jlink

Record linkage where you write the match rule in plain English.

import jlink

result = jlink.link(
    compustat, patents,
    entity="firm",
    on=[("conm", "assignee"), "state"],
    definition="A parent company and its subsidiary are different firms. "
               "A firm that changed its legal form (Inc. to LLC) is the same firm.",
    left_id="gvkey", right_id="assignee_id",
)

result.links             # gvkey, assignee_id, p (probability of a match), margin, ...
print(result.report())   # pairs compared, links made, what it cost
print(result.methods())  # a paragraph for your data appendix

String distance links "Acme Widgets Inc." to "ACME WIDGETS, INC". It does not link "IBM" to "International Business Machines", and it cannot know that you want a subsidiary kept apart from its parent. A research assistant can do both, slowly. jlink asks Jev, a decision model from TypeSafe, the question a research assistant would answer: given these two records and this rule, are they the same firm? Jev returns a probability in about a fifth of a second for about a thousandth of a cent. In the benchmarks below, 171,354 pairs cost $2.95 in total, at about 250 pairs a second. With no labels, jlink beats LinkTransformer's zero-shot models on eleven of twelve standard benchmarks and its fine-tuned ones on most of the product data (below).

jlink is built for the way applied economists link data:

• The rule is part of the method. You state what counts as a match, and that sentence goes in your appendix. Change the sentence and you change the linkage.
• Every pair gets a probability, so you can set a threshold, require a margin over the runner-up, or carry the uncertainty into estimation.
• You can check it. jlink draws a stratified sample for hand-labeling and turns your labels into precision, recall and a calibration table with confidence intervals.
• It reproduces. Every probability is stored. Rerunning costs nothing and returns the same links.
• It works from Python, the command line, Stata and R, on .csv, .dta and .parquet.

Before you start: where your data goes

jlink sends the fields you list in on, for each candidate pair, to an outside API (TypeSafe or OpenRouter). Do not use it on confidential or restricted-use data, such as Census RDC files, identified administrative records or anything under a data use agreement, unless that agreement allows it. Only the on fields leave your machine; blocking runs locally. With a local server (--api laya or --api diffusiongemma, below) nothing leaves the machine at all.

Install

uv add jlink            # in a project, for `import jlink`
uv tool install jlink   # the command line, which Stata and R also use

Both need uv. Add pyarrow if you read or write .parquet.

You need a key for one of two APIs. With keys for both, jlink uses TypeSafe's.

A key can also live in ~/.config/jev/typesafe.key or ~/.config/jev/openrouter.key.

Local servers (experimental)

--api diffusiongemma and --api laya, or JEV_API set to either name for the Python, Stata and R entry points, send the same pair questions to a System One server on your own machine, an OpenJev or laya-mlx process that you run separately. They are never chosen automatically, need no key, and count as $0 in the cost meter and a saved run's settings unless JEV_PRICE_PER_MTOK is set. The report still counts a local model's answers "by Jev" and scores.csv marks them source=jev; the report's Model: line, the model and provider columns of scores.csv and settings.json name the model that answered. The runtime's DiffusionGemma and Laya guides explain the setup; keep concurrency low while a local model warms up.

On a local server compares both with Jev on the five benchmarks, and the local-model results have the full record.

Try it

The repository ships a toy example: twelve Compustat-style firms and sixteen patent assignees.

jev-link link examples/compustat_sample.csv examples/patent_assignees.csv \
    --on conm=assignee --entity firm --left-id gvkey --right-id assignee_id --how many-to-many \
    --define "A subsidiary or division counts as its parent company. Different companies that merely share a word are not a match." \
    --block ngrams:conm=assignee:5 --block initials:conm=assignee --block exact:state -o links.csv

0.97  INTL BUSINESS MACHINES CORP   <- IBM                                      [initials]
0.98  INTL BUSINESS MACHINES CORP   <- International Business Machines Corporation
0.97  MINNESOTA MINING & MFG CO     <- 3M Company                               [exact:state]
0.94  UNITED TECHNOLOGIES CORP      <- Pratt & Whitney (United Technologies)
0.87  CORNING INC                   <- Corning Glass Works
0.80  EXXON CORP                    <- Exxon Research and Engineering Co.
0.69  FORD MOTOR CO                 <- Ford Global Technologies
      ... and five plain matches (Abbott Labs, The Boeing Company, ...)
not linked: Westinghouse Air Brake Company, Abbott Ball Company, General Electrodynamics Corp, Rockwell Automation

