Germany Occupation Direct Crosswalk v1

What it is

The 1,300 five-digit occupation units (Berufsgattungen) of the Klassifikation der Berufe 2010 – überarbeitete Fassung 2020 (KldB 2010), published by the statistics department of Germany's Bundesagentur für Arbeit (BA), are linked directly, not via ISCO-08, to an occupation master of 1,020 nodes. The nodes are the 1,016 O*NET-SOC codes in the O*NET® 31.0 Database (O*NET-SOC 2019 taxonomy) and 4 Germany-specific "other occupations" (ACW-DE-NNNN) for which AI Changing Work found no O*NET-SOC counterpart. Relations follow the definitions of the ESCO–O*NET crosswalk (2022) and always read from the O*NET-SOC side.

Three AI judges worked independently: two Anthropic Claude runs using different evidence, and one OpenAI Codex run. A link is confirmed when the third judge took part and a majority held, provisional when it rests on the two Claude runs (or on a rule or a second-round review applied to them) without a third-judge majority, and reviewed when the operator checked it. 1,263 of the 2,574 links (49%) are provisional.

This is AI Changing Work's own mapping (eigene Zuordnung). It is not a classification or correspondence of the Bundesagentur für Arbeit or of USDOL/ETA, and the Bundesagentur für Arbeit has not reviewed it.

• The links do not pass through ISCO-08 (International Standard Classification of Occupations). The occupation descriptions of the two classifications were compared directly.
• The fifth digit of the KldB is the requirement level (Helfer, Fachkraft, Spezialist, Experte). The judges chose down to the requirement level within the same four-digit group.
• Germany-specific occupations judged to have no O*NET-SOC counterpart were not forced into a match; they were added to the master as other occupations (ACW-DE-NNNN).
• Relation definitions follow the technical report of the ESCO–O*NET crosswalk (European Commission × U.S. Department of Labor, 2022), and relations always read from the O*NET-SOC occupation side.
• Used together with the direct crosswalks that link other countries' classifications (Korea, France, Spain, Japan) to the same O*NET-SOC code backbone, it connects those classifications through the same master. Other countries' other occupations are not included in this package.

Key figures

Occupation master

1,020

1,016 O*NET-SOC codes + 4 other occupations (Germany)

KldB units linked

1,300 / 1,300

0 units left unlinked

Links

2,574

match 1,302 · related only 1,272

Status

1,266 · 1,263 · 45

confirmed · provisional · reviewed

Provisional links have no majority that includes the Codex judge: either Codex did not judge them, or it did and no majority including it was reached. To use confirmed links only, filter the status column to confirmed and reviewed.

Of the 1,016 O*NET-SOC occupations, 951 have a KldB link (843 of them a match) and 65 have none. Of the KldB units, 904 have a match.

The English titles are copied unchanged from the BA's English edition. No Korean names are included. The "nearest O*NET-SOC occupation" is a reference; it is not a link (match).

Code scheme: ACW (publisher, AI Changing Work) · DE (ISO 3166-1 country code) · NNNN (serial number in order of listing). The serial number carries no meaning, and once assigned it is never changed or reused. The source code (KldB) is kept separately in the source_code column.

How it was made

• 1a Forward mapping — 923 O*NET-SOC occupations → five-digit KldB units (judges A and B)
• 1b Cross-check — all rows where A and B disagreed + a sample of rows where they agreed, then some low-confidence agreeing rows → judge X (678 rows)
• 1c Residual resolution — rules applied to the rows left after three-judge agreement + a second-round review (judge R) of 107 items
• 3 Orphan reverse mapping — the 468 units that no O*NET-SOC occupation pointed to in step 1 → the 1,016 O*NET-SOC occupations (judges A and B). Units for which no fitting occupation was found became other-occupation candidates
• 4 Review — the operator checked, one by one, the items without judge agreement and the other-occupation listings (25 items)

Step 1 covered 923 = 1,016 O*NET-SOC occupations − 74 "All Other" residual occupations − 19 military (SOC 55); these 93 were reopened as candidates in step 3. Step 3 is the same step as the "orphan validation" of the ESCO–O*NET crosswalk, and X was not used in it (judges A and B only). Because the KldB divides units finely by requirement level, there are many orphan units.

