a melody and this API does not decipher them. They are recorded as the codepoints the source printed; no pitch, interval or svara name is derived anywhere, because the Kauthuma decipherment authority is not held and the one that is published is for a different school. A reader can see how a verse was MARKED and still cannot learn how it was SUNG. Nor is the notated count a coverage claim about the Samaveda: no Samavedic figure anywhere in this API may be read as covering the Samaveda, and the work node's own scope sentence begins by saying so in capitals.
Measurements behind this limit
verses with source supplied notation
0
Verses carrying tone marks the source printed. Validated by the structural, semantic and adversarial gates before release.
verses with notation withheld
0
Verses the witness does not settle. Published beside the positive count on purpose: without it the notated figure reads as the whole corpus.
verses with no notation disposition at all
1,844
Verses neither notated nor typed as withheld. This must be 0: an untyped verse is one a reader cannot tell apart from a verified zero.
notation rows interpreted into pitch
0
Rows converting a mark into a pitch. 0, and it is a refusal rather than a gap: no decipherment authority for this school is held.
verse to saman edges
0
Edges from a verse to the saman that realises it. 0, because the object side is outside the bounded corpus and minting one to fill a denominator is forbidden.
excluded samavedic corpora
7
Bodies the Samavedic work declares it does not hold. Every one of them is part of what a reader means by 'the Samaveda'.
Ask this insteadThe notated verse text itself, served as a parallel witness beside the primary text, and the measured Rigvedic reuse within the arcika. All three are real, and all three are about the verse rather than the song.
Partly answerable
Which crops, metals and animals occur in which Vedas and in which contexts -- ritual versus non-ritual?
Partial coverage
Why
The axis exists and it resolves for a minority of the corpus. Every one of the 20,210 verses carries a ritual-context assignment derived from external ritual citation, but only 1,903 of them are placed in a rite; the other 18,307 carry a typed absence -- no citation found, or an opening shared with rival verses. So a crop-by-context or metal-by-context split is now possible over the resolved verses and is NOT served by any endpoint yet: no product surface reads ritual_context. What has changed is that the axis exists to build it on, and that a split built on it would have to carry the typed absence rather than collapse it into a zero.
What this is not
The coverage figure is NOT a resolution figure, and neither is a finding that Vedic crops and metals are non-ritual. A verse whose context is NO_RITUAL_CITATION_FOUND is one no surviving ritual manual cites, which is a fact about the manuals and not about the verse.
Measurements behind this limit
mantras
20,210
Every verse in the corpus. The population a context assignment would have to cover for the split to mean anything.
mantras linked to a rite
513
Verses linked to a rite. Real, and it is a same-verse mention of a rite rather than an assignment of the verse to a ritual setting.
mantras with a context assignment
20,210
Verses carrying a ritual-context assignment. Every verse carries one, which is a statement about coverage of the axis and not about how many of them resolve to a rite.
mantras placed in a rite
1,903
Verses the assignment actually places in a rite, certainly or probably. This is the figure a context split can be run on.
mantras whose context is a typed absence
18,307
Verses whose assignment is a typed non-answer -- no citation found, or an opening shared with rival verses that the citation cannot choose between. Counted here so the coverage figure above cannot be read as resolution.
Ask this insteadThe material-culture and metals views, which report occurrence by corpus with the lexical-minimum ceiling stated, and do not split by context. There is no context-split endpoint to point at: the axis is on the verses and nothing reads it yet.
Cannot be answered
Which weapons occur in which narratives?
Not yet modelled
Why
The graph holds 5 weapons and no narrative layer: 0 node labels name a narrative, myth or episode, and the 8 theme edges that exist attach a subject label to a verse rather than grouping verses into an episode. There is no way to say 'these verses tell the Vrtra-slaying', so a weapon cannot be placed in a narrative and no role in one can be recorded.
What this is not
This is NOT a finding that the Vedas contain no narratives. The corpus is full of them; this graph has no construct that represents one.
