VisualDCS

Created: 20-04-2026 · Last updated: 24-09-2026

Interactive frequency dashboards for the Digital Corpus of Sanskrit (DCS), built from corpus frequency data and rendered as standalone HTML files — no build step, no server, open directly in a browser.

---

What is this?

The DCS is the largest annotated corpus of Sanskrit texts, containing hundreds of thousands of morphologically tagged verb and nominal forms. VisualDCS turns that raw frequency data into visual, interactive tools for learners and researchers who want to understand what Sanskrit actually looks like in practice — which forms dominate, which are rare, and how coverage accumulates.

---

CSL Atlas DCS Handoff

VisualDCS is the home for DCS/corpus material moved out of csl-atlas. The initial atlas handoff landed on 2026-06-04 in VisualDCS PR #4, under docs/csl-atlas-migration/.

Those files are migration material only; they are not yet integrated into the runtime dashboards.

---

Research Archive (derived-data/, non-derived/) — tracked, not part of the dashboards

Two large folders sit alongside this repo's code: derived-data/ and non-derived/, together ~7.2GB / ~2,080 files. They're a personal Sanskrit-linguistics research archive (much of it V.V. Leonchenko's "Цифровой корпус санскрита" corpus-statistics work and related material) that predates and is broader than this repo's shipped dashboards. Both are git-tracked and pushed since 02-07-2026 (files >95MB as 7-Zip split volumes, see RESTORE_SPLIT_FILES.md; only the nested Zalizniak/GH/ git repos are excluded), but neither feeds the dashboard pipeline, which runs solely on src/DCS-data-2021/ and src/DCS-data-2026/ below. Treat them as a reference archive to mine for ideas or figures, not as live dashboard input.

• derived-data/ — the DCS-corpus half: datasets computed by statistically analyzing corpus text (frequency counts, distributions, collocations, verb/nominal-form analysis, compounds, synonyms, and the large Paralleli-v-tekstah-korpusa-SRC/ intra-corpus parallel-passage search). The most-frequent-Sanskrit-word / core-vocabulary lists specifically live in derived-data/Lexical-Cores/.
• non-derived/ — the non-DCS half: dictionaries (Kochergina, PWG/MW comparisons, the Saudamani electronic dictionary), the NCC manuscript catalog, lecture/conference material, translations, and cloned external tool repos under Zalizniak/GH/.

Both were reorganized 02-07-2026 — folder names transliterated Cyrillic→Latin, a legacy Works/Share/ export tree merged and deduplicated, and topics split DCS-vs-non-DCS. Each folder has its own README.md (context/how-to-work-with-it) and INDEX.md (per-folder size/file-count table + full change history) — start there rather than exploring blind.

---

Source Data

The repository draws on two upstream sources:

1. src/Распределение времен и наклонений.xlsx — an Excel file of raw frequency counts of Sanskrit verb forms across the DCS corpus. It powers the verb-form frequency dashboard:

Metric	Value
Total examples	781,616
Unique lemmas	55,032
Tense/mood categories	38

The two figures above are the Excel headline — what the dashboard HTML renders. The tracked CSV twin src/DCS-data-2021/timws.csv is the machine-readable source and yields 42 category codes summing 781,618 examples: the H1486/H2294 Aorist · Periphrastic Perfect · Perfect re-split turned pre-split buckets into separate codes, and the 2-unit gap to the Excel headline is the documented ±10 reconciliation (see CHANGELOG and .ai_state.md). Both numbers are correct for their own source; neither is hand-typed any more — scripts/verb_total_check.py re-derives the CSV pair from bytes on disk and fails if this README drifts from it.

2. src/DCS-data-2021/ — a raw dump of the DCS corpus (CSV/txt: 10.csv ≈ 4.57M annotated tokens, 7.txt, _8.csv, cs.csv, …). It is the source for the paradigm browser and the derived concordance/JSON assets. Because some files exceed GitHub's 100 MB limit, the dump is split into line-boundary parts (rebuild with src/DCS-data-2021/rejoin.bat) and stored as plain git blobs (converted out of Git LFS to avoid the storage quota); see DCS-data-CLEANUP.md for the full inventory and rationale.

