Working note · companion to the two proposals
The two proposals ask for scholarly reading time. This note describes the instrument that reading time is spent inside — who writes to it, what it refuses to do, and why it is not a wiki.
It is not itself a proposal. The instrument has to exist whether or not either proposal is funded, and most of it already does.
The argument in five lines
Both halves of the work already exist in public and have done for years. Neither half, alone, produces an index — and the reason is visible in two pages anyone can open.
Hindupedia, Dharmawiki and the Sanskrit-topic pages of the general encyclopaedias already carry a great deal of Paurāṇic material, much of it written with real care by serious scholars. The difference proposed here is not diligence. It is checkability, and it is worth demonstrating on a real page rather than asserting.
Hindupedia's article on Satyabhāmā, by Swami Harshananda, tells the story of her previous birth — that she was Guṇavatī, daughter of the brāhmaṇa Devaśarman, and was reborn as Kṛṣṇa's queen through her devotion. The article is accurate. Its References section, in full, reads:
1. Pārijāta means night jasmine.
The only source given for the article is a secondary work — the author's own Concise Encyclopaedia of Hinduism.
That source is followable: its publisher hosts the book online. Follow it, and the entry is four paragraphs of the same prose. No Purāṇa is named. No chapter. No verse. The chain runs from a wiki to a reference work and stops there, one step short of the text it is describing.
Two details of that copy are worth a moment, because they are the kind of thing this project exists to make visible. The printed entry reads “the pārijāta (night jasmine) tree” — so the wiki's entire scholarly apparatus turns out to be a parenthesis from its source, promoted to a numbered footnote. And the transcription is verbatim down to a misprint the two share, Kṛṣna for Kṛṣṇa. A common error is exactly how a copy is detected — it is how the descent of manuscripts is traced — and here it is doing that work between a wiki and a book, in public, with nobody watching.
None of which is a failure by Swami Harshananda. A concise encyclopaedia is not a critical apparatus and was never offered as one; omitting citations is the genre working as intended. It becomes a problem only when the genre is all there is — when the compression is the last surviving step, and the address it dropped was never recorded anywhere a reader can reach.
Meanwhile the passage itself is not lost at all. It stands at Kārtika Māsa Māhātmya 13, in Sanskrit, verse by verse, and in this corpus it has a measured onset in a recording.
Set the two side by side and something worse than a missing citation appears. The encyclopaedia gives the cause of her rebirth as “her purity of life, austerities and devotion to Viṣṇu.” The chapter's own Sanskrit heading gives it as कार्तिकव्रतपुण्येन — by the merit of the Kārtika observance — and the verse itself says she performed that vrata, पुण्यव्रतं कृतवती. In a work whose entire purpose is to glorify one observance, the retelling has dissolved that observance into general piety. The operative fact is the casualty.
And without the address, nobody can see that this happened. That is the case for citing to the verse, made by a single example rather than by argument. The story survived the retelling; its address did not; and because the address did not, neither did the claim.
And the sources that do give an address disagree: chapter-by-chapter translations place the episode at Skanda Purāṇa, Kārtika Māhātmya 13, other retellings say 21, and some attribute it to the Padma Purāṇa's Uttara Khaṇḍa instead. That disagreement cannot be settled without a verse address, and no amount of prose supplies one.
This is the structural difference, not a failure of care. On a wiki a page is a document and a citation is a string typed into it by hand. Skanda Purāṇa, Kārtika Māhātmya 13.4 looks like an address, but nothing dereferences it, nothing tests it, and nothing fails when it is wrong. There is no build that can go red.
This is not a hypothetical distinction, and it can be checked in about ten seconds. Wisdom Library's page for Viṣṇu assembles some fifty-nine definitions drawn from many texts and dictionaries across eight traditions — Vaiṣṇava, Purāṇic, Śaiva, Jain, Buddhist, and six language lexicons — each snippet carefully attributed to where it came from. It is substantial, genuinely useful work over a corpus far larger than this one, and it is free.
And it lists the god, the author of a dharmaśāstra, and fire in one undivided run, because separating them is not a thing aggregation can do. That is a concordance, and an excellent one. The step it does not take is the step this proposal is entirely about.
