H1 — does sense granularity inflate over time?
Each point is one dictionary: x = publication year, y = mean sense-units per entry
over its whole corpus; colour = lexicographic family. Generated from data/lexico/r2_h1.json
by npm run build-r2-pages.
Fixed 30-lemma panel (nouns present in all 11 dicts, per-lemma aggregation): removes MW headword-splitting confound. MW 1899 rises from 1.15/entry (full corpus) to 14.9/lemma (panel). Pearson r = 0.093 — year-trend stays flat even after deconfounding; H1 NOT supported.
Within-family controls — the sharper test
Year is a dict-level covariate with only 11 values across ~7 families, so the pooled Pearson r conflates time with family composition. The honest test holds family fixed: the four families covered by ≥2 dicts each give a within-family trend (panel units, which also remove the MW splitting confound).
| Family | Span | earlier → later (panel units/lemma) | Δ total | units / yr | verdict |
|---|---|---|---|---|---|
| indigenous | 1822–1873 | skd 1.767 → vcp 1.033 | −0.734 | −0.014 | decrease |
| Petersburg | 1855–1928 | pwg 5.167 → sch 1.567 | −3.600 | −0.049 | decrease |
| Monier-Williams | 1872–1899 | mw72 24.333 → mw 14.900 | −9.433 | −0.349 | decrease |
| Apte | 1890–1957 | ap90 10.667 → ap 8.867 | −1.800 | −0.027 | decrease |
Every family declines (mean −0.11 units/lemma·yr) — the opposite of inflation. A family-controlled OLS (panelUnits ~ year + family, n=11) gives year's partial slope −0.051/yr, 95% CI [−0.186, +0.084] — negative, but the CI spans zero because 11 dicts over 7 families leave only 3 residual df (near-saturated: the slope is bounded, not measured). Within the single western-parsed regime (n=9) the bivariate r is −0.11, so the family-granularity gap is lexicographic, not a parser artifact. New block: h1Controlled in data/lexico/r2_h1_panel.json; per-(lemma,dict) rows in data/lexico/r2_h1_panel_lemma.json (330 obs).
Trust Block. Evidence: data/lexico/r2_h1.json (full-corpus per-dict means, 11 dicts; Pearson r = 0.036 vs archived 0.06) and data/lexico/r2_h1_panel.json (fixed 30-lemma panel; pooled Pearson r = 0.093 vs archived 0.01, plus the h1Controlled within-family block). The pooled r is underidentified, not a measured null; held family-fixed, granularity is flat-to-decreasing over time, so H1 (sense granularity inflates over time) is contradicted in direction, not merely unsupported. Limitations: each within-family slope rests on as few as 2 dicts; the OLS is near-saturated (CI bounds the year effect, does not estimate it); panel biased toward medium-to-high sense-count nouns; MW 1872 outlier (24.3) reflects sparse nominal coverage. Validation: npm run build-r2-h1-panel + npm run build-r2-pages; 187/188 tests pass (the one failure is the unrelated sanskrit-util sibling drift-guard). Owner repo: csl-atlas.
Finding. Granularity clusters by FAMILY, not year: enumerating traditions (Apte, Benfey) pack many sense-units/entry; lumping traditions (Monier-Williams, Petersburg, indigenous) ~1. The pooled year-trend is near-zero (Pearson r = 0.036 full-corpus, 0.093 deconfounded panel), but that r is underidentified — year is pinned to 11 dict-points spread over ~7 families. Held family-fixed, the trend is not just absent but negative: all four multi-dict families decline over time (mean −0.11 panel-units/lemma·yr), and the family-controlled year slope is −0.051/yr (95% CI [−0.186, +0.084], near-saturated). So H1 (pure temporal inflation) is contradicted in direction; sense granularity is a tradition / marking-style trait, which Paper L must CONTROL FOR — and any future claim of inflation is falsifiable as a single family showing a clear per-year increase. The variance is captured by family (means: Benfey 2.53, Apte 2.35, Monier-Williams 2.06, Wilson 1.8, Cappeller 1.35, Petersburg 1.15, indigenous 1.09).