Methodology & sources

Singulariki is a source-backed encyclopedia of work. Every page is a projection of one labor graph that joins the public occupational record to the current research on how AI touches work. Nothing here is hand-written opinion: each figure traces to a named public dataset and the exact release shown below and stamped on every page.

What this is

The underlying object is not a set of articles — it is a graph: role × task × skill × knowledge × ability × tool × industry × wage × outlook × AI exposure × observed AI usage. A page is an aperture onto that graph. A role page renders everything the graph knows about one occupation; an industry, skill, or task page renders the same relations from a different vertex. Because the pages are projections of shared data, the numbers reconcile across the site by construction.

Data sources

Each page also stamps its own dataset and version inline, so you can always see the exact provenance of a specific figure. This table is the master list. Data exported as of June 2026 (snapshot 96022747d398) — Singulariki recomputes on each export, not live.

What the AI numbers mean — and what they don't

The single most-misread idea in the AI-and-work conversation is treating exposure as a forecast of job loss. This site keeps the distinctions explicit.

• AI exposure is not automation, adoption, or job loss. It measures where AI could assist or overlap with an occupation's tasks. High exposure most often means the work is amenable to augmentation — people working faster with AI — not the role disappearing. Exposure says nothing about timeline, cost, regulation, or whether a task will actually be handed off.
• The comparable AI-exposure band is built the same way everywhere. The Low / Moderate / High band shown across roles, job families, occupation groups, career clusters, education levels, job zones, and fields of study is the average of two published studies (\u201CGPTs are GPTs\u201D and AIOE) over an occupation's tasks, then ranked into percentiles. Same two studies, same percentile method — so the bands are comparable across every grouping by construction.
• Industry exposure is employment-weighted; occupation grouping bands are not. For an industry we weight each member occupation's exposure index by its employment in that industry (the BLS national industry-occupation matrix), so the figure reads the sector's real labor structure rather than a flat average of its occupations. The occupation-grouping lenses above (clusters, education levels, job zones, and so on) use an unweighted mean across member occupations. Both are percentile-banded, but the industry method is deliberately different — stated on the page and on the AI exposure hub — and the two are not directly comparable to each other.
• Every occupation and industry also carries a single exposure index. Alongside the band, each occupation has a unified exposure index (the same two studies folded together) with a percentile across all scored occupations, surfaced on role pages, the occupations index, search, and the by-occupation ranking. It is a convenience ordering of the same underlying signal, not a new measurement.
• Observed AI usage is a sample, not a census. The Anthropic Economic Index figures are shares of measured Claude.ai conversations mapped to an occupation's tasks — not a measure of the whole workforce, the whole market, or every AI tool. A task showing a high usage share means it appears often in that sample, weighted toward people who already use an AI assistant.
• Augmentation vs. automation is a behavior label, not an outcome. Anthropic rates each conversation's collaboration pattern. We group directive and feedback-loop use as automation-leaning and iteration, learning, and validation as augmentation-leaning. It describes how people worked with the model, not what happened to the job.
• Projections are pre-AI models. BLS employment projections are the official decade outlook, built before AI's full effect is known. They are the best available baseline, not a prediction that accounts for AI.
• Percentiles are within their reference set. A \u201C75th percentile\u201D wage or exposure rank is relative to the occupations (or dimensions) being compared on that page, and the reference set is stated next to the number.

Coverage and known limits

• O*NET rates 1,016 detailed occupations. Some derived measures cover a smaller subset — for example a given skill may be rated across roughly 894 occupations. Where a denominator differs, the page names it (\u201Crated occupation subset\u201D) rather than implying full coverage.
• Not every occupation has an observed-AI-usage profile: the Anthropic sample only matches tasks for occupations that appear in it. Pages omit the section rather than show an empty or zero-signal table.
• Wages are national and cross-industry (OEWS, May 2024). Local pay varies; the site does not yet carry metro-level wages.
• Every figure is a reference estimate from public data, not personalized advice. Treat it as a starting map, not a decision.

How a page is built

The pipeline lands each source in an R2 Iceberg lakehouse, joins them on the O*NET SOC code into governed gold tables, and projects bounded render models to the edge. Pages are served from Cloudflare. The same generated data drives every cross-link, so a claim on a role page and the matching claim on an industry or skill page come from one number, not two.

Found a figure that looks wrong or a caveat that should be sharper? That is exactly the kind of correction this project wants — the whole point is that the data, not the prose, is in charge.