Lineage Sankey

The central visual evidence for the kosha-collapse finding: six PWG <ls> labels (left) flow through the actual kosha works they name (middle), then converge into MW's single <ls>L.</ls> hedge (right).

Trust Block

• Evidence: hard-coded PWG and MW citation-flow counts from the documented kosha-collapse analysis.
• Limitations: this is a focused explanatory visual, not a complete source graph or a generated review queue.
• Validation: checked by npm run build; counts should be regenerated before paper-final use.
• Owner repo: csl-atlas.
• Next use: treat the chart as structural evidence, then check companion docs before making a lineage claim.

What the diagram shows

Stage 1 (left, six green nodes): PWG's six most-cited named kosha labels with their actual citation counts.

Stage 2 (middle, six pink nodes): the kosha works behind those labels. Hemacandra has two — Abhidhānacintāmaṇi (the main lexicon) and the Anekārthasaṃgraha (his polysemy supplement).

Stage 3 (right, one blue node): MW's single <ls>L.</ls> hedge — collapses all six PWG flows into one. 40,212 citations, 13.95% of all MW entries.

The collapse is the story. PWG distinguished six named koshas. MW (and PWK before it) dropped the named-source apparatus and replaced it with L. — gaining typographic compactness, losing bibliographic precision.

Shared-corrected-error overlay (APPARATUS-NOT-ERRORS)

The kosha-collapse Sankey above is about named sources. The overlay below is about mechanical errors: headwords that were corrected on both a Petersburg dictionary (PW / PWG) and MW. Edge width is the forensic shared-correction count (shared_corrections.csv, n=). The copy-detection verdict is not the edge weight — it is the Ahlborn direct test and co-correction null model already computed as F4b.

Trust Block (overlay)

• Evidence: data/forensic/shared_corrections.csv, data/forensic/f4b_report.json, packet src/data/corrections/shared_error_overlay.json (npm run build-correction-lane-overlays).
• Limitations: sample of co-corrected headwords, not the full C3 form-keyed loci join; edge weight is raw count, not residual above a confusion-aware null; in_pwgissues multi-dict fix bundles inflate co-correction by design.
• Validation: F4b numbers reproduced in docs/A10_REPRODUCIBILITY_AUDIT.md; packet + page via npm test / npm run build.
• Owner repo: csl-atlas (APPARATUS-NOT-ERRORS).
• Next use: treat co-correction lift as convergence + editorial coupling, not inheritance of mistakes; pair with Correction loci and Correction front.

Verdict (F4b): MW shares the PWG headword error in only / (%) cases. Co-corrected headwords are elevated (lift ), but that is the same hard words receiving different fixes — not copied mistakes.

Source: CDSL pwg.txt + mw.txt 2026-05-23. Static SVG: sankey-en.svg. Shared-error overlay: forensic F4/F4b (H1579). CC-BY-SA-4.0.