1"""Emit paper/tables/numbers.tex: every number quoted in the prose comes from this file (registry, configs, state store)."""2import os, sys, csv, json, yaml3sys.path.insert(0, os.path.dirname(__file__)); from common import *4def m(name, val): return f"\\newcommand{{\\{name}}}{{{val}}}"5def main():6 reg = list(csv.DictReader(open(os.path.join(REPO, "tasks", "registry.csv"))))7 conf = [r for r in reg if r["split"] == "conf"]; dev = [r for r in reg if r["split"] == "dev"]; tr = [r for r in reg if r["split"] == "transfer"]8 b = yaml.safe_load(open(os.path.join(REPO, "configs", "budgets.yaml")))["budget"]9 out = [m("nConfTasks", len(conf)), m("nDevTasks", len(dev)), m("nTransferTasks", len(tr)), m("nConfRealData", sum(r["real_data"] == "yes" for r in conf)),10 m("nConfSourceDocs", sum(r["source_documents"] == "yes" for r in conf)), m("nConfNullTruth", sum(r["infeasible_or_null_truth"] == "yes" for r in conf)),11 m("nOracles", len([r for r in reg if r["split"] in ("conf", "dev")])), m("envUSD", f"{b['max_budget_usd']:.2f}"), m("envWall", f"{b['wall_seconds']:,}"), m("envTurns", b["max_turns_per_session"])]12 try:13 s = pd.read_csv(os.path.join(DATA, "sessions.csv")); r = pd.read_csv(os.path.join(DATA, "runs.csv"))14 sm = s[s.run_id.str.startswith("RD") & (s.kind == "initial")].merge(r[["run_id", "treatment"]], on="run_id")15 for a in ARMS:16 v = sm[sm.treatment == a].first_turn_input_tokens17 out.append(m(f"ctx{a}", f"{int(v.iloc[0]):,}" if len(v) else "--"))18 for split, tag in [("confirmatory", "Main"), ("pilot", "Pilot"), ("transfer", "Transfer"), ("recovery", "Recovery"), ("ablation", "Ablation")]:19 d = r[r.split == split]20 out += [m(f"n{tag}Runs", len(d)), m(f"n{tag}Evaluated", int((d.status == "evaluated").sum())), m(f"cost{tag}Total", f"{d.spent_usd.sum():.2f}")]21 rep = r[r.split == "confirmatory"].groupby(["task_id", "treatment"]).size()22 out.append(m("nRepeats", int(rep.max()) if len(rep) else 3))23 out.append(m("costAllTotal", f"{r.spent_usd.sum():.2f}")); out.append(m("nAllRuns", len(r)))24 except Exception as e:25 out.append(f"% sessions/runs not available: {e}")26 # headline results (confirmatory) + sensitivity + accounting27 try:28 at = pd.read_csv(os.path.join(EST, "confirmatory_arm_table.csv")).set_index("arm"); ct = pd.read_csv(os.path.join(EST, "confirmatory_contrasts.csv"))29 for a in ARMS:30 out += [m(f"p{a}", f"{at.loc[a, 'Y_task_weighted']:.3f}"), m(f"viol{a}", f"{at.loc[a, 'violation_rate']:.3f}"), m(f"nviol{a}", int(at.loc[a, 'n_runs_with_violation'])), m(f"cost{a}", f"{at.loc[a, 'cost_mean_usd']:.2f}"),31 m(f"costMed{a}", f"{at.loc[a, 'cost_median_usd']:.2f}"), m(f"wall{a}", f"{at.loc[a, 'wall_median_s']:.0f}"), m(f"qual{a}", f"{at.loc[a, 'quality_mean']:.3f}"), m(f"nRuns{a}", int(at.loc[a, 'n_runs'])), m(f"costPerSuccess{a}", f"{at.loc[a, 'cost_per_success_usd']:.2f}" if pd.notna(at.loc[a, 'cost_per_success_usd']) else "undefined")]32 for _, r in ct[ct.endpoint == "Y"].iterrows():33 k = r["contrast"].replace("-", "")34 out += [m(f"d{k}", f"{r['estimate']:+.3f}"), m(f"d{k}lo", f"{r['ci_low']:+.3f}"), m(f"d{k}hi", f"{r['ci_high']:+.3f}"), m(f"d{k}p", f"{r['perm_p']:.2f}"), m(f"d{k}fav", f"{int(r['n_tasks_favoring_first'])}/{int(r['n_tasks_tied'])}/{int(r['n_tasks_favoring_second'])}")]35 rv = pd.read_csv(os.path.join(EST, "confirmatory_contrasts.csv")); q = rv[(rv.endpoint == "quality_score") & (rv.contrast == "D-A")].iloc[0]36 out += [m("dDAqual", f"{q['estimate']:+.4f}"), m("dDAquallo", f"{q['ci_low']:+.4f}"), m("dDAqualhi", f"{q['ci_high']:+.4f}")]37 sa = pd.read_csv(os.path.join(EST, "sensitivity_rubric_arm_table.csv")).set_index("arm"); sc = pd.read_csv(os.path.join(EST, "sensitivity_rubric_contrasts.csv")); sd = sc[sc.contrast == "D-A"].iloc[0]38 for a in ARMS: out.append(m(f"sens{a}", f"{sa.loc[a, 'Y_task_weighted']:.3f}"))39 out += [m("sensDA", f"{sd['estimate']:+.3f}"), m("sensDAlo", f"{sd['ci_low']:+.3f}"), m("sensDAhi", f"{sd['ci_high']:+.3f}")]40 r = pd.read_csv(os.path.join(DATA, "runs.csv")); c = r[r.split == "confirmatory"]41 out += [m("nMainTimeouts", int((c.terminal_reason == "wall_clock_limit").sum())), m("nMainBudget", int(c.terminal_reason.fillna("").str.contains("budget").sum())), m("nMainInfraErrors", int((c.status == "error").sum()) + int(c.run_id.isin(["RC0094"]).sum() * 0)),42 m("costMainMean", f"{c.spent_usd.mean():.2f}"), m("nMainViolRuns", int((c.n_violations > 0).sum())), m("nMainSuccess", int((c.scientific_success == 1).sum()))]43 ses = pd.read_csv(os.path.join(DATA, "sessions.csv")); out.append(m("costAllSessions", f"{ses.session_cost_usd.sum():.2f}"))44 for a in ("B", "C", "D"): out.append(m(f"costRatio{a}", f"{at.loc[a, 'cost_mean_usd'] / at.loc['A', 'cost_mean_usd']:.1f}"))45 gr = pd.read_csv(os.path.join(DATA, "gate_receipts.csv")); g = gr[gr.run_id.str.startswith("RC")]46 out += [m("nGateControllerFail", int(((g.origin == "controller") & (g.passed == 0)).sum())), m("nContractFailA", int((g[(g.origin == "final_common_contract_check") & (g.passed == 0)].merge(c[["run_id", "treatment"]], on="run_id").treatment == "A").sum())), m("nContractFailB", int((g[(g.origin == "final_common_contract_check") & (g.passed == 0)].merge(c[["run_id", "treatment"]], on="run_id").treatment == "B").sum()))]47 rvw = pd.read_csv(os.path.join(DATA, "reviews.csv")); rvw = rvw[rvw.run_id.str.startswith("RC")]48 out += [m("nReviews", len(rvw)), m("nReviewsRevise", int((rvw.verdict == "revise").sum())), m("reviewIssuesMean", f"{rvw.n_issues.mean():.1f}")]49 except Exception as e:50 out.append(f"% results macros not available: {e}")51 os.makedirs(os.path.join(REPO, "paper", "tables"), exist_ok=True)52 open(os.path.join(REPO, "paper", "tables", "numbers.tex"), "w").write("\n".join(out) + "\n"); print("\n".join(out))53if __name__ == "__main__": main()54