Candidate recall limits the best possible result
The README explains that the firm example was limited by candidate search. A correct judge cannot recover a true pair that blocking never supplied.
Inspect the supporting source ↗REPOSITORY ANALYSIS
jlink
A record-linkage tool that asks Jev whether candidate records refer to the same entity under an explicit user-defined matching rule, then resolves links in code.
E2 · source reviewedPublic material or source evidence was reviewed. A fixed-version source review requires checking a pinned revision; this is not a reproduction, performance endorsement, or safety guarantee.
- Stars
- 4
- Forks
- 0
- License
- MIT
- Last commit
- 2026-09-23
What Jev does here
Jev judges whether two candidate records refer to the same entity under a user-supplied match definition; local code generates candidates and resolves accepted matches.
Separates blocking, pair judgments, thresholded resolution and manual audit samples so missed candidates and wrong judgments can be investigated separately.
Fixed-version documentation and implementation material were inspected on 2026-10-24. No record linkage or benchmark was rerun. Records are sent for remote scoring; anonymization and permitted data use remain application responsibilities.
What task does this project fit?
Link records under an explicit definition of what counts as the same entity.
Material analysed：
From input to action
- Blocking creates a tractable set of candidate record pairs.
- One Noul per pair asks the user’s matching rule against the two records.
- Local resolution applies thresholds, optional margins and relationship constraints; audit samples support manual labeling.
Implementation details that matter
A threshold is tied to the data and match definition
The judge includes the supplied entity definition. Subsidiary-as-parent and typo-only matching are different tasks; one published operating point should not be copied without labels.
Inspect the supporting source ↗When to choose it
Useful when semantic identity matters more than string resemblance. For typo-only data, compare a simple normalized or string-similarity baseline before paying for pairwise judgments.
How to inspect it yourself
Read question in judge.py alongside README’s block→judge→resolve flow. Inspect an example definition and the candidate table; count how true pairs could be absent before interpreting judge accuracy.
Limits of this review
- No record linkage or benchmark was rerun.
- Records are sent for remote scoring; anonymization and permitted data use remain application responsibilities.