# Contract — F6-conf-03

Question. How strongly does the quarterly change in the US unemployment rate respond to quarterly real GDP growth (Okun's law, difference form), 1960Q1–2019Q4, and is the response stable across 1960Q1–1983Q4 vs 1984Q1–2019Q4?

Estimands. β (and α) in Δu_t = α + β g_t + ε_t, with u_t = mean of 3 monthly UNRATE values, Δu_t = u_t − u_{t−1}, g_t = 100·Δln GDPC1_t (not annualised). OLS estimates a linear projection coefficient; it is a descriptive/reduced-form association (not causal), meaningful under stationarity of (Δu, g) within the sample. HAC Newey–West (Bartlett) SE; 95% CI. Subperiod slopes; classical Chow SSR F-test, df (2, n−4), valid under homoskedastic, serially uncorrelated normal-ish errors (a known caveat).

Outputs. out/results.json (keys: sample_start, sample_end, n_sub1, n_sub2, baseline_metric, model_metric, effect_estimate, intercept, effect_se, hac_lags, effect_ci_low, effect_ci_high, beta_sub1, beta_sub2, chow_f_stat, chow_p_value, stability_conclusion, conclusion, task_id, feasibility, feasibility_reason, data_issues, n_input_rows, n_analysed, random_seed); out/figure1.png; out/table1.csv (period,n,intercept,slope,slope_se; 3 rows); out/report.md (6 sections); out/analysis.py; out/claims.json; plus protocol state files and out/exclusions.csv.

Information rules. Only data/GDPC1.csv and data/UNRATE.csv. Sample exactly 240 quarters 1960Q1–2019Q4 (1959Q4 levels used only for first difference). Nothing after 2019Q4 enters any estimate.

Invalidity/impossibility. Missing in-sample observations, hash mismatch, wrong sample count (≠240, 96+144), use of post-2019 data, CI not containing estimate.