3. src/DCS-data-2026/ — the current DCS distribution (CoNLL-U / Universal Dependencies, CC BY 4.0), imported into a queryable SQLite master by a documented, validated pipeline. The full corpus — 270 texts · 5,688,416 tokens · 754,726 sentences · 74 treebank texts — is pinned to gasyoun/dcs-conllu @ 04e0778 (a submodule at src/DCS-data-2026/conllu) and published as a SQLite GitHub Release (dcs-full-2026-03-05, 287 MB gz). Pipeline: parse_conllu → import_dcs_conllu → coverage_diff → export_master → validate → regen_widgets, validated end-to-end (cross-walk 0 mismatches; CI re-runs the suite on push). See DCS_CONLLU_IMPORT_PLAN.md and src/DCS-data-2026/reports/ for the pipeline, the verb tense/mood code map, and the 2021→2026 deltas.

Cross-repo consumer asset — dcs_lemma_summary.json

dcs_lemma_summary.json (83,239 lemmas) is the canonical DCS lemma-frequency summary that other repos in the org are meant to consume directly instead of re-parsing the CoNLL-U corpus — VisualDCS is registered as the org's DCS corpus/morphology ingest owner (family 8 in SHARED_CODE.md). It is produced by gen_dcs_lemma_summary.py from the DCS-2026 master and checked by validate_dcs_lemma_summary.py (CI-gated via validate-lemma-summary.yml). It is the feed behind the csl-atlas DCS adapter.

Aorist ≠ Perfect — the Tense=Past re-split (H1486)

UD's Tense inventory has no Aorist or Perfect value, so Sanskrit's two past tenses both surface as Tense=Past (111,167 tokens). DCS's own feat_formation carries the past-stem formation, and regen_widgets.py now uses it to re-separate them inside the finite past indicative (93,329 tokens — the only sub-bucket where the feature is populated):

class	rule	tokens	% of bucket
Aorist	feat_formation ∈ root·them·red·s·is·sis·sa (Whitney's seven aorist types)	12,054	12.92%
Periphrastic Perfect	feat_formation = peri	4,046	4.34%
Perfect	feat_formation IS NULL — DCS leaves the simple perfect unmarked	77,229	82.75%

Only the last row is a default rather than an observation, so it is measured, not assumed: validate_past_tense_resplit.py runs four independent checks and writes reports/past_tense_resplit_validation.md.

Read this before quoting either number. Both classes are bounds, not exact counts: Aorist is a lower bound (≥1.13% more aorists sit untagged inside Perfect, provably — the same surface forms appear tagged elsewhere in the same bucket), and Perfect is an upper bound additionally carrying ≥3.54% forms attested elsewhere as Tense=Impf, an upstream tense-tagging inconsistency the re-split cannot repair.

This supersedes the repo's former claim that feat_formation is "present on <2% of verbs, too sparse to re-split them". That divided 16,100 tags by all ~1.01M verb tokens; against the finite past indicative — the only bucket the feature applies to — coverage is 17.25%. The split is applied in verb_forms_ud.json / verb_forms_38cat.json; it is deliberately not propagated into visual/paradigm_attested.json (see the paradigm-trainer section).

The other past moods cannot be split at all (H3878). The re-split stops at Mood=Ind because that is where the data stops: across the whole pinned distribution Formation occurs 17,440 times and never once outside Mood=Ind — 16,100 on Tense=Past and 1,340 on Tense=Fut (the periphrastic future, a different feature that must never be summed with the periphrastic perfect). The 8,726 non-indicative past tokens (Jus 4,067 · Imp 1,700 · Sub 1,317 · Opt 1,065 · Prec 577) and the 9,112 past participles are unformationed upstream, so no query change here can recover a tag. Propagating from identical surface forms attested elsewhere would reach 296 of them (3.39%), which is why that inference layer was measured and rejected rather than built. Quote these moods by mood only — Mood=Jus on Tense=Past is the augmentless injunctive, and whether it rests on an aorist or an imperfect stem is exactly what DCS declines to say. Evidence and method: reports/past_nonindicative_formation_audit.md, regenerable by audit_past_nonindicative_formation.py and guarded by test_past_nonindicative_formation.py.

