Citation truncation & hapax overlap

Headword containment says the Petersburg dictionaries and Monier-Williams share an enormous amount of stock, but containment is symmetric — it cannot say who copied whom. This page carries the asymmetric test named in docs/LEXICOGRAPHY_ROADMAP.md §"Citation truncation patterns" (hypothesis PET-MW-CITE): a dictionary can shorten an ancestor's reference — PWG Rv. 1.22.16 → MW RV. — but it cannot invent locator precision it never had. On the sources both sides cite, the more specific side is the plausible ancestor.

Beside it sits the content-side half: hapax overlap. A shared common lemma is evidence of nothing — every Sanskrit dictionary has agni. A lemma recorded by exactly two of these dictionaries and by no other is the rare-stock analogue of a shared error.

Apparatus style, before any pair is formed

Locator depth is the number of numeric or i/v/x roman components a citation carries after its siglum: RV. = 0, RV. i, 1 = 2, Rv. 1.22.16 = 3. The two lanes are separated by this alone.

MW leaves % of its citations at a bare siglum, against % for PWG. Truncation is not a rare event in MW — it is MW's default.

The per-dictionary census

Truncation asymmetry — the directional test

Each row is one dictionary pair. evidence(A→B) is the count of A's citations, on the sources both cite at least times each, whose locator depth exceeds B's mean depth for that same source. The within-lane rows are controls — descent is not in question there, so they show what the measure reads on ordinary house-style difference.

Where the truncation happens — per shared source

Pick a pair to see the sources behind its number: how often each side cites the source and how deep its locators run. A positive delta means the first dictionary is the more specific one on that source.

Points below the diagonal are sources the first dictionary cites more precisely. The dot area is the pair's combined citation volume for that source, so a big dot far below the line is where the truncation claim carries the most weight.

Hapax overlap — the rare-stock half

shared hapax(A,B) counts lemmas recorded by exactly two of the in-scope dictionaries, and those two are A and B. rare Jaccard restricts both sets to lemmas in at most two dictionaries; rare/all is its ratio to the all-headword Jaccard, which is structurally ≤ 1 — read it as how much of the pair's agreement survives when common stock is removed, not as a lift.

Chart Trust Block

Generated by npm run build-citation-truncation. See docs/LEXICOGRAPHY_ROADMAP.md §"Citation truncation patterns" and docs/HYPOTHESIS_INDEX.md (PET-MW-CITE). CC-BY-SA-4.0.