The model tracks two things: labor shares — how many people work in each territory, globally — and replaceability — how technically capable deployed AI or robotics is at performing the core tasks of each occupation, scored 0 to 100.

These are separate measurements, and v5 keeps them strictly separate. Replaceability = tech capability × economic viability × commercial availability. A high score means the technology exists and works, the economics are plausible at realistic wage points, and a product is on the market. Replacement — how fast that capability translates into actual workforce displacement — is a different calculation that sits in the replacement formula. It lags replaceability by years or decades depending on regulatory gates, human preference, capital cycles, and institutional drag. A territory can score 90 on replaceability while employment has barely changed; that gap is not a flaw in the model, it is one of the most important things it shows.

Every occupation is decomposed into tasks (2–14 tasks per occupation, 4,818 total across 480 occupations). Each task is assigned to one of six capability vectors — routine cognitive, generative cognitive, physical automation, selective physical, unstructured physical, system engineering — and given a difficulty threshold and a time weight. A task is "replaceable" when its vector's capability value at that year exceeds its difficulty. The occupation's replaceability is the time-weighted share of replaceable tasks. Territory replaceability is the employment-weighted mean of its in-territory occupations.

Labor data from 1800 to 1870 is reconstructed from Bairoch, Maddison, and Mitchell. From 1870 to 1990 the GGDC 10-Sector Database + ILO historical estimates are splice-adjusted at 1991 to ILOSTAT. From 1991 onward, data uses ILO modelled estimates at the ISIC Rev. 4 section level across 189 countries, with World Bank / IISS Military Balance filling Governing & Protecting. Task decompositions draw from O*NET, BLS OES/OOH, and live 2025–2026 LinkedIn scrapes.

The 2026 capability values (C_R=0.76, C_G=0.57, P_A=0.75, Phi_S=0.46, Phi_U=0.15, S_E=0.35) were set by two independent recalibration instances under an honest deployment lens. The 1970 → 2041 trajectory extending these values was produced by parallel forecasters + reconciler at both ends — forward (2026 → 2041, Phase 7) and backward (1970 → 2025, Phase 12); 2025 was re-anchored on August 15, 2026 by a three-instance research protocol across two model families (see the full methodology §6.3). The replacement formula (conversion rate 0.30, piecewise 15-year lag schedule, barrier-stratified multipliers) was empirically fit against seven historical automation cases and the Anthropic Economic Index cross-section.

The dataset was produced through a reviewer-directed 13-phase pipeline using adversarial multi-instance separation — builders never validate their own work; separate instances do auditing, reconciliation, and documentation. Three Phase 11 audits removed 825 scrape artifacts and reclassified hundreds of mis-vectored tasks. The full methodology documents every parameter and every source; every number in the dataset should be explainable, and where it isn't, that is a known limitation documented explicitly.

The model is deliberately reductive and intentionally incomplete. It does not model job creation, policy responses, or economic restructuring. It shows the displacement pressure. What happens next is a political and societal question, not a technical one.