Measurements behind this limit
weapons in the registry
5
Weapons the graph knows. The half of the question that exists.
passage theme edges
8
Verses carrying a theme. The nearest thing to a narrative, and a theme is a subject label rather than an episode with participants.
narrative labels
0
Node labels naming a narrative, myth or episode. This zero is the layer the question is built on, searched for by name rather than assumed absent.
Ask this insteadWeapon occurrence by corpus from the material-culture view, with its lexical minimum stated.
Cannot be answered
Which communities emerge when the graph is restricted to high-confidence edges?
Not yet modelled
Why
Half of the benchmark's reason no longer holds and the other half is decisive. The high-confidence tier now carries 8,129 verse-to-entity assertions, so it is not purely bibliographic as the benchmark recorded -- although 0 of them reach a concept, so restricting to high confidence still deletes the concept layer outright. The second clause is what closes it: 0 subjects carry a community membership at any tier, so there is no partition to restrict.
What this is not
This is NOT a finding that high-confidence Vedic evidence yields no structure. No partition exists to be restricted, and the restriction itself would silently drop the concept layer, which a returned answer would not have disclosed.
Measurements behind this limit
tier a passage to entity edges
8,129
High-confidence edges from a verse to an entity. Non-zero, which reverses the benchmark's first clause: the high-confidence tier is no longer purely bibliographic.
tier a passage to concept edges
0
Of those, the ones reaching a concept. Still zero: the concept layer sits entirely at TIER_B, so a high-confidence restriction removes it completely.
subjects assigned to a community
0
Subjects carrying a community membership. The second clause, and it fails absolutely whatever tier is chosen.
Ask this insteadDeity co-occurrence with the tier of each contributing edge stated, so a reader can see what a confidence filter would remove.
Cannot be answered
Which graph paths are meaningful versus accidental -- can a path be scored for interpretability?
Not yet modelled
Why
A path can be traversed and each of its 276,660 candidate edges carries a grade, but there is no scoring function with a stated basis, 0 stored path metrics, and 0 human-reviewed edges to calibrate one against. Composing per-edge grades into a per-path number would produce a sortable ranking that nothing validates, which is worse than refusing: the ranking would look like a finding.
What this is not
This is NOT a finding that graph paths here are meaningless. It is the absence of any instrument that could tell a meaningful one from an accidental one.
Measurements behind this limit
stored path metrics
0
Stored metrics of a path kind, under any name. Searched by name rather than assumed absent.
human reviewed edges
0
Edges a human has reviewed. This zero bounds every calibration question in this product: with no judged sample, no score can be calibrated against anything.
graded edges
276,660
Edges carrying a quality grade. Large, and a grade per edge is not a score per path: multiplying grades along a path would invent a number.
Ask this insteadThe path endpoint, which returns the path with every edge's own grade and evidence and leaves the judgement to the reader.
Technical detail
The translation typology, the notation figures and the Ask benchmark come from the completeness record certified against release 50a40429103f, read from the knowledge service on this request. The per-edition rows, their exclusions, the corpus and translation totals and every recitation figure are read from the knowledge service when this page is built, and the question-level probes above run on the request.
Every limit here is a fact about this graph, never about the Vedas. NOT_ANSWERABLE means the dimension the question asks about was not built; it does not mean the corpus is silent, and an answer of 'we cannot establish this' is the honest outcome rather than a shortfall.
measured
This catalogue covers every one of the 21 questions the 100-question benchmark graded NOT_ANSWERABLE, each with a probe that runs on this request. It is still not a complete account of what this product cannot do: a question the benchmark never asked is not thereby answerable, and the benchmark is 100 questions rather than every question.
measured
Each card carries the frozen benchmark's grade beside the live one. Where they differ, the graph has moved since the benchmark was frozen and the live verdict is the one to read: a limitation copied forward from a stale grade would tell you this product cannot do something it can.