Jlink

Usually it wins, and the baselines are set up to give string methods their best shot: exact matching after normalization scores 0.26, 0.41, 0.00, 0.00 and 0.22 across five datasets, while jlink reaches 0.73 F1 on firms and 0.996 on DBLP-to-ACM. Each baseline got the single threshold maximizing its F1. But on synthetic typo data it loses: 0.96 against TF-IDF cosine's 0.998 on FEBRL4.

Mostly the model, but recall is capped by candidate search. On the firm run the forward top-10 n-gram pass proposed 67% of known links, and jlink found 93% of the true links it was given. Since 0.4.0 it also searches in reverse from each right record to its nearest left records, and the report counts records no pass paired with anything.

As a strong ranking, with the middle band audited. On the firm benchmark pairs above 0.8 were true matches 95 to 98% of the time and pairs below 0.2 were true 0.1% of the time, but the 0.5 to 0.8 band was overconfident: mean stated 0.64, actually true 39% of the time. On FEBRL4 it went the other way and was underconfident, so calibrate per dataset.

Not unless that agreement allows it. jlink sends the fields you list in on, for each candidate pair, to TypeSafe or OpenRouter; blocking runs locally and only those listed fields leave your machine. The README says not to use it on confidential or restricted-use data such as Census RDC files, identified administrative records, or anything under a DUA. Experimental local server modes keep everything on your machine.

With uv: uv add jlink in a project to import it, or uv tool install jlink for the command line that Stata and R also call; add pyarrow for .parquet, and it works on .csv, .dta and .parquet.

Write it in plain English as the definition argument, for example that a parent and subsidiary are different firms or that a legal-form change is the same firm; the sentence is part of the method and belongs in your data appendix.

jlink draws a stratified sample for hand-labeling and turns your labels into precision, recall and a calibration table with confidence intervals; every probability is stored so reruns cost nothing and return the same links.