Fifty pairs, $0.0007. Jev knows that 3M is Minnesota Mining and that Westinghouse Air Brake is not Westinghouse Electric. It can only judge pairs that blocking proposes, though: "3M Company" shares no letters with "Minnesota Mining & Mfg", so it was found only because the exact:state pass paired firms within a state.

Benchmarks

Five public datasets with known matches, run end to end (blocking, judging, resolving) on 2026-09-18 with Jev 1.13 through OpenRouter. jlink used its default threshold of 0.5 everywhere. Nothing was tuned on the answers. Each string baseline, by contrast, was given the single threshold that maximizes its F1 on the answers, so the baseline column is an upper bound on what that method can do.

The string baselines are best-match Jaro-Winkler and best-match TF-IDF cosine; the table shows the better of the two. Exact matching after normalization scores 0.26, 0.41, 0.00, 0.00 and 0.22.

What the table says:

• Where names carry meaning, jlink wins without tuning. On firms, products and publications it beats string similarity that was handed its best threshold.
• Where records differ only by typos, you do not need it. FEBRL4 is synthetic person data with character-level corruption, and TF-IDF cosine is nearly perfect there and free. jlink made no false links (precision 1.00) but was too cautious at 0.5; at a threshold of 0.3 its F1 is 0.98.
• The firm run was limited by its candidate search. The default forward top-10 n-gram pass proposed 67% of known links, and jlink found 93% of the true links it proposed. The 67% is measured blocking recall, not a ceiling for name-based methods, though ownership links such as "Homogeneous Metals Inc" to "United Technologies Corp" can be difficult to retrieve from names. Candidate search describes reverse search, larger k and their pair-count cost. A sliver of the error is the benchmark's: 11 Compustat names appear under two IDs, which accounts for 14 of jlink's 331 false links.
• Amazon to Google is hard for everyone, because listings differ in version and edition details that the records often omit.

The firm rule was revised once, after reading the errors of a 300-record pilot, as a user would: the first draft did not tell Jev that CPY means Company in these data. That changed F1 by two points. The final rule is two sentences: "Ignore legal-form suffixes (Inc, Corp, Co, CPY, Ltd, GmbH, N V). A subsidiary or division counts as its parent company."

Are the probabilities calibrated? Roughly, and it depends on the data. On the firm benchmark, pairs scored above 0.8 were true matches 95 to 98% of the time and pairs scored below 0.2 were true 0.1% of the time, but the 0.5 to 0.8 band was overconfident (mean 0.64, true 39% of the time). On FEBRL4 Jev was underconfident: pairs in the 0.2 to 0.5 band were true matches 66% of the time. Treat p as a strong ranking and check the middle band with an audit sample before using it as a literal probability.

Speed. The firm run judged 45,567 pairs in 176 seconds (259 pairs a second, 64 calls in flight, median latency 214 ms, one retry). Blocking 100,000 by 100,000 records takes about four minutes and 1 GB on an M3 Ultra. Blocking time grows roughly with the square of the data, so beyond that size a reliable shared field such as state or year can restrict the search: jlink.block.within(jlink.block.ngrams("name"), "state") searches separately inside each group. A separate exact("state") pass is different: it adds pairs to the union and does not split the existing search. Grouping can lose matches when group values disagree or are missing; see grouping, reverse search, and pair limits.

Reproduce everything: uv sync --group bench, uv run python bench/prepare.py, then uv run python bench/live.py nber-firms --budget 1.00. Baselines and data provenance are in bench/BASELINES.md and bench/FIRM_DATA.md.

On a local server

On 2026-09-22 the two local servers, on an Apple M3 Ultra, judged the candidate pairs of 300 left records from each benchmark, the same pairs Jev judged in the full runs above. The local-model results have every number, dataset by dataset.