---

Dashboards

Four standalone HTML tools, best entered via the landing page.

sanskrit_index.html — landing page / tool map

The recommended starting point. A single page that lays out a Stage 1 → 4 learning path (which roots and forms to study first for the fastest corpus coverage) and a filterable grid of tool cards. It advertises 14 tools; 13 of the cards open real widgets — 12 standalone files in this repo plus the live WhitneyRoots app — and one card (TANS: 193 текста) remains a planned widget: no genre/date dataset for those texts exists in the repo yet.

sanskrit_verb_form_dashboard.html — verb-form frequency

Distribution of Tenses, Moods, and Participles. An interactive single-page dashboard with three charts, built from the Excel source (38 categories, 781,616 examples):

Chart	What it shows
Bar + Pareto curve	Frequency of each verb form category with cumulative % overlay
Pareto detail line	Cumulative coverage by form rank (top 5 → 77.6%, top 11 → 94.9%)
Lemma density bars	Unique lemmas per category (breadth of vocabulary per form)

Key findings:

• Past Passive Participle leads with 233,079 examples (29.8%)
• Present Indicative follows with 157,003 (20.1%)
• Just 5 forms cover 77.6% of the entire corpus
• Just 11 forms cover 94.9% of the entire corpus

sanskrit_pxn_v4.html — paradigm browser

An interactive paradigm browser for 6 roots (√kṛ, √bhū, √as, √gam, √vac, √dā) across 9 tenses × 9 person/number cells (87 cells), built from the raw DCS corpus (745,394 verbal uses). Each cell is colour-coded by corpus frequency and clickable for real corpus examples. Features: stem+ending colour split, root comparison, verb-class labels, a "what to study next" coverage route, flashcard mode, a zero-cell filter, and CSV/Markdown export. Full feature documentation is in sanskrit_pxn_v4_docs.md (Russian).

sanskrit_paradigm_trainer.html — paradigm trainer, attested verb space (H1299)

Scales the paradigm browser from 6 hand-picked roots to 7,689 corpus-attested roots (frequency floor >=2 total VERB tokens; top 100 = "full" tier, the rest = "attested-only" long tail, cards with no hand-curated notes). Root picker with search + frequency rank, attested-only cell rendering by construction (every cell shown carries at least one real corpus form + count), and a frequency-weighted flashcard trainer mode with a JSON deck export for downstream SRS. Built by src/DCS-data-2026/gen_paradigm_attested.py from the full 2026 DCS master (dcs_full.sqlite, 270 texts, 1,007,361 VERB tokens); regression-pinned by src/DCS-data-2026/test_paradigm_attested.py. Loads visual/paradigm_attested_data.js via a plain <script src> (no fetch(), no server) -- keep both files together when copying. The 6-root deep view (sanskrit_pxn_v4.html) is untouched and still the tool for those 6 roots' full textbook paradigm with P./A. columns.