• The judges did not see each other's answers. The KldB material was the German and English source text of the BA files and the O*NET material the English source text, both as published (no translation of our own).
• The judges were given, as candidate hints, the KldB units reached mechanically along institutional correspondence tables (O*NET-SOC code → ISCO-08 → KldB, via the BA's Umsteigeschlüssel). The hints carried unit codes and titles only, with no relations, ISCO codes or paths, and the judges were instructed to search beyond the hints. The hint material and ISCO-08 titles are not included in this table.
• The published files contain codes and titles only. They do not contain Band 2 definition or task statements, O*NET definition or task statements, or the judges' reasons. KldB titles are copied byte for byte from the BA files (no normalization of dashes, spaces or letter case). No Korean names of German occupations are included.

29% of the step-1 forward-mapping links (523 of 1,790) are provisional. They have no cross-model (Codex) majority, so they are less reliable than confirmed links.

• First choices X disagreed with (KEEP): for 67 items, X did not choose the first-choice unit that A and B both chose, and the second-round review (R) kept all of them. Since a judge of the same model family upheld the judgment of the same model family, these are 2/3 but kept as provisional.
• Rule adoption (split, step 1): match relations chosen by X alone that lie on an institutional correspondence path (through the BA's Umsteigeschlüssel) were adopted. That path was also a hint in the judges' input, so it is not independent evidence. All of these links are provisional.
• Other-occupation listings: of the units the judges found to have "no corresponding O*NET-SOC occupation", only those checked by the operator were listed (reviewed).
• Rule checks: 0 violations of "at most one exact per O*NET-SOC occupation". 0 of the 1,300 KldB units left unlinked.

Review

1. Judge self-checks (every batch): row count = input · codes exist (1,300 five-digit KldB units / O*NET-SOC codes in the O*NET 31.0 Database) · allowed values only · count of departures from the two-question rule table · whether the chosen unit was actually opened with the lookup tool (call log check) · BA titles quoted in the reasons only byte for byte, no copying of Band 2 text.
2. Cross-model check: rows where A and B disagreed were sent to X. X did not see the other judges' answers.
3. Human review: orphan reverse-mapping links without judge agreement, low-confidence additions from the second-round review, and other-occupation listings.
4. Load and publication check: the rows in this package were extracted again from the production database, and a separate checker not used for loading or for building the published files rebuilt them row by row from the judgment ledger and the source files (KldB main table and English edition · O*NET 31.0 titles) and compared them. KldB titles were compared byte for byte with the source files.
5. Rule check: 0 violations of "at most one exact per O*NET-SOC occupation".

Limitations

• This is a judgment mapping. It has not been validated against employment statistics. o_in_r and r_in_o (the judges' estimate, made before choosing a relation, of what share of workers in one occupation fall within the other side) are the judges' estimates, not measured headcount ratios.
• The judges are AI models. Human review was limited to items without agreement and to the other-occupation listings. The step-1c rules are not a row-by-row review.
• Provisional links are 49% of the total (1,263 of 2,574). Agreement between two judges of the same model family is not strong evidence. The second-round review (R) is also from the same model family. For Germany, even in the sample of rows where A and B agreed, the share in which X gave a different first choice was higher than for the other countries.
• The requirement level (fifth digit) has no corresponding axis in the O*NET-SOC taxonomy, so it is the judges' estimate. Choices between requirement-level siblings within the same four-digit group are easily split.
• The candidate hints chain institutional correspondence tables, so they do not preserve relations. The judges were required to search beyond the hints.
• Military occupations (SOC 55) and the "All Other" residual occupations were left out of step 1 and were candidates only in step 3.

Changes

• v1 (2026-10-09): First release.

Attribution required when you share this data

AI Changing Work (2026). Germany Occupation Direct Crosswalk v1 (KldB 2010 ↔ O*NET-SOC 2019). https://aichanging.work/data/de-direct-crosswalk. AI Changing Work's contributions licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/); KldB codes and titles © Statistik der Bundesagentur für Arbeit, not licensed under CC BY.

Terms of use and sources

Below is the full text of the package's LICENSE file, reproduced as published. In short: only AI Changing Work's contributions (links, relation labels, match class, stage, consensus and status values, the ACW-DE-* codes and the nearest-occupation references) may be used under CC BY 4.0, to the extent AI Changing Work holds rights in them. O*NET-SOC codes and titles remain under the O*NET® 31.0 Database license (CC BY 4.0) with the attribution below.

KldB codes and titles, including the English titles of the other occupations, are © Statistik der Bundesagentur für Arbeit and are included unchanged under the BA's terms of use. They are not licensed under CC BY: they may be stored and passed on with the source stated, but must not be altered, and commercial use requires the BA's permission. AI Changing Work's non-commercial operation does not exempt your use from this condition.

Download

These files are byte-identical to the published package. You can check a downloaded file against its SHA-256, which is also listed in CHECKSUMS.sha256.