# Step 195 robustness results (archived)

Source: GitHub Actions run 41934425228, printed to the job log by the v196 workflow and archived in
`data/processed/ai_adoption_interaction_robustness_v195.json`.

| Check | n / runs | β (adoption × exposure) | p |
|---|---|---|---|
| Full sample | 848 | −0.545 | 0.315 |
| Matched country–occupation cells (368 cells) | 736 | −0.686 | 0.473 |
| Leave one country out | 13 runs | −0.839 … −0.201 | 0.082 … 0.712 |
| Leave one occupation out | 39 runs | −0.818 … −0.332 | 0.099 … 0.581 |

Reading: the sign is negative in all 54 specifications and significant at 5% in none. The point estimate is
stable to dropping any single country or occupation and to restricting to matched cells, so no single unit
drives it, but the data cannot distinguish it from zero. This supports the existing wording: a weak,
non-significant negative association; no calibration and no causal claim (`calibration_allowed=false`,
`causal_claims_allowed=false`). Only two annual transitions are available.