Data ceiling -- read before trusting a cell as "resolved": DCS tags unaccented text, so it cannot distinguish verb class I vs VI, or class IV vs the passive formation, at the root-class level -- no Panini class number is shown anywhere in this dataset. Tense=Past conflates aorist and perfect in UD, but DCS's own feat_formation re-splits the finite past indicative, and since H2294 this dataset carries that split too — Aorist / Periphrastic Perfect / Perfect are separate per-root cell categories, and the E46 reconciliation was re-run unchanged (6,454 matching roots, 0 disagreements). The split is bounded and one-sided, and the dataset says which way: Aorist and Periphrastic Perfect are read off feat_formation, while Perfect is the unmarked default — DCS does not tag the simple perfect, so every Perfect cell is 100% inferred rather than observed. The dataset therefore ships a cellEvidence map (formation-attested vs defaulted) and the trainer badges it on every cell, flashcard and exported deck card: the defaulted share of a cell is exactly 0% or 100% — never in between, asserted at build time — which is what makes a per-category marker exact rather than an approximation. Aorist is a lower bound, Perfect an upper bound (>=1.13% aorist leakage, >=3.54% imperfect contamination); prose quoting either as an exact corpus count is a defect. feat_voice only tags Passive vs non-passive -- parasmaipada vs atmanepada is not separately tagged by DCS, so non-passive finite forms for a cell are pooled together (unlike the 6-root RD grid's P./A. columns, which come from external grammatical knowledge, not this corpus tag set). Full report: reports/paradigm_attested_build.md; cross-check against csl-observatory's E46 paradigm-cell-coverage census: reports/e46_reconciliation.md.

sanskrit_nominal_dashboard.html — nominal paradigm, case × number (H1472)

The nominal twin of the verb paradigm browser, and the tool roadmap.md had been calling "the most obvious next step" since 2026-06: the 8 case × 3 number grid per declension class, over 2,263,192 cased NOUN + ADJ tokens (against 781k verbal ones). Pick a stem class (-a, -ā, -i, -ī, -u, -ū, -ṛ, -an, -in, -ant, -as, -is, -us, plus an explicit other consonant residue) and a token gender; each cell is colour-coded by its share within that gender and opens the attested surface endings, the top attested forms, and real corpus sentences with text + reference. Built by src/DCS-data-2026/gen_paradigm_nominal.py from the pinned 2026 DCS master; render-tested headlessly by tests/test_nominal_dashboard.js (node tests/test_nominal_dashboard.js). Loads visual/paradigm_nominal_data.js via a plain <script src> (no fetch(), no server) — keep both files together when copying.

Data ceiling — read before trusting a cell:

• Declension class is not a corpus tag. DCS tags case, number and gender, never a declension class. Every class label here is an orthographic heuristic over the lemma's citation form, reusing the tag list of SanskritGrammar's Sangram G2 asset (H1048) and its "citation form ≠ stem" caveat, extended with the -ant class G2 leaves in its residue.
• Compound members are not a case. 724,676 NOUN/ADJ tokens carry feat_case='Cpd' — no case at all. They are excluded from the grid and counted separately per class, so the dashboard can never pass the grid off as the whole nominal picture. Grid + Cpd + case-untagged = 2,996,410 = the entire NOUN/ADJ universe, asserted at build time.
• A gender column is not a gender paradigm. The axis is the token's tagged gender: an adjective cited in its masculine -a form contributes its feminine tokens (paramayā) to the -a class's Fem column, where they inflect as ā-/ī-stems.
• Endings are surface residues, not morphemes — what is left after stripping the class-marked stem; where the form does not start with that stem (strong/weak alternation, rājan → rājñā) the token is counted as not segmentable and left out of the endings view rather than force-split. The segmentable share is shown per class and per cell.
• The form shown is DCS's unsandhied analysis, not manuscript surface — 68.0% of grid tokens are flagged m_unsandhiedreconstructed.
• Known unsplit conflations, labelled per class: -ī pools the devī/nadī type with the monosyllabic śrī/strī type; -an pools -an/-man/-van; -ant pools -ant/-vant/-mant with the master's own -at/-vat/-mat citations of the same stems; other consonant is a residue bucket, not a class.
• Two of those conflations are now split (H3984) on an external lexical signal, not a character rule: -ī by IAST vowel-nucleus count (17 lemma_ids / 6,373 tokens monosyllabic vs 4,477 / 58,295 polysyllabic), -ant by MW+PWG csl-json headword entry-id sets (191 / 8,851 one-lexeme-two-spellings, 0 two-headwords, 36 / 70 left pooled with no dictionary witness). NOUN universe only — the ADJ mass is declared unreached, not estimated. Builder src/DCS-data-2026/split_pooled_nominal_classes.py, reports reports/nominal_pooled_class_split.md
  • reports/nominal_g2_reconciliation_split.md (reconciles: true, no G2 total moves), payload visual/paradigm_nominal_class_split.json. The paradigm_nominal.json buckets themselves are unchanged — the split is an additive side-car. See GAPS §14 and FINDINGS §630.
• The -ant pool's ADJ half is now measured too (H4011) — the surplus H3984 declared unreached. Sourced per-lemma from the DCS master (read-only) instead of the NOUN-only G2 CSV: 36,768 ADJ tokens over 1,880 lemma_ids, adjudicated by the same MW+PWG entry-id signal into 34,809 / 1,015 one-lexeme-two-spellings, 0 two-headwords, 661 / 193 at_only, 95 / 13 ant_only, 1,203 / 659 left pooled. The pinned bhagavat (lemma_id 48482, ADJ, 3,238 tokens) is finally classified: one lexeme, two spellings. The -ī class's 664 ADJ tokens / 54 lemma_ids are split by the same syllable count (31 / 3 monosyllabic vs 633 / 51 polysyllabic). With both halves in, the -ant pool is 100 % inventoried and 97.4 % of its token mass carries a verdict; every NOUN and Sangram G2 total is re-derived and unmoved. Builder src/DCS-data-2026/split_pooled_nominal_classes_adj.py, report reports/nominal_pooled_class_split_adj.md, payload visual/paradigm_nominal_class_split_adj.json — a sibling, not an overwrite.

