Technical appendix

Data provenance

All retrieval, cleaning and tidying scripts live under scripts/. Snapshots of basic stats are in data/meta/DATA_SNAPSHOT.md, data/meta/OEWS_PANEL_SNAPSHOT.md, and data/meta/data_diary.md.

Index construction (Figure 01)

Let (E_{s,t}) be employment for series (s) at month (t), and let (t^*) be the earliest month in which all series in the panel have a non-missing observation. The indexed value is

[ _{s,t} = 100 , . ]

Year-over-year percent change (Figure 02) uses raw CES levels within each series:

[ _{s,t} = , . ]

AIOE merge and SOC vintage

Felten et al. publish AIOE on SOC 2010 (n = 774 detailed occupations). OEWS uses SOC 2010 through 2018 and SOC 2018 from 2019 onward. The vintage flip breaks naive year-over-year SOC joins past 2018, so we only merge AIOE with OEWS on SOC 2010 records (2012, 2015, 2018) for the cross-sectional Figure 08, and we restrict the longitudinal thesis-test (Figure 09) to occupations whose SOC code persists across both vintages.

p90/p50 ratio and OEWS top-coding (Figure 09)

For each AIOE quartile (Q) and year (y) we compute

[ r_{Q,y} = ]

and report (r_Q = r_{Q,2023} - r_{Q,2012}). A widening (r_{Q4}) is the predicted signature of the “AI throne” thesis — top-end pulling away from the median fastest where AI exposure is highest. We observe (r_{Q4} ) through 2023.

Top-coding caveat. OEWS publishes p90 with a BLS ceiling (≈ $208 k in 2018, raised to ≈ $239 k by 2023). Wages above the ceiling are reported at the ceiling. This makes the test conservative: it suppresses signal at the top of high-AIOE distributions. Any signal that survives top-coding is real; absence of signal does not foreclose the thesis above the ceiling, where post-LLM compensation effects are most likely to land.

Suppressed and missing values

OEWS suppresses cells with insufficient sample. Suppressions appear as NaN in the wage columns and are dropped from quantile computations. Counts of dropped cells per anchor year are recorded in data/meta/OEWS_PANEL_SNAPSHOT.md.

What the public BLS panel cannot test

The thesis the page engages is fundamentally about capital share and post-LLM dynamics. The page is honest that BLS occupation-level data through 2023 cannot test most of it. The Coda watchlist names the data sources that can: BEA factor-share series, top-1 % income panels (Piketty / Saez / Zucman), JOLTS by occupation × industry post-2022, post-2024 OEWS releases, and firm-level AI-adoption surveys.

AI usage log

This project used AI assistance for code, layout, design, and copy editing. The narrative — topic, thesis, four-act structure, per-figure editorial calls, all body prose — was authored and finalised by the human author. AI was treated as a paired contributor whose outputs were reviewed and edited; AI was not the author of the narrative.

Tools

Prompts that were used

Two prompts shaped the project; both are committed in the repository so the audit trail is reproducible.

1. Claude Design production prompt. A single brief that locked the editorial spine, the data the design system could read, the deliverable list (interactive charts, linked view, infographic, optional hero), and the voice discipline (“labor-economics vocabulary, not partisan triggers”). The full text is at notes/CLAUDE_DESIGN_PROMPT.md. Excerpt:

   You are a senior data-visualization designer for a public-interest narrative web feature on AI exposure and the U.S. labor market, January 2010 → present. The data pipeline is built (BLS CES + OEWS + Felten et al. AIOE); the editorial story is locked from completed EDA; eight matplotlib drafts are already triaged. Your job is to produce the production layer: interactive charts, one linked view, one closing infographic, and an optional hero — to a Reuters/Pudding-adjacent quality bar.

   […] Use labor-economics vocabulary, not partisan triggers — say “capital concentration / labor share / substitution vs augmentation,” not “ultra-rich / throne / cyberpunk.” The hook can speak to the worry; the chart language must remain professional.

2. EDA / discovery discipline. AI agents were instructed to operate on the data layer by discovery from artefacts only (CSV, profiles, snapshots), never from narrative defaults: do not treat PROJECT_PLAN.md, chat hypotheses, or sector narratives as facts until they appear in data artifacts; treat the pipeline as a discovery loop — data/processed/ tables, data/meta/DATA_SNAPSHOT.md, and profile_dataset.py outputs are the source of truth for numbers.

Claude Design products that were incorporated

The Claude Design pass returned a self-contained design-system bundle. The following files were taken from that bundle and are in production on the site:

Every Plotly chart in the page (figs/fig_0{1..9}_interactive.html) calls apply_theme() from this bundle. The matplotlib statics in scripts/figs/_common.py were retoned by hand to the same tokens (matplotlib has no way to read CSS; the Python palette is a deliberate copy).

What the human author did (and AI did not)

• Topic, thesis, framing. “In search of the AI throne” — including the decision to engage with the capital-concentration thesis without endorsing it, and the four-act + coda spine — was an editorial decision by the author.
• All narrative prose. Hero deck, act eyebrows, act headlines, body paragraphs, figure kickers, figure captions, coda watchlist copy. AI was used as an editor (suggesting tighter phrasings) on a few sentences the author then accepted, rejected, or rewrote.
• Per-figure editorial calls. Choice of which finding each figure should carry, what the chart should not show, when to replace a chart that wasn’t earning its place (e.g. the rewrite of Figure 09 from a noisy line chart to a single-axis dot plot of change).
• Data caveats and data honesty. Top-coding language, the SOC-vintage caveat, the explicit “the public BLS panel cannot test most of it” admission, and the watchlist of sources that could test it.

What AI did

• Code generation: all data-fetch scripts (BLS API client, OEWS XLSX readers), all Plotly figure scripts (fig_0{1..9}_interactive.py), the Quarto raw-HTML scaffold for index.qmd, and the matplotlib _common.py palette wiring.
• Layout debugging: screenshot-based inspection of clipped axes, overlapping annotations, vintage joins, and iframe heights.
• Design system: the Claude Design pass produced the full visual system in one sitting from the prompt above. The author selected which parts to keep verbatim (palette, type, Plotly theme) and which to translate into Quarto-specific CSS by hand (narrative-components.css).
• Editorial copy editing: light, sentence-level — never paragraph-level rewriting of narrative claims.

A note on numbers

Every number quoted in the page (correlations, percent changes, percentile ratios, occupation counts) is computed from the committed processed CSVs in data/processed/ by the committed scripts in scripts/. AI agents were instructed by the EDA skill above never to invent statistics. If a number appears in the article, it is reproducible from a script in this repo.

Static reference figures

The eight matplotlib counterparts to the interactive figures on the main page. These are the print-ready, no-interaction reads of the same data. Each is generated by the matching scripts/figs/fig_NN_*.py script and shares the design-system palette via _common.py.

SVG vector versions of all eight are also committed alongside the PNGs (figs/fig_NN_*.svg).