DiffusionGemma comes within 0.03 of Jev's F1 on firms, publications, products and software and decides 94 to 99% of candidate pairs the way Jev does, at 14 to 18 minutes for each 3,000 pairs. On FEBRL4 it is stricter about typos: at precision 1.0 its recall is 0.72, against Jev's 0.93. Laya puts 90 to 93% of the candidate pairs for firms, products and software at 0.5 or above and is not a substitute for Jev in record linkage. LinkTransformer's pretrained models, with a threshold tuned on labels, match DiffusionGemma within 0.03 on firms and publications, beat it on FEBRL4 and fall far behind on the two product sets, where a cosine cutoff keeps many near-duplicates. Fine-tuned on labels, it links FEBRL4 perfectly and still trails Jev on the other four under the same link rules (Against LinkTransformer). Jev remains the default; DiffusionGemma is the option when records may not leave your machine.

Against LinkTransformer

LinkTransformer (Arora and Dell) is the embedding linker economists reach for: it encodes each record with a sentence transformer, links nearest neighbours by cosine similarity, runs locally and costs nothing per pair. On 2026-09-20 jlink was compared with it three ways. jlink used its default threshold of 0.5 everywhere and saw no labels; LinkTransformer ran locally, pretrained and fine-tuned.

With no labels, jlink beats LinkTransformer's zero-shot models on eleven of the twelve DeepMatcher benchmarks and is level on the twelfth (Beer, where 14 matches make one pair worth three points of F1). These are the 25,235 labelled test pairs behind Table A-5 of the LinkTransformer paper (arXiv:2309.00789), each judged once at 0.5 with a one- or two-sentence rule, for $0.46. Fine-tuned on each benchmark's training labels, LinkTransformer pulls ahead on eight rows, by up to seven points; jlink stays ahead on four, three of them product data, where it also beats Ditto, a fine-tuned language-model matcher. F1 in percent; the published columns are as printed in the paper, and bold marks where jlink beats the fine-tuned LinkTransformer.

The publication rows are pairwise classification: every pair is judged alone, two versions of one paper look like a match, and jlink's precision there is 0.83 to 0.93. Linked end to end with one-to-one assignment, the same judgments score 0.9965 on DBLP to ACM (above).

On firm names, jlink beats LinkTransformer's purpose-built company model. On the NBER to Compustat split, each method was allowed the 2,537 links jlink made and scored on listed matches recovered: jlink 2,034, LinkTransformer's Wikidata company model 1,960, LinkTransformer fine-tuned on the development labels 1,931, character TF-IDF 1,889. The lead holds on the records outside jlink's rule-wording pilot (1,912 against 1,841). The product splits are the closer contest. With development labels to fine-tune on and a threshold tuned on them, LinkTransformer edges past jlink on Abt to Buy (0.954 against 0.942 with one-to-one assignment, where character TF-IDF also reaches 0.951) and on Amazon to Google (0.736 against 0.721), and on FEBRL4 typos it reaches 0.999 where jlink stops at 0.960, as the string baseline above already does.

Only jlink changes its answer when the definition changes. Two small fixtures, hand-labelled from public filings by this repository's author (HP, IBM and Kyndryl, Meta, Kellogg, Kraft, Alphabet and others), ask the same pairs under four definitions: legal entity, corporate group at the record dates, physical site, operating business. jlink answered 30 of 33 and 39 of 40 test decisions correctly. A similarity score is one number per pair whatever the definition, so no threshold on any score can exceed 23 of 33 and 30 of 40. LinkTransformer's company model and all-mpnet-base-v2, with a threshold chosen on development pairs, reached 18 and 23. Pairs within one corporate family are correlated, so the gap is the finding, not the decimals.

These are public benchmarks that Jev, like LinkTransformer's encoders, may have met in training, and each comparison is one split with one seed, so differences of a point or two are noise. Reproduce it with uv run --group bench python -m bench.deepmatcher_pairs --live --budget 0.60 (the Table A-5 pairs) and bench/lt_compare.py with bench/linktransformer_scores.py (the splits). LinkTransformer is GPL-3.0 and runs in its own environment; jlink never imports it.