Full report: reports/paradigm_nominal_build.md; cross-check against Sangram G2's declension-cell coverage (57,144 lemma_ids, exact agreement on tokens and attested cells): reports/nominal_g2_reconciliation.md.

sanskrit_nominal_trainer.html — per-lemma nominal trainer (H2321)

The nominal twin of sanskrit_paradigm_trainer.html: 31,753 NOUN lemmas (frequency floor ≥ 2 tokens; top-100 = full tier) with a search list, an 8 case × 3 number grid of attested cells only, frequency-weighted flashcards, and JSON deck export. Built by src/DCS-data-2026/gen_paradigm_nominal_lemmas.py; render-tested by tests/test_nominal_trainer.js (node tests/test_nominal_trainer.js). Loads visual/paradigm_nominal_lemmas_data.js via <script src> — keep both files together.

Load-bearing constraint — coverage is not re-derived. The 24-cell bitstring for every lemma is taken from SanskritGrammar Sangram G2's lemma_cell_coverage.csv (H1048; 57,144 lemmas, already reconciled exactly against this repo's class dashboard). Forms and per-cell counts come from the pinned master, and only for cells G2 already marks attested. The generator refuses on any token-count or cell-set drift vs G2 (measured: 0 / 0).

Data ceiling: same as G2 + H1472 — median cells/lemma is 1; a full 24-cell paradigm is almost never attested; stem_final is a citation-form tag; forms are DCS unsandhied analyses (often reconstructed). Hapax lemmas below the floor stay in G2 but are out of this trainer asset.

Build report: reports/paradigm_nominal_lemmas_build.md.

sanskrit_anki_decks.html — Anki-deck flashcard widget (H1504)

A standalone flashcard viewer over the already-computed visual/anki_compact.json — 200 cards across 4 learning stages (Stage 1: √kṛ, √vac, √bhū, √as; Stage 2: √gam, √dṛś, √dā, √śru, √sthā, √brū, √ah; Stage 3 and 4 widen to the long tail). Data is embedded directly in the page (no fetch(), no companion file — the whole deck is ~16 KB of JSON), so the file is self-contained and opens directly from disk. Stage filter, click-to-flip cards, a CSV export in Anki's plain-text import format (#separator:Comma / #html:true / #columns:Front,Back,Tags header, HTML-formatted backs, per-card tags), and a Markdown export grouped by stage for Obsidian. The landing page's Anki-деки (D2) card links to it directly.

