A directed graph of 77 public-domain authors (Homer to the 1920s), candidate edges permitted only chronologically forward in time. Every edge carries two independent similarity scores, never merged: stylistic (word choice, syntax — TF-IDF) and conceptual (ideas, themes — embeddings). Held-out against real, independently-documented influence claims, conceptual similarity is significant and replicates across two separate validation sources; stylistic similarity does not hold up under a proper sweep (see caveats below). Click any author to trace their edges. Full method: design doc.
What this is: chronologically-valid candidate edges with a measured similarity score, not proof of influence — the accent-colored edges are the subset independently documented (LLM-enumerated critical consensus, or Wikidata's structured "influenced by" property; never used to build the graph, only to check it after the fact). What's solid: conceptual similarity between documented pairs is significantly higher than a shuffled-timeline null, twice over, on two independent sources (z=9.47 on 130 held-out pairs, z=7.16 replicated on 102 independent Wikidata pairs) — and it survives a density-confound check on the best-represented authors (z=6.25). What's most likely null: stylistic similarity is not significant on the full held-out sample (z=0.91), and the two narrower checks that are significant (Wikidata z=2.45, well-represented subset z=2.97) turn out not to sit on a trend. Raising the minimum books per author and re-testing at every level — each against a null drawn from that same subset — gives 0.9, then −0.7, then 0.3, then a single excursion at ≥4 books that decays again. Conceptual, run identically as a control, is significant at every level and declines smoothly as pairs are lost. The published trio were three non-nested samples, not a series. n=77 authors is real but modest scale.