Global Automation Atlas

Public replication package for Global Automation Atlas by Prashant Garg, Tommaso Crosta, and Jasmin Baier.

Atlas website | Paper | Data downloads

This repository contains retained analysis data, source-data extracts, and code for checking the results reported in the paper. It starts from the retained task-country labels and their accompanying non-empty rationales, then rebuilds downstream numerical checks, tables, and figures. Exact submitted versions are retained as reference snapshots, while the manifest distinguishes direct builds from numerical or source-data check renders.

The package does not include exploratory drafts, old figure variants, raw API logs, annotation batches, or embeddings. For the rationale-concept analysis reported in the paper, it includes the analysis-ready rationale corpus and compact source-data tables for the fixed concepts, discovery scores, concept-fidelity scores, and heldout paired estimates for exposure, substitution-only, and augmentation-only contrasts.

Quick start

From the package root, first check that the expected files are present:

Rscript code/check_replication_package.R

Then run the reproduction workflow:

Rscript code/make_all.R

If the Python requirements are installed in a virtual environment, point the workflow to it with, for example, PYTHON=.venv/bin/python Rscript code/make_all.R.

The workflow writes rebuilt files to reproduced/. It also runs a numeric audit of paper-presented values. If any audited number fails to match after rounding, the script stops and points to reproduced/checks/numeric_claim_audit.csv.

What the scripts check

• code/check_replication_package.R checks the core inventory and prints row counts for the task-country label file.
• code/make_all.R rebuilds summary tables and selected figures from the included data.
• code/04_numeric_audit.R checks headline values reported in the paper, including country exposure ranges, validation correlations, gender-gap summaries, and fixed-effect coefficients.
• code/08_rebuild_gender_table_B8.R re-estimates Supplementary Table B.8 from the released two-digit employment weights and country-cell exposure panels.
• code/05_rationale_concept_audit.R checks the HypotheSAEs discovery, fidelity, and heldout paired-estimator fields for all three paper-facing contrasts.
• code/06_refresh_audit.R writes manifest.csv, mapping each active manuscript asset and reported statistic to its public reproduction check, package artifact, source data, and checksum where applicable.
• code/manuscript_figures/build_fixed_concept_appendix_figures.py directly rebuilds the three 20-concept HypotheSAEs appendix figures from the released 60-row discovery/fidelity/heldout table.
• code/manuscript_figures/build_combined_rf_rationale_main_figure.py directly rebuilds current Figure 6 from the released five-seed TreeSHAP/ALE and pooled-rationale tables.
• code/manuscript_figures/build_openai_observed_use_iwa_validation.py rebuilds the observed-use validation from the public OpenAI Signals files, O*NET task hierarchy, and retained U.S. task labels.
• code/manuscript_figures/build_current_check_renders.py writes current numerical/source-data checks, including the five-seed direction audit and gender/informality panels.
• docs/figure_asset_registry.csv is the authoritative 40-active-figure mapping. reproduced/checks/figure_reproduction_status.csv validates it against the manuscript snapshot.

At the time of this package build, the numeric audit checks 59 rounded paper values.

Folder guide

• prompts/ contains the country-conditioned, income-group, and context-free prompt protocols, plus model notes.
• data_intermediate/ contains the retained measurement outputs used downstream. The central file is task_country_labels_analysis.parquet, with 2,330,776 task-country observations across 124 countries and 18,797 tasks.
• data_intermediate/task_country_rationales/ contains the 2,320,863 non-empty short rationales accompanying those labels, stored in eight ordinary Parquet partitions.
• data_analysis/ contains smaller panels used directly in the country, channel, predictor, occupation, industry, and gender analyses.
• outputs/source_data/ contains source data for paper and supplementary figures and tables where a compact source file is available.
• outputs/source_data/rationale_concepts/ contains the current 60-concept discovery, fidelity, and heldout estimates, the 18 pooled families shown in Figure 6, the place-mask vocabulary, and source data for the illustrative examples table.
• outputs/figures/ and outputs/tables/ contain exact reference snapshots of the active manuscript artifacts. Their capture and original production provenance are documented in docs/provenance/REFRESH_PROVENANCE.md; they are never substituted for a rebuilt result without that distinction being explicit.
• reproduced/ is created by code/make_all.R and contains rebuilt checks, tables, and figures.
• docs/data_dictionary/ gives column descriptions and types for the main public data files.
• data_raw_public_metadata/source_inventory.csv lists public source datasets and links.

Language-model labelling step

The package documents the language-model labelling protocol but does not rerun it. The API stage depends on model availability, model versions, and paid external services. For reproducibility, the package supplies the retained labels used in the paper and the prompts needed to inspect the measurement protocol.