How it works

1. Block. Comparing every record with every other is wasteful, so jlink first proposes candidate pairs on your machine, at no cost. By default each left record is paired with its ten nearest right records by character n-grams. Add passes for what n-grams miss: jlink.block.initials("name") pairs "IBM" with "International Business Machines", jlink.block.exact("state") pairs everything within a state, and jlink.block.window("year", 1) pairs records whose years differ by at most one.
2. Judge. Each candidate pair goes to Jev with your rule, including equal names: identical text need not identify the same entity. If equal compared fields establish identity in your data, explicitly enable Linker(..., exact_shortcut=True) to accept complete normalized equalities without a call. Likelier pairs are judged first; cached scores remain available after the budget runs out.
3. Resolve. Choose links from the probabilities: one-to-one (the default; the best overall assignment with no record used twice), many-to-one, one-to-many or many-to-many, with a probability threshold and an optional margin over the runner-up. result.relink(...) tries other rules without paying again.
4. Audit. result.audit_sample(200) draws pairs across the probability range, links and non-links alike, with both records side by side. Label them in a spreadsheet, then jlink.evaluate(labeled, mode="selected") evaluates the delivered links and pair-score calibration.

linker = jlink.Linker(
    entity="firm", on=[("conm", "assignee"), "state"], definition="...",
    blockers=[jlink.block.ngrams(("conm", "assignee"), k=10), jlink.block.initials(("conm", "assignee"))],
)
linker.estimate(compustat, patents, left_id="gvkey", right_id="assignee_id")   # blocking only, no API calls
# {'left': 4585, 'right': 2488, 'pairs': 45567, 'dollars': 0.6316, 'seconds': 227.8,
#  'assumptions': {'tokens_per_pair': 330, 'price_per_million_tokens': 0.042,
#                  'pairs_per_second': 200, 'token_basis': 'short_records',
#                  'throughput_basis': 'short_records'}}
# (the real run on these data cost $0.62 and took 176 seconds)

result = linker.link(compustat, patents, left_id="gvkey", right_id="assignee_id", budget=2.00)
strict = result.relink(threshold=0.9, min_margin=0.3)     # no new calls
panel = strict.merged()                                   # both tables side by side, plus p
result.save("linkage/")                                   # links.csv, scores.csv, settings.json

budget=0 allows cache hits and explicitly enabled exact shortcuts only; budget=None is unlimited. A positive budget stops new requests at the observed cost, but calls already in flight can overshoot it. Saved runs retain input fingerprints, blocker parameters, and model identities, including cached answers. See budget semantics and run provenance.

Relations, not only identity

By default the question put to Jev is "Record A and record B refer to the same firm", followed by your definition. Some linkages are not identity. A news article is not a police incident, but it can report one. style="rule" makes your definition the whole proposition: "Record A and record B satisfy the following match rule. ..." The two tables then rarely share columns, so an on item may be one-sided: ("text", None) is shown on the left record only and (None, "neighborhood") on the right record only.

published_soon_after = jlink.block.window(          # article 0 to 3 days after the incident,
    ("published", "occurred"), between=(0, 3), unit="days")           # never before it
result = jlink.link(
    articles, incidents, style="rule",
    definition="Record A is a news article that reports the shooting incident in record B.",
    on=[("text", None), ("published", None),
        (None, "occurred"), (None, "neighborhood"), (None, "victim_age_group"), (None, "fatal")],
    blockers=[jlink.block.within(published_soon_after, "borough")],   # if both tables have one
    left_id="article_id", right_id="incident_id", how="many-to-one",
)

This is a sketch of a design, not a result: it has been run against a fake model in the tests and never against Jev, so nothing is known yet about how well Jev judges this relation. The date logic sits in blocking on purpose. jlink.block.window compares dates and numbers exactly on your machine, so the rule need not ask Jev to do arithmetic, and only pairs inside the window are paid for. how="many-to-one" lets several articles report one incident.

entity is optional under style="rule" because the question no longer names one. result.methods() then describes a relation defined by your rule and does not say the records are the same entity. Identity and rule answers are cached under different questions and never mix. One-sided fields are shown to the judge only; a blocking pass needs a column on each side and says so if given one. See relation linking and windows on dates and numbers.