sanskrit_concordance.html — concordance search (H1505)

A standalone search widget over the 6,423-form concordance that had, until now, only been consumed inline by other dashboards' example panels. Type a root or an inflected form (diacritic-insensitive, so ASCII kuryat matches kuryāt) and get every matching form with its total corpus occurrence count and up to 5 real example strophes, each with a source citation and the query highlighted in place. Built by gen_concordance_data.py, which packs the existing visual/conc_part1/2/3.json into a single visual/conc_data.js window.* script twin (≈5.3 MB, no fetch(), no server — works from a double-clicked file:// page); render-tested headlessly by tests/test_concordance.js (node tests/test_concordance.js). The landing page's "Конкорданс" card now opens it directly.

sanskrit_paradigm_builder.html — paradigm-ending heat map (A27)

The landing page's Парадигм-строитель card, previously a described widget only: a heat map of the 25 tense/mood categories by corpus share (99.1% of the 781k verbal examples together) and, per category, the person × number grid (sg/du/pl × 1/2/3) of the top-5 attested endings with occurrence counts, plus a side-by-side A-vs-B comparison of any two categories. Data is a verbatim embedded copy of the published visual/paradigm_endings.json (2021 DCS dump instrument; where segmentation is impossible the surface form itself is published as the "ending") — the embedded copy is checked against the JSON twin by tests/test_paradigm_builder.js (node tests/test_paradigm_builder.js). Self-contained; no companion file needed.

sanskrit_collocations.html — collocation navigator (A27)

The landing page's Коллокации card, previously a described widget only: a searchable, POS-filterable list of the 800 most frequent lemmas (n ≥ 317), each with its collocates (same-sentence lemma co-occurrence) grouped by part of speech — verbs, nouns, adjectives, other. Clicking a highlighted collocate navigates to its own card when that lemma is itself in the top-800 (with a back button); dimmed chips are outside the list. Data ships as visual/coll_data.js, a window.COLL_DATA script twin of the published visual/coll_compact.json packed by gen_collocations_data.py (no fetch(), no server — keep both files together); render-tested by tests/test_collocations.js (node tests/test_collocations.js).

Whitney Roots — external live app (gasyoun/WhitneyRoots)

The landing page's Whitney Roots card links out to the live app at gasyoun.github.io/WhitneyRoots: a crosswalk of Whitney's 930 roots (Whitney × Monier-Williams × Apte × DCS frequency/PPP) — 713 of them DCS-attested — built and maintained in its own repo; the top-6 roots already cover 25% of the attested root-token mass (19 → 50%, 68 → 75%). Not duplicated inside VisualDCS.

dcs_corpus_dashboard.html — corpus / genre statistics

DCS Corpus Statistics (Russian UI). A single page summarising the corpus by text and genre, consuming the tracked JSON assets visual/corpus_stats_widget.json, visual/dcs_genres.json (18 genre profiles), and visual/dcs_texts_clean.json (288 texts with tense profiles). The landing-page cards Корпусная статистика and DCS: 17 жанров open it.

The two verb dashboards report different headline totals (781,616 vs 745,394) because they use different aggregations of the corpus — the Excel's 38 tense/mood categories vs the browser's 87 person×number / non-finite cells. A third figure exists and is not a contradiction: the tracked CSV twin timws.csv carries 42 category codes summing 781,618 (post-re-split, ±2 of the Excel headline) — checked by scripts/verb_total_check.py.

