H2/H3R — sense survival & drift on inheritance edges

A 28-noun panel across three measured inheritance edges (WIL→SHS, WIL→YAT, AP90→AP). H2 tests whether cited ancestor senses survive more often than uncited ones. H3R measures sense-unit drift (copy, condense, revise) along each edge using mean senses per lemma and gloss overlap. Generated from data/lexico/r2_h2h3.json; see docs/R2_FINDINGS.md.

Trust Block. Generated from data/lexico/r2_h2h3.json (H2: cited 0.762 vs uncited 0.705; archived 0.7/0.54). Limitations: panel reconstructed from nouns in all 5 dicts (documented drift in R2_REBUILD_CONTRACT.md); SHS/YAT senses split by inline `N.` markers. Validation: npm test; all unit tests pass. Owner repo: csl-atlas.

H2 — Citation-survival (Supported)

H3R — Sense-drift per edge

Data

Edge	Senses anc→des	Drift	Gloss overlap	Pattern	Archived
Wilson 1832 → Śabda-Sāgara 1900	9→9	0	0.91	near-verbatim copy	7.9→8.5 (overlap 0.82)
Wilson 1832 → Yates 1846	9→5.679	-3.321	0.26	condensation (semicolon-aware count, #126 promotion: YAT abridges ~9→5.7, not the artifactual 9→1)	7.9→1.1 (overlap 0.15)
Apte 1890 → Apte 1957	10.821→7.75	-3.071	0.56	revision, no expansion	15.5→11 (overlap 0.61)

Summary

H2 supported: Cited senses survive at 76% (n=84) vs uncited at 71% (n=723); gap = 0.06. Well-sourced senses are stickier.

H3R not supported (no net-addition): Derivatives copy or condense. Wilson 1832 → Śabda-Sāgara 1900: near-verbatim copy (overlap 0.91). Wilson 1832 → Yates 1846: condensation (semicolon-aware count, #126 promotion: YAT abridges ~9→5.7, not the artifactual 9→1) (overlap 0.26). Apte 1890 → Apte 1957: revision, no expansion (overlap 0.56).

Robustness — the citation effect is edge-concentrated

The H2 chart above is the unadjusted gap, and it is fragile. Cited ancestor senses are not spread across the panel: Wilson 1832 carries almost no <ls> citations, so 82 of 84 cited senses sit on a single edge — Apte 1890 → 1957 (the other two edges have one cited sense each). That concentration makes the pooled controlled regression unreliable: refitting survived ~ cited + centrality controls + edge fixed effects (lemma-cluster-robust) gives a citation OR of 3.0 (p = 0.011), but the estimate moves with unrelated parsing choices — the YAT semicolon promotion (#126) alone shifted it from ≈1.75 to ≈3.0 without touching a single cited sense. The pooled OR is therefore not a trustworthy citation effect.

The clean test is within the one citation-bearing edge, where parsing of the others is irrelevant. On Apte 1890 → 1957: cited senses survive at 0.768 (n = 82), uncited at 0.661 (n = 221) — a two-proportion z = 1.80, p = 0.072, not significant. So citation co-varies with survival but is not established as an independent predictor; a larger, multi-edge panel with citations on more than one edge is needed. (h2Controlled.withinPrimaryEdge + citedByEdge in data/lexico/r2_h2h3.json.)

Threshold sensitivity (referee item M4). The pooled OR is significant at every Jaccard cutoff from 0.10 to 0.25 — but that stability just inherits the edge-concentration fragility above; it is not evidence for H2. (Note the naive gap itself collapsed to ~0.03–0.06 once YAT is counted properly, from ~0.17 before.)

Jaccard cutoff	cited rate	uncited rate	naive gap	pooled OR	pooled p
0.10	0.798	0.740	0.058	2.32	0.038
0.125	0.786	0.730	0.055	2.64	0.028
0.15 (ref)	0.762	0.705	0.057	3.04	0.011
0.175	0.738	0.696	0.042	2.58	0.025
0.20	0.738	0.694	0.044	2.84	0.017
0.25	0.714	0.686	0.028	2.70	0.024

Verdict. Treat the pooled OR as unreliable (edge-concentrated); the honest H2 signal is the within-Apte-edge test — suggestive, not significant (z = 1.80, p = 0.072). Generators: h2Controlled / h2ThresholdSensitivity in data/lexico/r2_h2h3.json. (The generated "H2 supported" summary above reflects the unadjusted pooled rates only.)