"""Recompute every statistic referenced in the paper from the released CSV snapshot. Single source of truth for paper updates. Outputs: - Spearman rho_s + two-sided p for Stars / VS Code installs / SO questions - Pearson r on log-targets (Appendix robustness check 2) - Leave-one-out range on the GitHub-stars correlation (robustness check 1) - Single- vs combined-factor sub-composite (robustness check 3) - Dirichlet bootstrap (1000 draws from Dir(3.5, 2.5, 2.0, 2.0)) - Inter-factor Spearman matrix on n=50 """ import csv, json, math, random, sys from pathlib import Path from reproduce_table3 import ( spearman, two_sided_p, load_latest_scores, load_latest_signals, load_registry, benchmark_sentiment_subcomposite, ) def pearson(xs, ys): n = len(xs) mx, my = sum(xs)/n, sum(ys)/n num = sum((xs[i]-mx)*(ys[i]-my) for i in range(n)) dx = math.sqrt(sum((xs[i]-mx)**2 for i in range(n))) dy = math.sqrt(sum((ys[i]-my)**2 for i in range(n))) return num/(dx*dy) if dx and dy else float("nan") def main(): csv_dir = Path(__file__).parent.parent / "data" / "csv" scores = load_latest_scores(csv_dir) signals = load_latest_signals(csv_dir) registry = load_registry(csv_dir) eligible = [n for n, c in registry.items() if c.get("github_repo") and n in scores and n in signals] print(f"\n=== Sample size ===") print(f"50 agents in registry, n = {len(eligible)} have github_repo + scores + signals") bs, stars_log, installs_log, so_q = [], [], [], [] for name in eligible: s = signals[name] bs.append(benchmark_sentiment_subcomposite(scores[name], s)) stars_log.append(math.log10((s.get("github_stars") or 0) + 1)) installs_log.append(math.log10((s.get("vscode_installs") or 0) + 1)) so_q.append(s.get("so_questions") or 0) # --- Table 3 (Spearman) --- print(f"\n=== Table 3: Spearman correlations (n={len(bs)}) ===") for label, ys in [("GitHub stars (log)", stars_log), ("VS Code installs (log)", installs_log), ("Stack Overflow question volume", so_q)]: rho = spearman(bs, ys) p = two_sided_p(rho, len(bs)) sig = "p<0.001" if p<0.001 else f"p={p:.3f}" print(f" {label:<32} rho_s={rho:+.3f} {sig}") # --- Pearson on log targets (Appendix Robustness check 2) --- so_log = [math.log10(q + 1) for q in so_q] print(f"\n=== Robustness check 2: Pearson on log targets ===") for label, ys in [("GitHub stars (log)", stars_log), ("VS Code installs (log)", installs_log), ("Stack Overflow questions (log)", so_log)]: r = pearson(bs, ys) print(f" {label:<32} r={r:+.3f}") # --- Robustness check 3: alternative sub-composites --- print(f"\n=== Robustness check 3: alternative sub-composites vs stars ===") bench_only, sent_only = [], [] for name in eligible: sub = json.loads(scores[name].get("sub_scores") or "[0.5, 0, 0]") bench_only.append(sub[0] if sub else 0.5) sent_raw = float(scores[name].get("sentiment") or 0.0) sent_only.append(max(0.0, min(1.0, (sent_raw + 1.0) / 2.0))) print(f" Benchmark only: rho_s={spearman(bench_only, stars_log):+.3f}") print(f" Sentiment only: rho_s={spearman(sent_only, stars_log):+.3f}") print(f" B+S combined: rho_s={spearman(bs, stars_log):+.3f}") # --- Robustness check 1: leave-one-out --- print(f"\n=== Robustness check 1: leave-one-out (GitHub stars) ===") rhos = [] for i in range(len(bs)): bs_loo = bs[:i] + bs[i+1:] sl_loo = stars_log[:i] + stars_log[i+1:] rhos.append(spearman(bs_loo, sl_loo)) print(f" n iterations: {len(rhos)}") print(f" range: [{min(rhos):.3f}, {max(rhos):.3f}]") print(f" median: {sorted(rhos)[len(rhos)//2]:.3f}") full = spearman(bs, stars_log) print(f" full sample: {full:.3f}") print(f" max |delta from full|: {max(abs(r-full) for r in rhos):.3f}") # --- Dirichlet bootstrap --- print(f"\n=== Dirichlet bootstrap (1000 draws from Dir(3.5,2.5,2.0,2.0)) ===") # We perturb the four-factor weights, then re-derive a B+S sub-composite using # the perturbed B and S weights renormalized to sum to 1 over {B, S}. random.seed(12345) def dirichlet(alphas): ys = [random.gammavariate(a, 1.0) for a in alphas] s = sum(ys) return [y/s for y in ys] rho_samples = [] for _ in range(1000): w_B, w_A, w_S, w_E = dirichlet([3.5, 2.5, 2.0, 2.0]) # Resampled B+S sub-composite, renormalized over {B, S} denom = w_B + w_S if denom == 0: continue wb, ws = w_B / denom, w_S / denom bs_pert = [wb * bench_only[i] + ws * sent_only[i] for i in range(len(eligible))] rho_samples.append(spearman(bs_pert, stars_log)) rho_samples.sort() lo, hi = rho_samples[25], rho_samples[975] # 95% interval median = rho_samples[len(rho_samples)//2] print(f" 95% interval: [{lo:.3f}, {hi:.3f}]") print(f" median: {median:.3f}") print(f" min, max: [{min(rho_samples):.3f}, {max(rho_samples):.3f}]") print(f" fraction >0: {sum(1 for r in rho_samples if r>0)/len(rho_samples):.3f}") # --- Inter-factor Spearman on n=50 (Table 2) --- print(f"\n=== Table 2: inter-factor Spearman correlations (n=50) ===") # Build per-agent factor vectors all_agents = [n for n in registry if n in scores] Bs, Ss, As, Es = [], [], [], [] for name in all_agents: sub = json.loads(scores[name].get("sub_scores") or "[0.5, 0, 0]") Bs.append(sub[0] if len(sub) > 0 else 0.5) sent_raw = float(scores[name].get("sentiment") or 0.0) Ss.append(max(0.0, min(1.0, (sent_raw + 1.0) / 2.0))) As.append(float(scores[name].get("adoption") or 0)) Es.append(float(scores[name].get("reliability") or 0)) pairs = [("B-A", Bs, As), ("B-S", Bs, Ss), ("B-E", Bs, Es), ("A-S", As, Ss), ("A-E", As, Es), ("S-E", Ss, Es)] for label, x, y in pairs: print(f" {label}: rho={spearman(x, y):+.3f}") print(f" n = {len(all_agents)}") if __name__ == "__main__": main()