---

Data Assets

The repository also tracks a set of derived JSON and reference files used to power future dashboards and widgets:

File	Contents
docs/csl-atlas-migration/	DCS/corpus handoff material migrated out of csl-atlas
sanskrit_verb_forms.md	Obsidian reference for the top 100 roots with paradigms
visual/dcs_texts_clean.json	288 texts with tense profiles
visual/dcs_genres.json	18 genre profiles — 17 named families + Other (weighted averages)
visual/dcs_scatter.json	170 data points for diachronic charts
visual/form_lookup.json	7,873 verb forms → root / tense / rank
visual/coll_compact.json / coll_data.js	800 lemmas × collocates by part of speech; coll_data.js is the packed window.COLL_DATA twin consumed by sanskrit_collocations.html (A27)
visual/paradigm_endings.json	25 tenses × attested endings from the corpus; embedded verbatim by sanskrit_paradigm_builder.html (A27)
visual/paradigm_attested.json / paradigm_attested_data.js	H1299: 7,689 roots × attested finite/non-finite cells (top-100 full tier + long tail), consumed by sanskrit_paradigm_trainer.html
visual/paradigm_nominal.json / paradigm_nominal_data.js	H1472: 14 declension classes × token gender × 24 case·number cells — counts, surface endings, top forms, corpus examples; consumed by sanskrit_nominal_dashboard.html
visual/paradigm_nominal_lemmas.json / paradigm_nominal_lemmas_data.js	H2321: 31,753 NOUN lemmas × G2-attested cells + top forms; coverage from G2, forms from pin; consumed by sanskrit_nominal_trainer.html
visual/paradigm_nominal_class_split.json	H3984: the -ī and -ant pools re-bucketed on an external lexical signal (syllable count; MW/PWG entry-id sets) with per-lemma verdicts and the G2 reconciliation totals; additive side-car to paradigm_nominal.json, no dashboard consumer yet
visual/paradigm_nominal_class_split_adj.json	H4011: the -ant pool's ADJ half (36,768 tokens / 1,880 lemma_ids) inventoried per-lemma from dcs_full.sqlite and adjudicated on the same MW/PWG entry-id signal, plus the -ī class's non-NOUN residue; sibling of the H3984 payload, which it does not modify
visual/corpus_stats_widget.json	Summary morpho-statistics for widgets
visual/anki_compact.json	200 Anki flashcards
visual/conc_totals.json	6,423 forms → total occurrences in corpus
visual/conc_part1/2/3.json / conc_data.js	Concordance: 6,423 forms × ≤5 examples (2,141 forms per part); conc_data.js is the packed window.* twin consumed by sanskrit_concordance.html (H1505)
verb_classes.json	13 verb classes with P/Ā distribution
tense_case_data.json	Form frequencies + case data (from cs.csv)
morph_pn.json	Person × number by tense (from 10.csv)
prefix_clean.json	Prefix productivity scores
passage_library.json	40 curated passages from the corpus

JSON files live in two places — most under visual/, but verb_classes, tense_case_data, morph_pn, prefix_clean, and passage_library sit at the repo root.

---

Other tracked material

Beyond the dashboards, source data, and research archive, the repo also holds:

• papers/ — the A38 data-descriptor paper for the DCS-2026 SQLite release (A38_dcs2026_release_paper.md) and its A38_release_checklist.md (Zenodo deposit + DOI-mint steps, still pending the upstream CC-BY sign-off).
• derived-parametric-core/ — a dictionary-side parametric core of Sanskrit extracted from PWG (Voronezh-school parametric-lexicology method of V. T. Titov / A. A. Kretov), cross-checked against V. V. Leonchenko's corpus core in derived-data/Lexical-Cores/. A separate, additive dataset — the Leonchenko sources in derived-data/ are untouched.
• mockups/ — non-destructive dark data-app restyles of the three shipped dashboards (CSS-only; the dashboard scripts are byte-identical). Design exploration, not the live pages.

---

Methodology

Dashboards use Pareto % (cumulative frequency %) to show how corpus coverage concentrates in a small number of high-frequency forms.