Dedupe: one table against itself

events = jlink.dedupe(
    articles, style="rule", on=["text", "published"], id="article_id",
    definition="Both news articles report the same shooting incident.",
    blockers=[jlink.block.within(jlink.block.window("published", 3, unit="days"), "borough")],
)
events.clusters                          # id, cluster_id, cluster_size: one row per article
events.labeled()                         # the articles, with cluster_id appended
looser = events.recluster(threshold=0.4) # no new calls, like relink

jlink.dedupe blocks a table against itself, never pairs a record with itself, judges each unordered pair once (earlier row as record A), and groups records into clusters. Joining every pair above the threshold lets one wrong pair chain two unrelated groups together, so the default is average linkage in which every pair between two clusters votes, a pair that blocking never proposed counting as a non-match. That resists chaining and can split a true group that blocking covered only in part; unproposed="ignore" and linkage="components" are the other two rules, and switching is free. report() counts the high-probability pairs the rule left apart, and result.split_pairs() lists them. Dedupe measures both failure modes on synthetic scores, and states what is not known: nothing here measures Jev on a dedupe task, or whether it answers (A, B) and (B, A) alike.

Optional semantic candidate search

Install uv add 'jlink[embeddings]' (or uv sync --extra embeddings in this checkout), then combine local embeddings with character matching:

passes = [
    jlink.block.ngrams(("conm", "assignee"), k=10),
    jlink.block.embeddings(("conm", "assignee"), k=10,
        model="sentence-transformers/all-MiniLM-L6-v2",
        revision="1110a243fdf4706b3f48f1d95db1a4f5529b4d41"),
]
linker = jlink.Linker("firm", [("conm", "assignee")], definition="...", blockers=passes)

The embedding model runs locally; its weights download on first use. The union can recover aliases that character similarity misses, at the cost of more candidate pairs. This is an optional retrieval method, not a claim that any particular encoder beats other systems. See semantic retrieval and benchmark instructions.

Checking the links

sample = result.audit_sample(n=200)
sample.to_csv("audit.csv", index=False)      # fill in is_match with 1 or 0, then:

labeled = pd.read_csv("audit.csv", dtype={"left_id": str, "right_id": str})
ev = jlink.evaluate(labeled, mode="selected")
print(ev.summary())
print(ev.to_markdown())                      # a table for the appendix

The sample is stratified by probability, so the uncertain middle is covered and not only the easy ends, and the estimates are weighted back to all judged pairs. Selected mode uses the sample's saved membership in the final links; the default mode="threshold" instead assesses p >= threshold, before assignment and margin filtering. Recall covers judged candidates only, excluding unjudged pairs and true matches lost in blocking. Brier and calibration always assess pair scores. You may leave labels blank, but keep those rows in the file: the labeled pairs of each probability bin are reweighted to stand for the whole bin, so skipping most of the unlikely pairs does not inflate recall. Blanks that fall on the hard pairs within a bin can still bias the result, and a bin with no label at all leaves the estimates undefined. See evaluation modes and limitations for blank labels, bootstrap assumptions and the exported table contract.

For a local side-by-side review page with accept/reject/unsure decisions, durable history, and offline recomputation, see Local human review. Start with jlink.create_review(result).write_html("review.html") or jev-link review --help.

Command line, Stata and R

jlink estimate compustat.dta patents.csv --on conm=assignee --on state
jlink link compustat.dta patents.csv --on conm=assignee --on state --entity firm \
      --define "A parent company and its subsidiary are different firms." \
      --left-id gvkey --right-id assignee_id --block ngrams:conm=assignee:10 --block initials:conm=assignee \
      -o links.csv --scores scores.csv --report report.md --save linkage/
jlink audit scores.csv --links links.csv --left compustat.dta --right patents.csv --on conm=assignee \
      --left-id gvkey --right-id assignee_id -n 200 -o audit.csv
jlink evaluate audit.csv --mode selected --markdown