The prompt files are:

• prompts/country_conditioned_prompt.md
• prompts/income_group_prompt.md
• prompts/context_free_prompt.md
• prompts/model_config.json

The retained label file keeps the columns needed for the paper analyses: exposure level, labour margin, channel, AI materiality, AI function, country identifiers, and task identifiers. The accompanying rationale text is stored separately in data_intermediate/task_country_rationales/ to avoid duplicating label and metadata columns. The compact rationale-concept source bundle in outputs/source_data/rationale_concepts/ documents the rationale-derived results reported in the paper. Raw API logs, annotation batches, and embeddings are not required by the public reproduction workflow and are not included.

Data dictionaries

The main dictionary is docs/data_dictionary/schema_with_descriptions.csv. It reports file names, column names, inferred column types, and descriptions for the main files in data_intermediate/ and data_analysis/. A shorter hand-written dictionary is also available at docs/data_dictionary/data_dictionary.csv.

Figure and table reproduction

The package distinguishes three figure reproduction classes.

First, outputs/figures/ contains the exact figure files used by the paper. These files are included so that readers can recover the submitted visual evidence exactly.

Second, exact_reproducible_build figures are generated by package-relative adaptations of the original plotting scripts. The three fixed-concept figures, combined Figure 6, and OpenAI observed-use alignment are checked against the active snapshots at 150 dpi; PDF byte hashes can differ because regenerated files carry different metadata and font-object encodings.

Third, numerical_source_data_check_render figures reproduce the values from released source data while allowing expected differences in layout, cartography, typography, or composition.

The same logic applies to tables. Final manuscript table files are in outputs/tables/; rebuilt check tables are in reproduced/tables/. Supplementary Table B.8 is re-estimated from the released inputs in outputs/source_data/ilostat_gender/table_B8_inputs/ and checked against gender_fe_table_current.csv. The gender-gap decomposition figure and its reported component means are checked from the two released decomposition files in outputs/source_data/ilostat_gender/.

Data sources

The analysis uses public source data from O*NET, the World Bank, Penn World Table, Barro-Lee, Worldwide Governance Indicators, CEPII BACI, ILOSTAT, the IMF AI Preparedness Index, Eurostat, and external exposure measures cited in the paper. The file data_raw_public_metadata/source_inventory.csv lists source links and roles in the analysis.

Requirements

The reproduction scripts use R and Python. R dependencies are arrow, dplyr, readr, tidyr, ggplot2, scales, stringr, and fixest; Python dependencies are listed in code/requirements.txt. Poppler's pdftoppm is used for the rendered-pixel comparison. No API keys are needed.

Software checklist details

The package provides source code rather than compiled standalone software. It includes the retained task-country labels, compact analysis panels, and source-data files needed to demo and reproduce the reported checks.

Tested environment:

• macOS 26.1
• R 4.5.3
• arrow 23.0.1.2
• dplyr 1.2.1
• readr 2.2.0
• tidyr 1.3.2
• ggplot2 4.0.2
• scales 1.4.0
• stringr 1.6.0
• fixest 0.13.2

No non-standard hardware is required. A normal desktop or laptop is sufficient. No API keys are required for the public reproduction workflow.

To install the required R packages, run:

install.packages(c("arrow", "dplyr", "readr", "tidyr", "ggplot2", "scales", "stringr", "fixest"))

python3 -m pip install -r code/requirements.txt

Typical install time on a normal desktop computer is under 10 minutes when binary packages are available, but can be longer if arrow must be compiled from source.

To run the demo inventory check:

Rscript code/check_replication_package.R

Expected output reports the number of task-country labels, figures, tables, and retained rationale concepts. This check normally runs in under one minute.

To reproduce the included tables, selected figures, and numerical audits:

Rscript code/make_all.R

Expected output is written to reproduced/, including rebuilt tables, selected rebuilt figures, reproduced/checks/numeric_claim_audit.csv, reproduced/checks/figure_reproduction_status.csv, and reproduced/checks/rationale_hypothesaes_audit.csv. On a recent desktop, the workflow usually runs in a few minutes.

Known limits

The package is designed to reproduce the paper results from retained labels and analysis-ready rationales, not to repeat the original paid API labelling and annotation calls. Raw API batches, embeddings, exploratory outputs, and restricted third-party microdata remain outside the package.

Citation

If you use this package, please cite the paper. Citation metadata are provided in CITATION.cff. The current paper version is available at automationatlas.org/paper.

Licence and third-party data terms

The code in this package is released under the MIT License; see LICENSE. The data files combine original constructed measures with cleaned extracts or derived variables from public third-party sources. Those source datasets remain governed by their own terms. See docs/DATA_USE_NOTICE.md for source-specific notes and attribution guidance.