First test: data and results (inconclusive)

The data and results of the first test, as published on the results page. Everything here is CC0, derived from OpenAlex.

Result: inconclusive. Of 28,114 pairs of topics from different domains that had barely cited each other up to 2017, 305 started citing each other by 2023. The citation network around each pair ranked them about 14 times better than chance. Adding what the papers say, measured by the words they share, did not measurably improve on it (difference in average precision -0.010, 95% interval -0.050 to +0.045). The plan, written before the analysis, is research/PREREGISTRATION.md.

Files

• pairs_2017.csv: the 28,114 evaluation pairs, one row each: who the two topics are, the links before and after 2017, the label, the stricter label, every feature the models use, and each model's score. The scores reproduce the published average precisions exactly.
• pairs_2011.csv: the 28,710 pairs the models were fitted on, the same columns without labels from the stricter check or scores.
• backtest_e1_k3_tfidf.json: the full report.
• robustness.json: the robustness checks listed in the plan.
• topic-records-seed2026.zip: every topic record the analysis used, as collected from OpenAlex in October 2026, so the numbers can be reproduced exactly even after OpenAlex changes.

Every column and field is explained in docs/DATA.md. The numbers were produced by commit 45c3048; the tables by python -m undiscovered_research.export at this release's commit.