Read the right way round, that is the strongest available argument for funding the reading. The automated half of this work has already been demonstrated at scale, publicly and for free, by someone else. It did not produce an index — because sense-division is a judgement, and judgements are made by people who have read the passage. Sorensen made those judgements for one text, a hundred years ago, and no one has made them since for anything larger.
Put the two together and the shape of the problem is exact. A reference work carries the judgement and drops the address. An aggregation carries the address and declines the judgement. Neither is an entry, and each is missing precisely what the other has.
An index entry is the two joined: a claim with an address attached. Sorensen settled that form in 1904. Nobody has redone it since — for any corpus larger than the one he read.
An entry that carries both — the judgement and the address — and, where a recording exists, the second at which the verse is chanted.
The store is an append-only log over the text. It holds two kinds of row, and keeping them apart is the whole design.
Testimony is never adjudicated. An entry records that Sorensen states X is the son of Y at III, 65 — never that X is the son of Y. Judging whether the authority is right is a later, human, separate act.
Rulings are adjudicated, and they supersede rather than delete. A junior's reading survives a senior's correction, because the disagreement between them is a measurement the project needs, not a conflict to be tidied away.
Parikṣit shows why both are needed, because he presents one of each.
One person, two treatments. The Mahābhārata's Parikṣit — cursed by Śṛṅgi, bitten by Takṣaka — and the Bhāgavata's Parikṣit, the listener of the seven-day discourse, are the same man differently expounded. Both accounts stand. Nobody rules, and no entity is split: a name is never divided by which book it stands in.
Four people, one spelling. Sorensen's index separates four distinct Parikṣits — a son of Abhimanyu, a son of Avikṣit, a king of Ayodhyā, a son of Anaśvan — and the works spell them alike. Here a machine has no business deciding, and a scholar must.
A search engine cannot tell these two situations apart. It is the entire reason the reading is worth funding.
An index that never decides is a concordance. Sorensen's worth is that he did decide — four Parikṣits, seven Ugrasenas, divided and glossed — and that a century later you can see what he decided and on what evidence.
Take two of the best-known figures in the tradition, and two passages about them that are told again and again without anyone saying where they are told.
She asks Kṛṣṇa what she can possibly have done to deserve him — दानं व्रतं तपो वापि किं नु पूर्वं कृतं मया — and he answers by telling her of her former birth. The chapter's own Sanskrit heading says so outright: that she obtained the standing of Kṛṣṇa's wife through the merit of the Kārtika observance in a previous life.
The story itself is not obscure; it is retold in reference works and across the web. What is almost never given is where it stands. Here it is a verse, in Sanskrit, with a number — and, in the atlas, a second in a recording. That is the whole of what this project adds to a story the tradition already knew.
In the recording, the verse where he begins — शृणुष्वैकमनाः कान्ते, listen with a single mind, beloved — is seated in the waveform to within a second. Open it in the atlas and hear it →
A brahmin named Dharmadatta is granted a boon by Viṣṇu's messenger: that in another birth he will be a king renowned in the Solar line by the name of Daśaratha — and that Viṣṇu himself will be born as his son. The Rāmāyaṇa's premise, given here as the fruit of a month's observance, in a Purāṇa that is not the Rāmāyaṇa.
This one is the honest case. The verse naming him — नाम्ना दशरथस्तत्र, Kārtika Māsa Māhātmya 25.24 — is exact and numbered, and it is not in either recording. Musiri expounds this chapter as far as 25.21 and goes on into the next without reading the boon itself. So there is no second to offer, and the atlas offers none: a citation that resolves in the text and nowhere in the audio is a different thing from one that resolves in both, and the point of the atlas is not to blur them.
Why these two carry the argument. Both are figures a reader already knows, which is precisely what makes the passages hard to find. Search either name and the famous material buries everything else — Satyabhāmā across the epic and the Kṛṣṇa cycle, Daśaratha across the whole Rāma tradition. These chapters are a needle in that. Yet neither required any inference to locate: the corpus states both stories in its own chapter headings, in Sanskrit, and has done all along. Nobody had read those headings as an index.