How Pareto % is calculated

Forms are sorted by descending frequency. Pareto %[N] is the share of the total covered by the top N forms:

Pareto %[N] = (Count₁ + Count₂ + … + CountN) / Total × 100

Example: (233,079 + 157,003) / 781,616 × 100 = 49.91%

Key thresholds in this corpus

Coverage	Forms needed
~50%	2 forms (PPP + Pres. Ind.)
~80%	5–6 forms
~95%	11 forms
100%	38 forms

The remaining 27 forms form a long tail — together they add only ~3.7% of coverage despite representing the majority of distinct categories.

For full methodology with term definitions, see pareto.md.

---

Roadmap

✅ Already shipped (in sanskrit_pxn_v4.html, plus the landing page):

• Landing page / tool map with a Stage 1 → 4 learning path — sanskrit_index.html

• Concordance integration — clicking a paradigm cell opens real corpus examples for that form

• Root comparison — a second root shown inline under each form

• Stem + ending colour split — invariant ending highlighted, stem greyed

• Flashcard mode — a cell as a question (root + person + tense → ?), answer on flip

• "What to study next" route — slider-driven, by corpus coverage gain

• Attested-only filter — hide paradigm cells with zero corpus examples

• CSV / Markdown export

• Per-root attestation counts, scaled to the whole attested verb space (H1299) — done via sanskrit_paradigm_trainer.html (7,689 roots, not just the 6 hand-picked ones; attested-only by construction, frequency-weighted trainer mode, JSON deck export). Supersedes the "Per-root attestation counts" item below for the general case; the 6-root deep view keeps its own richer P./Ā.-split cells for those 6 roots.

• Nominal paradigm dashboard, case × number per declension class (H1472) — done via sanskrit_nominal_dashboard.html (2,263,192 cased NOUN + ADJ tokens, 14 stem classes × token gender × 24 cells, surface endings + corpus examples per cell). This was the roadmap's "biggest unbuilt item".

• Anki-deck flashcard widget (H1504) — done via sanskrit_anki_decks.html (200 cards, 4 learning stages, stage filter, CSV export for Anki + Markdown export for Obsidian, data embedded so no companion file is needed). Landing page's D2 card flipped from type:'widget' to type:'file'.

• Concordance search widget (H1505) — done via sanskrit_concordance.html (6,423 forms, diacritic-insensitive search, total occurrence count + up to 5 example strophes with citation per form). The landing page's "Конкорданс" card flipped from type:'widget' to type:'file'.

• Print / PDF one-page export (H1536) — an on-screen "🖨 Печать / PDF" button (window.print()) plus an @media print stylesheet in sanskrit_pxn_v4.html and sanskrit_paradigm_trainer.html: Ctrl+P / clicking the button hides all nav/controls/side panels and prints only the current paradigm grid, one page.

• Per-lemma nominal drill-down trainer (H2321) — done via sanskrit_nominal_trainer.html (31,753 NOUN lemmas, G2 coverage oracle, frequency-weighted trainer). See the section above.

🔴 Still planned — high priority

• Split the -ī and -ant conflations — needs an external lexical signal (stem syllable count / a dictionary class tag); DCS alone cannot separate devī from śrī, and no guess is made in the current asset.

See roadmap.md for the original discussion.

---

Tech Stack

• Data: Microsoft Excel (.xlsx) for the frequency tables; raw DCS corpus CSV/txt under src/DCS-data-2021/ (plain git blobs, split into line-boundary parts — converted out of Git LFS) for the paradigm browser; derived JSON in visual/ and the repo root
• Dashboards: Vanilla HTML + Chart.js 4.4.1 — no build step, no dependencies, open directly in browser

---

License

• Code and dashboards (the HTML tools, Python pipeline, and derived JSON): Apache 2.0.
• The DCS-2026 SQLite master and its exports: CC BY 4.0, inherited from the upstream Digital Corpus of Sanskrit (Hellwig, CC BY 4.0). Cite it per CITATION.cff, and cite the underlying corpus annotation separately.

---

Dr. Mārcis Gasūns