Ghost stock — is dictionary-unique vocabulary real?

42% of the cross-dictionary union's lemmas appear in exactly one dictionary. Are they recorded language, or lexicographic sediment — inherited citation-forms, artefacts, ghost words? This page joins the union headword backbone against two independent modern witnesses: the Digital Corpus of Sanskrit attestation flags (via the frozen VisualDCS summary) and, for MW, the Sanskrit Heritage lexicon crosswalk (agenda hypotheses PH4 GHOST-STOCK + PH6 HERITAGE-WIT).

Of union lemmas, (%) are DCS-attested — but the rate climbs monotonically from % at n_dicts = 1 to 100% for lemmas shared by 13+ dictionaries. Dictionary-unique vocabulary is overwhelmingly corpus-invisible.

Attestation by multiplicity (PH4)

Share of lemmas DCS attests, per number of dictionaries listing the lemma. Whiskers are Wilson 95% score intervals (mostly narrower than the dots).

Which dictionaries' unique stock is real? (per-dict)

Among each dictionary's unique lemmas (n_dicts = 1), the share DCS attests. Note the honest surprise: the highest unique-and-attested shares belong to MW and MD (general dictionaries), not to the specialised lexica the agenda predicted — though DCS under-samples Buddhist and technical literature, which caps what BHS-style uniqueness can show (see Trust Block limitations).

Logistic model

Descriptive logistic regression of attested on n_dicts plus family-presence indicators (deterministic IRLS; McFadden pseudo-R² , n = ). Each additional witnessing dictionary multiplies the odds of corpus attestation by ≈ .

The Heritage triangulation (PH6) — MW lemmas only

For the union lemmas MW lists, a 2×2×2 cube: MW-unique × Heritage-covered × DCS-attested. Both witnesses point the same way — Heritage-uncovered lemmas are far likelier to be DCS-unattested.

Ranked ghost-candidate queue — inferred, not asserted

The triple filter MW-unique ∧ Heritage-uncovered ∧ DCS-unattested isolates candidates in two evidence tiers: explicit ( — Heritage's own MW↔DICO alignment saw the key and did not cover it) ranks above crosswalk-missing ( — key absent from the crosswalk, possibly a key artefact). These are review candidates for the H5-style QA queue: absence from two modern witnesses is evidence of rarity, not proof a word never existed.

Chart Trust Block

• Claim: dictionary-unique headwords are disproportionately corpus-unattested (PH4), and Heritage non-coverage triangulates the same ghost stratum for MW, yielding a ranked candidate queue (PH6).
• Evidence label: derived for the attestation rates, strata, cube, and model (deterministic joins + arithmetic on committed inputs); inferred for the ghost-candidate queue — no row is asserted as a ghost word until a human source read (H5 discipline) decides it.
• Source files: union_headwords.tsv ( normalized lemmas; owned by SanskritLexicography), mw_heritage_crosswalk.tsv (same owner), data/dcs/dcs_lemma_summary.json ( lemmas, frozen VisualDCS export); committed packet src/data/ghost-stock/ghost_stock.json with provenance in its .source.json sidecar.
• Generated by: npm run build-ghost-stock
• Validation: npm run validate-ghost-stock — strata/cube/queue arithmetic, Wilson-CI bracketing, logistic convergence, plus a full rebuild cross-check against the sibling checkout when present; the logistic fit is independently reproduced (NumPy IRLS) to 4 decimals.
• Known false positives: none at the join level — attestation and coverage are read directly off the committed witness files, not inferred.
• Known false negatives: DCS samples the transmitted literature — "unattested" means absent from that corpus release, not absent from Sanskrit; Buddhist/technical vocabulary is under-sampled, which is exactly why BHS's unique stock scores low here. Homonyms collapse onto one normalized key (attested if any homonym is). Heritage coverage is MW-keyed and mirror-derived, not a live INRIA query.
• Review status: machine-reviewed (deterministic validator); the ghost-candidate queue awaits human spot-check — route via the H5 anomaly review discipline before any correction or paper claim.
• Owner repo: csl-atlas (join + rendering); the union backbone and crosswalk are owned by SanskritLexicography; corpus attestation belongs to VisualDCS per the consumption contract.
• Next action: /spot-check-sample a stratified sample of the explicit tier against MW source lines; then fold verdicts into the H5 queue.
• External dependencies: SanskritLexicography (union + crosswalk), VisualDCS summary (in-repo frozen copy).
• Boundary note: no DCS ingestion here (VisualDCS contract); Heritage is cited and joined, never cloned (LGPLLR).