So what a scholar should be handed is not a corpus and not a search box. Handing someone eighteen Mahāpurāṇas and a search field is handing them the problem back. It is a dossier for each headword — Satyabhāmā's, as it stands today, before any new work:
| what is known | |
|---|---|
| In the Mahābhārata, under her own index entry | 62 attestations, 41 of them certified by a philologist |
| Across twelve Purāṇas — Skanda, Padma, Viṣṇu, Brahma, Brahmavaivarta, Bhāgavata and others | 155 candidates, none yet verified |
| Chapters whose own heading names her | including the one above, unreachable by name-search |
| Recordings in which she is heard | two reciters, onsets measured in the waveform |
That is roughly two hundred and twenty verses ranked by evidence, in place of three hundred and fifty thousand. A scholar writing on Satyabhāmā would not have thought to open the Brahma Purāṇa.
Four things go into such a dossier, and the order matters:
An upanyāsakar expounding a text does something no index does. He glosses each verse in the vernacular, names figures the verse leaves unnamed, and reaches across the corpus from memory — as the Rāmāyaṇa tells it, as the Bhāgavata has it — because that is what the tradition of exposition is.
The transcripts already catch this. Across the recordings processed so far the material is structured as Sanskrit verse followed by its vernacular exposition, alternating, so each gloss is bound to the verse it glosses:
| in the transcripts so far | |
|---|---|
| Episodes transcribed | 129 |
| Units of vernacular exposition | 13,375 |
| Episodes invoking the Gītā by name | 44 |
| …the Upaniṣads | 33 |
| …the Bhāgavata | 29 |
| …the Rāmāyaṇa | 28 |
At a hundred and twenty-nine episodes this is suggestive. At the few hundred the audio proposal would commission it becomes a corpus of pointers: the places a scholar did not know to look, named by someone who did.
It must be graded honestly. A pointer is a lead, not a citation. The expounder rarely gives chapter and verse; the transcript is machine output with machine errors; and the exposition inherits its place in the recording from the verse beside it rather than from its own measurement. So it ranks below an indexed citation — but it differs from a machine's guess in the way that matters: a named scholar said it, on a date, at a timestamp, in a recording anyone can play.
This is where the two proposals actually meet. Each produces something the other needs. The Nāma-Kośa supplies the entity layer the Audio Project lists among its own limits; the Audio Project supplies the growing body of expert cross-references that turns a search box into pointers.
The reading is the expensive part and cannot be automated. It can, however, be the same act as learning the text.
The wiki feeling belongs to the people making the index. Among themselves they want exactly what a wiki gives: a page for each name, links between them, somewhere to argue, and the sense of a thing taking shape.
Readers of the finished work get none of that model. They get an index — typed headwords, sense divisions, glosses, verse-addressed citations, and audio where it exists. Nobody outside sees a work queue or an undecided span, except where the index deliberately publishes its own frontier.
Three roles, and they are not three job descriptions so much as three different substances being contributed.
| role | contributes | working conditions |
|---|---|---|
| First-pass reader | Judgements — is this name standing in this verse? | Phone, twenty minutes between classes, unreliable data. One decision at a time. |
| Technical intern | Apparatus — extractors, guards, the generated pages | Laptop, version control, review. |
| Senior scholar | Rulings and the written entry | Sees only what the system has already marked hard. |
The Śāradā Pīṭham runs educational institutions at Kalady, New Delhi and Bangalore alongside its Veda Pāṭhaśālas. A very small number of interested students from each — not an assigned batch — is the right shape to begin with, and it settles three things at once.
It also makes one commitment concrete rather than aspirational: the interface has to work in Sanskrit and in the reader's own bhāṣā, which for this cohort means Malayalam, Hindi and Kannada. That is a list the recruitment gives you, not a guess.
The proposals ask for reading time, and reading time is expensive because it is normally bought. There is a second way to get it, and it is better than buying: set it as coursework. A student who reads the Virāṭa Parva closely enough to write about Uttarā has, in the same act, read it closely enough to say whether उत्तरा in a given verse is her. Those are not two tasks that happen to align. They are one task, described twice.
The assignment. Choose a figure nobody has written up yet, and write a note on them in which every claim is cited to a verse.
Why the constraint is the point. Left to themselves, students choose Arjuna, Bhīma, Draupadī — and those are precisely the entries an index already handles well. The value is in the tail. So the pool is not a list the teacher keeps; it is a state of the data: a headword attested in the text with no written entry behind it, and no ruling yet on who stands where. What is taken shows as taken, and the same query that assigns the coursework measures the coverage.
