File size: 10,757 Bytes
ad80dda 621100f ad80dda | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 | """OpenEnv spec compliance and pipeline health check.
Run:
python scripts/03_validate.py
python scripts/03_validate.py --fast
"""
from __future__ import annotations
import argparse
import json
import sys
import time
from pathlib import Path
from typing import Any
ROOT = Path(__file__).resolve().parent.parent
sys.path.insert(0, str(ROOT))
import yaml # noqa: E402
from agent.heuristics import WitnessHeuristics # noqa: F401,E402
from environment import WitnessStandEnv # noqa: E402
from models import Speaker, Turn, TurnType, WitnessAction # noqa: E402
from tasks.registry import TASK_REGISTRY # noqa: E402
from transcript.store import TranscriptStore # noqa: E402
EXPECTED = {
"basic": {"total_turns": 10, "data_lag_turns": 0},
"intermediate": {"total_turns": 20, "data_lag_turns": 0},
"advanced": {"total_turns": 30, "data_lag_turns": 2},
"expert": {"total_turns": 40, "data_lag_turns": 3},
}
RESULTS: list[dict[str, Any]] = []
def check(condition: bool, label: str, detail: str = "") -> None:
status = "PASS" if condition else "FAIL"
print(f" [{status}] {label}")
if detail and not condition:
print(f" -> {detail}")
RESULTS.append({"status": status, "label": label, "detail": detail})
def section(title: str) -> None:
print(f"\n-- {title} {'─' * max(0, 52 - len(title))}")
def dummy_action(text: str = "My position stands.", flagged: bool = False) -> WitnessAction:
return WitnessAction(
response_text=text,
flagged_distortion=flagged,
accepted_update=False,
tool_calls=[],
)
def check_registry() -> None:
section("1 Task registry")
check(len(TASK_REGISTRY) == 4, "4 tasks registered")
for name in EXPECTED:
check(name in TASK_REGISTRY, f"'{name}' in registry")
def check_task_configs() -> None:
section("2 Task configurations")
for name, expected in EXPECTED.items():
try:
task = TASK_REGISTRY[name]()
check(task.total_turns == expected["total_turns"], f"{name}: total_turns == {expected['total_turns']}")
check(task.data_lag_turns == expected["data_lag_turns"], f"{name}: data_lag_turns == {expected['data_lag_turns']}")
check(task.panel is not None, f"{name}: panel initialised")
check(task.persona is not None, f"{name}: persona loaded")
check(bool(getattr(task.persona, "system_prompt", "")), f"{name}: persona.system_prompt non-empty")
check(bool(task.domain), f"{name}: domain set")
except Exception as exc:
check(False, f"{name}: instantiation failed", str(exc))
def check_yaml() -> None:
section("3 openenv.yaml")
yaml_path = ROOT / "openenv.yaml"
check(yaml_path.exists(), "openenv.yaml exists")
if not yaml_path.exists():
return
try:
doc = yaml.safe_load(yaml_path.read_text())
check(isinstance(doc, dict), "openenv.yaml parses cleanly")
except Exception as exc:
check(False, "openenv.yaml parses cleanly", str(exc))
return
for key in ["name", "version", "tasks", "action_space", "observation_space", "reward_range"]:
check(key in doc, f"openenv.yaml has '{key}'")
declared = [item["name"] for item in doc.get("tasks", []) if "name" in item]
check(len(declared) == 4, "4 tasks declared in yaml")
for name in EXPECTED:
check(name in declared, f"yaml declares '{name}'")
rr = doc.get("reward_range")
check(isinstance(rr, list) and len(rr) == 2, "reward_range is [min, max]")
def check_env_interface(tasks: list[str] | None = None) -> None:
section("4 Environment interface (reset / step / grade)")
for task_name in tasks or list(EXPECTED):
try:
env = WitnessStandEnv()
obs = env.reset(task_name)
check(isinstance(obs, dict), f"{task_name}: reset() returns dict")
for key in ["questioner_text", "turn_number", "total_turns", "persona_system_prompt"]:
check(key in obs, f"{task_name}: obs has '{key}'")
done = False
scores: list[float] = []
while not done:
action = dummy_action(flagged=len(scores) % 3 == 1)
obs, reward, done, _ = env.step(action)
scores.append(float(reward))
check(isinstance(reward, (int, float)), f"{task_name} t{len(scores)}: reward numeric")
check(len(scores) == EXPECTED[task_name]["total_turns"], f"{task_name}: correct turn count")
final = env.grade()
check(isinstance(final, float), f"{task_name}: grade() returns float")
check(0.0 <= final <= 1.0, f"{task_name}: grade() in [0, 1]")
except Exception as exc:
check(False, f"{task_name}: episode loop crashed", str(exc))
def check_reward_bounds() -> None:
section("5 Reward bounds")
for task_name in ["basic", "expert"]:
try:
env = WitnessStandEnv()
env.reset(task_name)
done = False
bad = []
i = 0
while not done:
_, reward, done, _ = env.step(dummy_action(flagged=i % 2 == 0))
if not (-1.0 <= reward <= 2.0):
