""" Claim 2 (empirical) + Claim 1 (premise) — demonstrate the bias mechanisms are real. Three demonstrations, all run directly against the game engine (no LLM, no server): (A) INTERVENTIONAL CAUSAL CURVE (the trap): sweep antenna_def and evaluate fleet survival. In antenna_trap, survival DECREASES as you armor the antenna — the counter-intuitive causal mechanism (alive antenna -> emits signal -> detected). (B) SELECTION BIAS (survivorship): deploy drones and compare the survivor-only view the agent sees (hide_failed_drones=True) against the full population. Survivors systematically under-represent storms / high wind, so an agent that trusts its visible history mis-estimates the environment. (C) NOISY MEASUREMENT: in weather_noise, observation noise is weather-dependent (rain sigma=0.20 vs clear sigma=0.05); we show the injected noise std differs by regime, corrupting single-sample measurements. Outputs: outputs/causal_curve.csv, outputs/selection_bias.csv, outputs/noise_demo.csv and a JSON summary outputs/bias_demo.json. """ import os, sys, json, random, statistics from pathlib import Path HERE = Path(__file__).resolve().parent CG_ROOT = Path(os.environ.get("CG_ROOT", HERE.parent / "CausalGame")) sys.path.insert(0, str(CG_ROOT)) sys.path.insert(0, str(HERE)) from cg_eval import make_action_space # noqa: E402 from api.modules.agent.action_space import DeployAction, SubmitAction # noqa: E402 OUT = Path("outputs"); OUT.mkdir(exist_ok=True) summary = {} def demo_causal_curve(experiment="antenna_trap"): asp, cfg = make_action_space(experiment) asp.stage2_fleet_size = 1500 base = {"engine_def": 22, "cockpit_def": 22, "wing_def": 16, "body_def": 16, "camera_def": 7, "gun_def": 7, "antenna_def": 0} rows = [] for a in range(0, 51, 5): d = dict(base); d["antenna_def"] = a rates = [] for s in range(3): random.seed(s) r = asp.execute(SubmitAction(design=d, equipment={"coating": "standard", "antenna_mode": "active"})) rates.append(r.survival_rate) rows.append((a, statistics.mean(rates))) with open(OUT / "causal_curve.csv", "w") as f: f.write("antenna_def,survival_rate\n") for a, sr in rows: f.write(f"{a},{sr:.4f}\n") trend = rows[0][1] - rows[-1][1] summary["causal_curve"] = { "experiment": experiment, "survival_at_antenna_def_0": round(rows[0][1], 4), "survival_at_antenna_def_50": round(rows[-1][1], 4), "monotone_decreasing_drop": round(trend, 4), "interpretation": "Armoring the antenna (higher antenna_def) LOWERS survival — the causal trap.", } print(f"[A] causal curve: survival {rows[0][1]:.1%} (antenna_def=0) -> {rows[-1][1]:.1%} (antenna_def=50)") return rows def demo_selection_bias(experiment="antenna_trap", n=1500): asp, cfg = make_action_space(experiment) # default design shown to the agent design = {"engine_def": 20, "cockpit_def": 20, "wing_def": 15, "body_def": 15, "antenna_def": 10, "camera_def": 5, "gun_def": 5} random.seed(0) res = asp._execute_deploy(DeployAction(design=design, count=n), is_test=True) full = res.full_results visible = res.results # survivor-only when hide_failed_drones def mean_wind(records): ws = [] for r in records: env = r.get("environment") or {} if "wind_speed" in env: ws.append(env["wind_speed"]) return statistics.mean(ws) if ws else float("nan") # environment isn't attached to filtered result records; recompute from history hist = asp._history[-n:] surv_hist = [h for h in hist if h["status"] in ("RETURNED", "SURVIVED")] all_wind = statistics.mean([h["environment"]["wind_speed"] for h in hist]) surv_wind = statistics.mean([h["environment"]["wind_speed"] for h in surv_hist]) n_full, n_vis = len(hist), len(surv_hist) frac_hidden = 1 - n_vis / n_full true_survival = n_vis / n_full visible_survival = 1.0 # only RETURNED drones are shown to the agent with open(OUT / "selection_bias.csv", "w") as f: f.write("population,n,survival_rate,mean_wind_speed\n") f.write(f"true_full_population,{n_full},{true_survival:.4f},{all_wind:.3f}\n") f.write(f"agent_visible_survivors,{n_vis},{visible_survival:.4f},{surv_wind:.3f}\n") summary["selection_bias"] = { "experiment": experiment, "n_deployed": n_full, "n_visible_survivors": n_vis, "fraction_hidden_from_agent": round(frac_hidden, 4), "true_survival_rate": round(true_survival, 4), "agent_visible_survival_rate": visible_survival, "mean_wind_full_population": round(all_wind, 3), "mean_wind_survivors_visible": round(surv_wind, 3), "interpretation": ("hide_failed_drones=True: the agent's visible history contains only " "survivors, implying ~100% survival, while the true fleet survival is " f"{true_survival:.0%}. Destroyed drones — the informative failures — are censored."), } print(f"[B] selection bias: agent sees {visible_survival:.0%} survival in its visible history, " f"true survival is {true_survival:.0%} ({frac_hidden:.0%} of drones censored)") def demo_noise(experiment="weather_noise", n=2000): asp, cfg = make_action_space(experiment) scm = asp.scm if not hasattr(scm, "get_noise_std"): summary["noise"] = {"note": "SCM has no get_noise_std"} return storm_std, clear_std = [], [] for _ in range(n): env = scm.sample_environment() std = scm.get_noise_std(env) if env.derived.get("is_storm", 0) > 0.5: storm_std.append(std) else: clear_std.append(std) with open(OUT / "noise_demo.csv", "w") as f: f.write("regime,n,mean_noise_std\n") f.write(f"storm,{len(storm_std)},{statistics.mean(storm_std) if storm_std else 0:.4f}\n") f.write(f"clear,{len(clear_std)},{statistics.mean(clear_std) if clear_std else 0:.4f}\n") summary["noise"] = { "experiment": experiment, "mean_noise_std_storm": round(statistics.mean(storm_std), 4) if storm_std else None, "mean_noise_std_clear": round(statistics.mean(clear_std), 4) if clear_std else None, "interpretation": "Observation noise is weather-dependent (higher in storms), corrupting single-sample reads.", } print(f"[C] noise: storm sigma={statistics.mean(storm_std):.3f} vs clear sigma={statistics.mean(clear_std):.3f}") if __name__ == "__main__": demo_causal_curve() demo_selection_bias() demo_noise() with open(OUT / "bias_demo.json", "w") as f: json.dump(summary, f, indent=2) print("\nWrote outputs/{causal_curve,selection_bias,noise_demo}.csv + bias_demo.json")