How large the pool is. Counting only persons cited between five and a hundred times in the epic — small enough to finish in a term, large enough to have a story — and only those appearing in three or more parvans, so the note is a thread rather than a mention: 1,256 figures. It is not a shortage that has to be rationed. It is a queue that coursework is capable of draining.
What a good pick looks like. Not Arjuna, but Arjuna and Ulūpī. The Nāga king Kauravya's daughter takes him to her father's palace during his year of exile in the first book; twelve books later she incites his own son Babhruvāhana to fight him, and when Arjuna falls she revives him with the saṃjīvana gem and explains the curse that had to be worked out that way; and she is still present at the very end, in the book of the great departure. Twenty-eight verses, twenty-three of them already cited by a philologist, spread across four parvans out of eighteen. The references exist. The synthesis does not. That is a term's work with a real result at the end of it.
A different lesson, from a different pick. A student who takes Uttarā — Virāṭa's daughter — finds instead that seven separate entries claim the name she goes by. One is her brother. One is an epithet of Viṣṇu. One is the direction north. Before writing a sentence they have learnt what no lecture conveys as well: a name in Sanskrit is a set, not a label, and the tool refuses to choose for them. Following her own entry — wife of Abhimanyu, mother of Parikṣit — carries them from a dancing lesson in a disguised court to the man the whole epic is being recited to.
What the corpus gets. A headword that had no entry now has a drafted one, with every claim addressed to a verse; a set of bookmarked passages with a reading attached to each; and, where the student was sure, a proposed identification. A senior scholar is then editing a draft rather than starting from nothing — which is the whole difference between an index that gets built and one that stays proposed.
The assignment. Here are ninety-seven verses in which four different Uttaras are each claimed. Rule on them.
This is not the essay task in another form. It requires construing the verse — the case ending, the governing verb, who is being addressed — and it is exactly what a pāṭhaśāla trains. The student may also mark cannot determine, which is a real answer and is credited as one.
What the corpus gets. Signed adjudications: the substance of the index, produced by the readers best qualified to produce it, as part of their own study rather than instead of it.
The assignment. Parikṣit is treated one way in the Mahābhārata and another in the Bhāgavata. Document the difference without deciding which is right.
The seminar produces something the machinery specifically needs and cannot generate: a ruling that two accounts both stand, attributed to their sources and reconciled by neither. It is also where a student first meets the difference between a text disagreeing with another text and a machine having made a mistake.
The assignment. Build the view that answers: which other names stand in the verses where this one stands?
A self-contained piece of work on a real corpus, delivered to real users, which becomes part of what every reader above uses the following term. The portfolio piece is genuine because the artefact is.
A submission is read twice, to two different standards, and the distinction is not a formality.
So the classroom lane is quarantined by default. What can flow in immediately is the mechanical residue — which verses were looked at, which identifications were proposed — because each of those is a claim with an address that anybody can re-check. Prose and sense-divisions wait for a signature.
One rule follows from this and it is not negotiable: never assess by volume. If a mark depends on producing thirty citations, thirty citations will be produced. Assessment rests on the senior's sample, and I could not determine which Uttara this is earns full credit when the verse genuinely does not say.
The arrangement needs no defending on novel grounds, because it is not novel. In this tradition, teaching and transmission were never separate activities. A student who chanted the text was the mechanism by which it survived; a student who copied a manuscript learned it by copying, and the copy outlived them both. Nobody described that as extracting labour from students, because the thing the student gained and the thing the tradition gained were produced by one act. What is proposed here is the same arrangement with the addresses made checkable.
Its nearest modern ancestor is the reference work this one is modelled on. The Oxford English Dictionary was built by an appeal to ordinary readers to send in quotation slips, and its most prolific readers were named in its prefaces. Sorensen's index belongs to the same century and the same method. This is not a technology company's model imported into Sanskrit scholarship; it is how the great reference works of Sanskrit scholarship's own period were actually made.
