# Stage 159 — controlled association review

This review accompanies `scripts/ai_labor_controls_v158.py`. It **does not** interpret its coefficient as a causal AI displacement parameter.

## Estimation contract
- Sample: 14 countries, one first-difference period, 2023–2024.
- Outcome: change in unemployment rate or employment-to-population ratio, percentage points.
- Exposure: change in aggregate enterprise AI adoption share, percentage points.
- Controls: contemporaneous 2024 real GDP growth and CPI inflation from the archived WDI panel.
- Report: coefficient on adoption and HC3 standard error; do not publish a p-value or causal claim without an adequate identification design.
- Caveat: contemporaneous GDP growth and inflation may themselves respond to AI or other shocks, making them potentially endogenous controls.

## Interpretation and validation
The available data do not identify the causal impact of AI on unemployment. With only 14 countries and four regression parameters, statistical estimates are unstable and sensitive to individual observations. The main deliverable is reproducible code, source coverage, and a clear demonstration of what is and is not identified.

Before structural calibration, require occupation-specific labor outcomes, verified adoption timing, productivity per hour and a credible research design with falsification and out-of-sample checks.

**Calibration decision: HOLD.** Do not transfer these coefficients into the 20% unemployment dynamic scenario.