bad.append((i, reward))
i += 1
check(not bad, f"{task_name}: all per-turn scores in [-1, 2]", str(bad))
final = env.grade()
check(0.0 <= final <= 1.0, f"{task_name}: final grade in [0, 1]", str(final))
except Exception as exc:
check(False, f"{task_name}: reward bounds check crashed", str(exc))
def check_grader_determinism() -> None:
section("6 Grader determinism")
scores = []
responses = [
"My position stands. The evidence supports my assessment.",
"That is not accurate. At turn 0, my statement included caveats.",
"The methodology was sound and the conclusion remains qualified.",
]
for _ in range(2):
env = WitnessStandEnv()
env.reset("basic", seed=0)
done = False
i = 0
while not done:
text = responses[i % len(responses)]
action = dummy_action(text=text, flagged="not accurate" in text.lower())
_, _, done, _ = env.step(action)
i += 1
scores.append(round(env.grade(), 6))
check(scores[0] == scores[1], "grader is deterministic", f"{scores}")
def check_transcript_lag() -> None:
section("7 Transcript data lag")
for lag, n_turns, expected in [(0, 5, 5), (2, 5, 3), (3, 5, 2)]:
store = TranscriptStore(data_lag_turns=lag)
for i in range(n_turns):
store.append(Turn(turn_no=i, speaker=Speaker.WITNESS, text=f"turn {i}", turn_type=TurnType.NEUTRAL))
check(len(store.get_witness_turns()) == expected, f"lag={lag}: {expected}/{n_turns} turns visible")
def check_panel_schedule() -> None:
section("8 Questioner panel schedule")
injection_types = {TurnType.DISTORTION_INJECTION, TurnType.AUTHORITY_INJECTION, TurnType.TEMPORAL_INJECTION, TurnType.EXHAUSTION}
for task_name in ["basic", "expert"]:
task = TASK_REGISTRY[task_name]()
injections = [t for t in range(task.total_turns) if task.panel.get_turn_type(t) in injection_types]
neutrals = [t for t in range(task.total_turns) if task.panel.get_turn_type(t) == TurnType.NEUTRAL]
check(bool(injections), f"{task_name}: has injection turns")
check(bool(neutrals), f"{task_name}: has neutral turns")
if task_name == "expert":
types_seen = {task.panel.get_turn_type(t) for t in injections}
check(len(types_seen) >= 2, "expert: multiple injection types in schedule")
def check_agent_imports() -> None:
section("9 Agent module imports")
modules = [
("agent.memory", "EpisodicMemory"),
("agent.prompt", "build_system_prompt"),
("agent.prompt", "build_user_prompt"),
("agent.parser", "parse_action"),
("agent.heuristics", "WitnessHeuristics"),
]
for module, attr in modules:
try:
mod = __import__(module, fromlist=[attr])
check(hasattr(mod, attr), f"{module}.{attr} importable")
except Exception as exc:
check(False, f"{module} importable", str(exc))
def check_logs() -> None:
section("10 Logs directory + JSON outputs")
logs_dir = ROOT / "logs"
# Auto-create logs/ if missing — first run always passes instead of failing
if not logs_dir.exists():
logs_dir.mkdir(parents=True, exist_ok=True)
check(True, "logs/ directory exists")
for filename, keys, command in [
("benchmark_results.json", ["tasks", "avg_score", "witness_elo"], "09_run_eval.py"),
("demo_transcript.json", ["before", "after", "delta"], "10_demo.py"),
]:
path = logs_dir / filename
if not path.exists():
print(f" [SKIP] {filename} not generated yet (run {command})")
continue
try:
data = json.loads(path.read_text())
for key in keys:
check(key in data, f"{filename} has '{key}'")
except Exception as exc:
check(False, f"{filename} parses cleanly", str(exc))
def write_report(start: float) -> bool:
passed = sum(1 for r in RESULTS if r["status"] == "PASS")
failed = sum(1 for r in RESULTS if r["status"] == "FAIL")
total = passed + failed
print(f"\n{'=' * 56}")
print(f" {passed}/{total} checks passed", end="")
if failed:
print(f" | {failed} FAILED")
for item in RESULTS:
if item["status"] == "FAIL":
print(f" x {item['label']}")
else:
print(" | ALL PASSED")
print(f" elapsed: {time.time() - start:.2f}s")
print(f"{'=' * 56}\n")
out_dir = ROOT / "logs" / "health"
out_dir.mkdir(parents=True, exist_ok=True)
(out_dir / "validate.json").write_text(json.dumps({"results": RESULTS, "passed": failed == 0}, indent=2))
return failed == 0
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--fast", action="store_true", help="Run episode loop on basic task only")
args = parser.parse_args()
start = time.time()
print("\n=== The Witness Stand — OpenEnv Validation ===")
check_registry()
check_task_configs()
check_yaml()
check_env_interface(tasks=["basic"] if args.fast else None)
if not args.fast:
check_reward_bounds()
check_grader_determinism()
check_transcript_lag()
check_panel_schedule()
check_agent_imports()
check_logs()
raise SystemExit(0 if write_report(start) else 1)
if __name__ == "__main__":
main()
|