Descent axes — is the citation canon a fourth axis?

The atlas measures dictionary inheritance along three committed axes — content (headword-stock containment), convention (house-style fingerprint similarity), and microstructure (register/layer similarity) — and THREE-AXIS-INDEP showed they are statistically separate on the 13 documented L0 edges. This page tests PH2 CITE-4AXIS: is the citation profile — cosine similarity of two dictionaries' normalised citation vectors, "do they quote the same books?" — a fourth separable axis, or does it just travel with one of the three?

The citation vectors come from the canonical-siglum source matrix behind the Citation apparatus page; the dict×text graph behind the Citation canon explorer is reported as a sensitivity column (its MW coverage is degenerate — see the Trust Block).

n = testable edges of the documented L0 edges — only 7 of the 14 edge dictionaries have a validated <ls> citation adapter, so edges are excluded (listed below), not imputed. At this n only near-perfect collinearity would be detectable (critical |r| = at p=0.05); read every number as descriptive.

Verdict —

No citation~axis correlation reaches the n= p=0.05 threshold, and no exact label-permutation p (all permutations enumerated) falls below 0.05:

The one estimate worth watching is citation~microstructure (Pearson , permutation p ) — the strongest cross-axis coupling measured on any axis pair so far. It is not detectable at n=, but it is the first thing to re-test when the documented edge set grows.

Axis-pair scatters — one dot per documented edge

Each panel plots the testable edges with the citation cosine on the vertical axis against one of the three committed axes (Pearson r per panel: content , convention , microstructure ).

The 4×4 correlation matrix (same reduced edge set)

Pearson correlations of all four axes on the same n= edges — the three-axis cells here differ from THREE-AXIS-INDEP's n=13 values because the edge set is reduced; the committed n=13 stats remain the reference for the original three axes.

The testable edges

The sensitivity column is the same cosine computed on the data/citations/ dict×text graph. It rank-agrees with the apparatus cosine at Spearman , but its MW values are artifacts of resolver coverage (BEN→MW = 0.0 because the graph resolves MW to only a handful of texts) — which is why the apparatus matrix is the primary source.

Excluded edges (honest shrinkage)

Chart Trust Block

• Claim:
• Evidence label: ${data.evidenceLabel}; review status: ${data.reviewStatus}.
• n = testable edges of documented (honest shrinkage: edges excluded for missing <ls> adapters — fewer than the agenda's "9 of 13" estimate; nothing imputed). Small-n caveat: critical |r| at n= is (p=0.05) / (p=0.01); all verdicts are descriptive, not confirmatory.
• Source files: three_axis_comparison.json, citation-apparatus.json, ls_citation_edges.tsv (sensitivity only).
• Generated by: npm run build-four-axis-independence → src/data/lexico/four_axis_citation_independence.json; exact label-permutation p over all permutations per citation~axis pair.
• Validation: npm run validate-four-axis-independence (recomputation match, edge accounting, matrix symmetry, cosine bounds); npm test; npm run build.
• Known limitation: the citation vectors live in the apparatus top-source space (head of the citation-frequency distribution); MW is under-represented in the sensitivity ls-graph column, not in the primary apparatus column.
• Owner repo: csl-atlas.
• Next action: re-test citation~microstructure when the documented edge set grows (csl-atlas#89/#92); read the fourth-axis verdict into A03 §4 (paper_three_axes_descent.md).
• External dependencies: none at runtime; inputs are committed atlas artifacts.
• Boundary note: dictionary citation evidence only — no corpus passage search, no standards export.

Related: Citation canon explorer (the dict×text matrix this axis is derived from) · Citation apparatus (apparatus style) · Dictionary genealogy (the descent evidence the axes decompose).