--save linkage/ writes what result.save("linkage/") writes: the links, every candidate score, and settings.json with the question, blocking and provenance. jlink review create linkage/ ..., jlink.load and a replication package all read that folder, so the review page is reachable from the command line alone. Stata's rundir() and R's run_dir = forward it. The save folder must be separate from input and output paths, and existing reserved run members must be files; both are checked before blocking or judging.

jlink dedupe firms.dta --on name --entity firm --id gvkey -o clusters.csv --scores scores.csv groups the records of one file, and jlink cluster scores.csv --threshold 0.8 -o strict.csv regroups saved scores without API calls.

--style rule asks the relation in --define and makes --entity optional; --on text= and --on "=neighborhood" are the one-sided fields (quote a leading =: zsh, the macOS default shell, otherwise reads =word as a command lookup and stops before jlink runs). Stata's style(rule) and R's style = "rule" forward the same option. --block window:published=occurred:0..3d is the date window above, --block window:year:1 a numeric one, and --block within:borough:RULE runs any rule inside groups; --date-format reads dates that are not ISO 8601.

The Stata and R wrappers are single files in this repository, not part of the Python package: copy stata/jlink.ado and stata/jlink.sthlp to your personal ado directory (sysdir shows it), and source() r/jlink.R. Both call the installed command, so install the package first.

macOS ships a Java tool at /usr/bin/jlink. If jlink opens a Java prompt, use jev-link, which is the same program, or python -m jlink.

use compustat, clear
jlink using patents.dta, on(conm=assignee state) entity(firm) leftid(gvkey) rightid(assignee_id) ///
    define("A parent company and its subsidiary are different firms.") saving(links.dta)

source("r/jlink.R")
links <- jlink(compustat, patents, on = c("conm=assignee", "state"), entity = "firm",
               left_id = "gvkey", right_id = "assignee_id")

Limits

• These are a model's judgments. Audit a sample before you rely on the links.
• Jev can only judge pairs that blocking proposes. If the names share nothing, add a pass that brings the pair together some other way (exact on state, year or industry).
• Jev reads the fields you give it and nothing else. It does not look anything up, and what it knows about firms stops at its training data.
• Repeated calls return nearly but not exactly the same probability (within 0.03 in our tests). Saved scores make results exact: keep scores.csv with your replication files.
• The default model ID is an alias for the latest Jev. Pin one with model= or --model (for example typesafe/jev-1.13 on OpenRouter) and report it; result.methods() does.
• The Stata and R wrappers were run on macOS against Stata 19.5 and R 4.5.1. They do not support Windows yet, and they call link only: dedupe and the review page need the command line or Python.
• Rule-style relation linking, the window blocker and dedupe are tested against a fake model and synthetic data only. No benchmark in this README covers them.

Development

uv sync --group bench && uv run pytest   # offline, no key; Stata and R tests skip if absent

SPEC.md is the design contract the modules were built against. src/jlink/core.py names the providers jlink offers; the client, retries, answer cache and cost meter are the shared jevkit-runtime, which jgrep and the other JevKit tools use too. jlink's provenance records are documented in run provenance.

MIT license. The benchmark datasets keep their own terms; see bench/FIRM_DATA.md.

Shared JevKit development

This tool uses jevkit-runtime, imported as jevkit_runtime. Clone that repository beside this one as ../jevkit-core, then run uv sync. Core Python edits apply on the next invocation of this tool; restart long-lived Python processes after editing.

The distribution name is jevkit-runtime because jevkit-core on PyPI belongs to a different project. The runtime is available on PyPI. Use the sibling checkout for shared development, or uv sync --no-sources for a standalone source checkout. Existing published versions of this tool are unaffected by this source migration.

From the core checkout, python scripts/dev.py setup, check, and wheel-check set up and validate all five consumers in separate environments. CI checks out core tag v0.3.1. Prompts, question construction, and budget policies remain in this repository; answer identity, the answer store, transport, and metering are the runtime's. Runtime 0.2 keys and stores answers differently from 0.1, so a cache written by an earlier version is re-asked once after upgrading.