Forecast the Frontier, Not the Aftermath: Capability Fronts, Latent Displacement, and the Velocity of AI Occupational Change

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Artificial-intelligence labour-market measurement generally begins with occupations, tasks, observed adoption or realised employment outcomes. This paper begins further upstream: with the movement of the AI capability frontier itself. It argues that the observable labour event may occur substantially after the technological event that causes it. Task content can become economically compressible while headcount remains unchanged, producing latent displacement that becomes visible only when accumulated change crosses an organisational threshold. The result is a distinction between employment survival and occupational security: a role may remain staffed after its core human function has ceased to be structurally necessary. The paper develops a capability-gated model of occupational change in which role-functions are constrained by binding technical requirements rather than additive exposure scores. It separates core capability from assurance, robustness, deployment conditions and jurisdictional permission; distinguishes frontier broadening from directional cross-role transfer; introduces the nibble interval and adaptation backlog; and develops the concepts of responsibility compression and jurisdictional automation. A preregistered prospective pilot uses a seeded O*NET held-out sample and explicitly embargoes labour-market outcomes from capability scoring. Five prospective tests are locked: three role-specific capability crossings, one frontier-broadening test and one artifact-lineage directional-transfer test. Each carries predetermined PASS, FALSE or unresolved conditions and publicly observable validation instruments. The framework does not attempt to estimate an economy-wide percentage of jobs that will disappear. Its empirical proposition is narrower and falsifiable: if changing AI capability is a cause of occupational change, movements of the relevant capability frontier should contain information about subsequent labour adjustment before that adjustment is fully visible in labour statistics. The paper therefore proposes forecasting the frontier rather than fitting models to its aftermath. Supplementary research material The accompanying workbook, AI-Occupational-Change-Blind-Test-Sheet_v7_prereg_final.xlsx, contains the prospective empirical apparatus for this paper. It records the seeded O*NET held-out sample, capability ontology and stress testing, capability-state rules, calibration and held-out role-functions, public observability instruments, locked PASS/FALSE criteria, and the final preregistered prediction set. The preregistration was closed on 14 September 2026, before publication and before labour-market outcomes were opened for validation. It contains five prospective tests: three single-role capability crossings, one frontier-broadening test, and one artifact-lineage directional-transfer test. Subsequent labour outcomes are therefore reserved as downstream validation evidence rather than inputs to the capability assessment. The workbook is published alongside the paper to preserve the audit trail and permit later comparison between the locked capability-front predictions and realised outcomes.