Mercor Labor Index maps 29 benchmarks to 18,797 O*NET tasks, covering 10% of US working hours

The map groups 18,797 tasks into 22 job families and shades each occupation by share of tasks linked to benchmark prompts. Six families show zero occupations reaching coverage. Logistic regression on task-level data indicates coverage probability rises with occupational wage bill and with AI-exposure tags from Eloundou et al. 2023, producing systematic under-sampling of middle-income roles such as legal secretaries and court clerks despite comparable exposure scores to lawyers.

BLS wage and employment figures converted to annualized wage bills confirm the skew: high-wage-bill legal and software tasks receive dense prompt mapping while lower-wage administrative and clerical tasks remain near zero. This pattern matches the selection criteria listed—public availability, agentic format, economic realism—yet produces an index that tracks headline labor markets rather than total hours worked.

The resulting 90% uncovered share implies current leaderboards cannot forecast deployment impact on the majority of US tasks. Expansion plans stated by Mercor target the long tail, but the initial methodology already encodes a wage-bill prior that will require explicit correction to achieve representative coverage.

Operational consequence is straightforward: model evaluations remain informative only inside the 10% slice already benchmarked; outside it, performance claims rest on untested extrapolation.