File size: 7,098 Bytes
2b9a95b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 | """
Claim 1 — run LLM agents on CausalGame and measure whether they recover the causal
mechanism. Models are served via Hugging Face Inference Providers (provider 'hf' added
to run_agent.py / config/agent.json).
For each (model, scenario, mode, seed) we:
- (re)start the FastAPI backend bound to that scenario,
- run run_agent.py, capturing the machine-readable result JSON (survival, victory,
final design, reflection) written via REPRO_RESULT_JSON,
- derive a behavioral 'causal recovery' signal (did the final design adopt the
causally-correct lever for that family?),
- optionally LLM-judge the written report on a causal-reasoning rubric (CR proxy).
Outputs: outputs/llm_results.jsonl (one row per session) + outputs/llm_summary.csv.
"""
import os, sys, json, time, socket, subprocess, argparse, random
from pathlib import Path
HERE = Path(__file__).resolve().parent
CG_ROOT = Path(os.environ.get("CG_ROOT", HERE.parent / "CausalGame"))
OUT = Path("outputs"); OUT.mkdir(exist_ok=True)
# Behavioral 'causal recovery' rule per family:
# antenna family -> sacrifice antenna (antenna_def <= 5) => stealth mechanism found
# deployment family -> pick signal-filter/hardened enhancement AND shield => EMI comm mechanism
# weather_noise -> high antenna_def (>=25) => protect-antenna mechanism
def causal_recovery(scenario, design, equipment):
design = design or {}
equipment = equipment or {}
a = design.get("antenna_def", design.get("antenna", None))
if scenario.startswith("antenna_trap"):
return (a is not None) and (a <= 5)
if scenario.startswith("deployment_zone_trap"):
eqvals = " ".join(str(v).lower() for v in equipment.values())
used_filter = any(k in eqvals for k in ["signal_filter", "hardened", "adaptive", "aggressive", "thermal_shield"])
shield = design.get("shield_def", 0) or 0
return used_filter or shield >= 20
if scenario == "weather_noise":
return (a is not None) and (a >= 25)
return None
# run_agent.py connects to the base_url in config/server.json (localhost:8000),
# ignoring env overrides, so we always bind the backend to that fixed port and run
# sessions serially.
PORT = int(os.environ.get("CG_PORT", "8000"))
def kill_port(port):
subprocess.run(["bash", "-lc", f"pkill -f 'uvicorn api.app:app.*--port {port}' 2>/dev/null; "
f"lsof -ti tcp:{port} | xargs kill -9 2>/dev/null"], check=False)
time.sleep(1)
def start_backend(scenario, port):
kill_port(port)
env = dict(os.environ)
env["CAUSALGAME_EXPERIMENT"] = scenario
logf = open(OUT / f".server_{scenario}.log", "w")
proc = subprocess.Popen(
[sys.executable, "-m", "uvicorn", "api.app:app", "--port", str(port)],
cwd=str(CG_ROOT), env=env, stdout=logf, stderr=subprocess.STDOUT,
)
import urllib.request
for _ in range(60):
try:
urllib.request.urlopen(f"http://localhost:{port}/api/v2/mission_status", timeout=2).read()
return proc, logf
except Exception:
time.sleep(1)
proc.terminate()
raise RuntimeError(f"backend for {scenario} did not start")
def run_session(model, scenario, mode, port, seed, timeout=900):
# Absolute path: the run_agent.py subprocess runs with cwd=CG_ROOT, so a
# relative outputs/ path would resolve inside the repo and fail to write.
res_path = (OUT / f".res_{model}_{scenario}_{mode}_{seed}.json").resolve()
if res_path.exists():
res_path.unlink()
env = dict(os.environ)
env["REPRO_RESULT_JSON"] = str(res_path)
env["CAUSALGAME_EXPERIMENT"] = scenario
env["PYTHONHASHSEED"] = str(seed)
cmd = [sys.executable, "run_agent.py", "--model", model,
"--experiment", scenario, "--mode", mode]
logf = OUT / f".sess_{model}_{scenario}_{mode}_{seed}.log"
t0 = time.time()
try:
with open(logf, "w") as lf:
subprocess.run(cmd, cwd=str(CG_ROOT), env=env, stdout=lf,
stderr=subprocess.STDOUT, timeout=timeout)
except subprocess.TimeoutExpired:
pass
dur = time.time() - t0
row = {"model": model, "scenario": scenario, "mode": mode, "seed": seed,
"duration_s": round(dur, 1)}
if res_path.exists():
try:
data = json.load(open(res_path))
row.update({
"survival_rate": data.get("survival_rate"),
"victory": data.get("victory"),
"final_design": data.get("final_design"),
"equipment": (data.get("final_evaluation") or {}).get("equipment"),
"reflection": data.get("reflection"),
"tokens": data.get("tokens"),
"success": data.get("success"),
})
except Exception as e:
row["parse_error"] = str(e)
else:
row["error"] = "no result json"
row["causal_recovery"] = causal_recovery(scenario, row.get("final_design"), row.get("equipment"))
return row
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--models", nargs="+", required=True)
ap.add_argument("--scenarios", nargs="+", required=True)
ap.add_argument("--modes", nargs="+", default=["legacy"])
ap.add_argument("--repeats", type=int, default=1)
ap.add_argument("--timeout", type=int, default=900)
ap.add_argument("--out", default=str(OUT / "llm_results.jsonl"))
args = ap.parse_args()
results = []
outp = Path(args.out)
fout = open(outp, "a")
for scenario in args.scenarios:
port = PORT
proc, logf = start_backend(scenario, port)
print(f"\n### backend up for {scenario} on :{port}", flush=True)
try:
for model in args.models:
for mode in args.modes:
for seed in range(args.repeats):
print(f" -> {model} | {scenario} | {mode} | seed {seed}", flush=True)
row = run_session(model, scenario, mode, port, seed, timeout=args.timeout)
sr = row.get("survival_rate")
print(f" survival={sr} victory={row.get('victory')} "
f"causal_recovery={row.get('causal_recovery')} ({row.get('duration_s')}s)", flush=True)
results.append(row)
fout.write(json.dumps(row) + "\n"); fout.flush()
finally:
proc.terminate()
try:
proc.wait(timeout=10)
except Exception:
proc.kill()
logf.close()
fout.close()
# summary CSV
import csv
cols = ["model", "scenario", "mode", "seed", "survival_rate", "victory",
"causal_recovery", "duration_s"]
with open(OUT / "llm_summary.csv", "w", newline="") as f:
w = csv.writer(f); w.writerow(cols)
for r in results:
w.writerow([r.get(c) for c in cols])
print(f"\nWrote {outp} and outputs/llm_summary.csv ({len(results)} sessions)")
if __name__ == "__main__":
main()
|