Its nearest modern institution is the teaching hospital, and that is the comparison to hold the design to. Residents learn by treating patients and patients are treated; the institution exists because the two coincide. It also carries its safeguards with it — supervision, sign-off, and nobody practising unsupervised — which are exactly the rules set out above. The comparison you choose is the standard you are held to, and this one demands the quarantine and the signature rather than excusing them.
So: the student gets a corpus that answers back, a supervisor's correction on real textual judgement, and a named credit in a published scholarly apparatus — a route from study into research authorship that presently barely exists. The project gets reading time it could not otherwise afford, from readers who were going to do the reading anyway. Where a contributor is a minor, the credit and the consent are a guardian's to give, and the senior scholar's name stands on anything published.
Neither side is doing the other a favour. That is the only arrangement of this kind that lasts more than one term.
Beyond the index itself: what having the addresses lets this project give back, and how anyone can watch it being built.
A corpus that is addressed at the verse can settle other people's citations, and that is worth more to the field than another set of pages. The disagreement above took one lookup to resolve: a chapter-by-chapter translation's verses 1, 2–3 and 4–6 correspond exactly to this corpus's 13.1, 13.2 and 13.4, so chapter 13 is right and the rival attribution is not.
Note the shape of that correspondence, because it is the whole method. The chapter numbers agree but the segmentation does not — a translation prints as one unit what the text divides into three. Joining verse n to verse n therefore drifts silently at every merge, which is a defect this project has already shipped once, with a different translation, and had to fix by matching on the text instead. So the join is content-addressed: two passages are the same passage when the Sanskrit says so, never when the numbering does.
What that yields is a resolver — give it a loose citation or a quoted half-verse, and it returns the address, the Sanskrit standing there, and, where a recording exists, the second at which it is chanted. The same machinery makes the expounders' cross-references in VII followable, since those are loose citations too. It is a by-product of having built the index, not a second project.
Three limits keep it honest. A resolver links to another site's text pages, where a chapter and a translator are named; it must never take that site's aggregated name-pages as an index, for the reason set out in IV. Translations in copyright cannot be absorbed, only linked — which is the correct relationship anyway, and it means a Purāṇa this site carries no translation for can still hand a reader one, verse by verse, without holding a word of it. And a mapping to somebody else's page numbers has to be stored with the text that justified it, so that when their site is reorganised the link can be rebuilt rather than merely lost.
A citation is text in a paragraph, verified by whichever reader happens to check. A page is authored and edited; provenance means knowing who last changed the wording. Where two traditions disagree, the last editor decides.
A citation is an address into a machine-readable edition, and the build refuses to publish one that does not resolve. A page is derived — rebuilt nightly from evidence. Provenance means knowing which decisions produced a claim, by whom, on what grounds. Disagreement is recorded with both readings standing.
And one thing no wiki can offer at all: where a recording exists, the citation plays. Kārtika 13.2 is not only a verse whose Sanskrit can be printed. It is a named reciter, at a measured second, in a recording anyone can open — which is what the atlases on this site already do.
This is not a book sent to the press, with a copy shared when it arrives. The output is digital and close to real time, so what a supporter — or a contributing institution, or a reader — should get is a nightly rebuild with a report of what changed: decisions signed and by whom, headwords closed, escalations opened, agreement between readers, and how far the frontier has moved.
The unit that closes should be small: a headword, or a single adhyāya — never the Harivaṃśa. Begin where an existing index already supplies the answers, so that units start closing in the first fortnight. If the first closed unit is eight months away, nobody watches.
It matters more for the readers than for anyone funding it. A student who can see their name against four hundred signed decisions, and watch a chapter go from grey to complete, comes back next week.
Exists. The corpus, addressed at the verse across the epic and all eighteen Mahāpurāṇas. The name layer over it, with graded evidence and rival claimants kept as sets rather than guesses. The transcripts described in VI. The two atlases on this site, and the publish-and-verify step that refuses to ship a page whose data is broken. The dossier in V is a query away.
Does not exist. The reading client itself, and the queue behind it. The discussion threads. The nightly report. The index of chapter headings. Nothing here is large, and none of it is the hard part.
Neither software nor solvable by it. Who assigns the work; what happens when a reader stalls; who chases an unanswered escalation; how a reader is calibrated before going live and checked again afterwards. That is the part which decides whether any of this works, and no amount of design settles it in advance.