Occupation navigator

Reproduce the paper analysis

Documentation and computational audit: OpenAI Codex, 6 October 2026; not independent human validation.

Rebuild the real benchmark rankings and six recovery comparisons from released occupational scores, public O*NET relations, frozen candidate mappings and aggregate outcomes. Reconstruct the Census crosswalks and render the public paper figure.

The posting-task worked example uses clearly labeled synthetic job-months. Full reconstruction of the posting-task matrix requires licensed NLx aggregates; the supplied reconstruction was tested locally against authorized aggregates and matched the released matrix. The package contains the actual research source code, but no NLx postings or survey person records.

• Download complete public package
• Instructions, reproduction boundaries, equations and data dictionary
• Full occupational-relatedness score matrix — 1,016 specialties, compressed NumPy archive
• Excluded occupation-task pairs and both model decisions — 19,682 pairs; public taxonomy labels only
• Two-model review protocol · review counts and provenance
• Original public source releases to benchmark outcomes — executed and verified against 260,491 observed pairs
• Fresh matrix-to-ranking-to-benchmark code · synthetic construction example · actual crosswalk rebuild
• Completed verification · File checksums · Original production source paths and hashes

python reproduce.py --output results reconstructs all 2,720 model-by-origin lists before checking frozen results. Expected answers are assertions, not calculation inputs.

Results are retrospective external-data comparisons. RCA selection used benchmark feedback. Different source populations, time periods, occupation systems and outcome definitions limit generalization.