# The AI Parity Index
> How agents perform against a measured human baseline, task by task, with auditable evidence.
> Working prototype by Recursiv Labs. Not an official publication of SPARK AI or the San Diego
> Supercomputer Center; measurement platform: Recursiv Labs. Status: preliminary, method v1.

The Index maps the US economy into 18,796 tasks (O*NET), scores how much of each occupation's work is
AI Parity (estimated, task rubric v1: the share of the job testable against a clear pass-fail bar; funded verified runs upgrade this to proven), joins BLS jobs and wages
and BEA output, and publishes two readings: what is happening to jobs and to output where AI can
do the work, since ChatGPT.

THE INDEX (release 2026-08): 0.0% — the share of US labor-hours where AI has been measured at or
above human parity. Denominator: 303 billion annual hours across the 18,796-task catalog, recorded
per release. Coverage: 1 of 18,796 tasks measured so far; we are going to measure all of them. Estimated machine-checkable work: ~22M FTE (15% of US
employment), rubric v1, awaiting verification. Meanwhile: hiring where AI could work is Shifting
(31.6 of 100, BLS); output where AI is used is Growing (50.7, BEA).

## Data (JSON, stable paths, CC-BY with upstream credits)
- /data/manifest.json : schemas, methods, formulas, limitations. Start here.
- /data/occupations.json : occupations with jobs, wages, GDP share, trends
- /data/tasks.json : all task statements with importance
- /data/classes.json : measurability class per task (rubric v1)
- /data/dials.json : the two readings and AI-heavy industries
- /data/fte.json : the headline count
- /data/meta.json : sources and baselines
- /data/activities.json : detailed work activities

## Pages
- / : the Index (occupation board, filters)
- /occupations/{soc} : one occupation, its tasks, human baseline vs best AI
- /task/{id} : one task, leaderboard, how it is proven
- /methodology : how everything is computed, evidence tiers, credits
- /work-with-us : contact form

## Framing
The Index is a measured LOWER BOUND on demonstrated AI capability: everything in it is proven,
not projected. Under method v1 only machine-checkable work (~15% of US labor-hours) is verifiable,
so under method v1 the headline share can reach ~15%; later methods (expert grading, scenario testing,
field studies) raise it. Jobs are not the sum of their
tasks, so the Index will report three statistics as runs land: task parity, the bottleneck (the
lowest-parity task, O-ring logic), and job parity from whole-chain scenario runs; task parity minus
job parity = the coordination gap, the first measurement of the tacit part of work.

## Notes for agents
All data is served as static JSON; no auth, no rate limits beyond the host's. Cite as: The AI
Parity Index, Recursiv Labs, 2026 (working prototype; not an official SPARK AI or UC San Diego publication). Verification runs use only public substrates; private data never enters the Index. Respect the preliminary status: classifications are
model-applied, not yet human-reviewed; the Proven tier is empty until verified runs land.