File size: 16,263 Bytes
b759ee5 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 | """
Quilltale — Automated Evaluation Runner
Plays through a fixed scenario and measures three metrics:
1. invalid_transition_rate — how often the GM proposes invalid state changes
2. memory_utilisation_rate — how often NPC memories actually shape narration
3. factual_consistency_rate — how often narration contradicts world state
Run with:
python eval_runner.py
Outputs:
eval_results/report_<timestamp>.json
eval_results/report_<timestamp>.txt
"""
import json
import os
import logging
from datetime import datetime
from pathlib import Path
from dataclasses import dataclass, field
from src.world.state import WorldState
from src.agents.game_master import GameMasterAgent
from src.llm import get_llm
from src.llm.base import BaseLLM
logging.basicConfig(level=logging.WARNING)
### Evaluation scenario ########################################
### A fixed sequence of actions designed to stress-test all three metrics.
### Covers: movement, item interaction, NPC conversation,
### NPC memory escalation, invalid attempts, multi-step routing.
EVAL_SCENARIO = [
### Turn 1-3: basic movement and scene establishment
"look around the tavern carefully",
"examine the wanted notice on the wall",
"examine the rusty dagger on the table",
### Turn 4-6: NPC interaction to help build Marta's memory
"talk to Marta the barkeep",
"ask Marta about the dagger",
"ask Marta about the chest upstairs",
### Turn 7: memory escalation
"threaten Marta to tell you what she knows",
### Turn 8-9: item interaction
"pick up the dagger",
"pick up the wanted notice",
### Turn 10: invalid attempt to pick an item not here
"pick up the strange coin",
### Turn 11-12: movement
"go north to the street",
"go east to the market",
### Turn 13-14: NPC interaction in new location
"talk to Aldric the merchant",
"ask Aldric about the strange coin",
### Turn 15: multi-step routing back
"go back to the tavern",
### Turn 16: verify Marta still remembers the threat
"talk to Marta",
### Turn 17-18: upstairs exploration
"go upstairs to my room",
"examine the locked chest",
### Turn 19: use key on chest
"use the old iron key on the chest",
### Turn 20: invalid movement attempt
"go south",
]
### LLM Judge ########################################
JUDGE_SYSTEM = """
You are an objective evaluator assessing AI game master output quality.
You always respond with valid JSON only. No preamble or explanation outside the JSON.
"""
def judge_memory_utilisation(
llm: BaseLLM,
narration: str,
npc_memories: list,
npc_name: str,
) -> dict:
"""
Ask the LLM judge: does this narration reflect the NPC's recorded memories?
Returns {"reflects_memory": bool, "confidence": float, "reason": str}
"""
if not npc_memories:
return {"reflects_memory": True, "confidence": 1.0, "reason": "No memories to reflect."}
memory_text = "\n".join(
f" - Turn {m.turn} ({m.emotional_tone}, sig={m.significance}): {m.description}"
for m in npc_memories[:3]
)
prompt = f"""
{npc_name} has these recorded memories of the player:
{memory_text}
The game master produced this narration:
"{narration}"
Does the narration reflect any of these memories through {npc_name}'s behaviour,
tone, dialogue, or reaction? Even subtle reflection counts (e.g. wariness, coldness,
gratitude shown through action rather than stated directly).
Respond with JSON:
{{"reflects_memory": true/false, "confidence": 0.0-1.0, "reason": "one sentence"}}
"""
try:
raw = llm.generate_json(prompt, JUDGE_SYSTEM)
return json.loads(raw)
except Exception:
return {"reflects_memory": False, "confidence": 0.0, "reason": "Judge call failed."}
def judge_factual_consistency(
llm: BaseLLM,
narration: str,
world_context: str,
) -> dict:
"""
Ask the LLM judge: does the narration contradict the world state?
Returns {"is_consistent": bool, "confidence": float, "violation": str}
"""
prompt = f"""
The current world state contains these facts:
{world_context}
The game master produced this narration:
"{narration}"
Does the narration contradict any recorded facts? Look for:
- Items mentioned that aren't in the current location or inventory
- NPCs described as present when they are not listed
- Movement described to locations not reachable from current exits
- Health or inventory states that differ from recorded values
Respond with JSON:
{{"is_consistent": true/false, "confidence": 0.0-1.0, "violation": "describe any contradiction, or empty string if none"}}
"""
try:
raw = llm.generate_json(prompt, JUDGE_SYSTEM)
return json.loads(raw)
except Exception:
return {"is_consistent": True, "confidence": 0.0, "violation": "Judge call failed."}
### Metrics collector ########################################
@dataclass
class TurnRecord:
turn: int
action: str
narration: str
state_update: dict
changes_applied: list[str]
rejected_transitions: list[str]
memory_judgement: dict = field(default_factory=dict)
consistency_judgement: dict = field(default_factory=dict)
npcs_present: list[str] = field(default_factory=list)
npc_memories_present: bool = False
@dataclass
class EvalReport:
scenario_name: str = "default_world"
total_turns: int = 0
timestamp: str = ""
### Raw counts
total_transitions_attempted: int = 0
total_transitions_rejected: int = 0
total_turns_with_npcs: int = 0
total_turns_memory_reflected: int = 0
total_turns_memory_judged: int = 0
total_turns_consistent: int = 0
total_turns_consistency_judged: int = 0
### Derived metrics
invalid_transition_rate: float = 0.0
memory_utilisation_rate: float = 0.0
factual_consistency_rate: float = 0.0
### Turn-by-turn records
turns: list[TurnRecord] = field(default_factory=list)
### Failure examples
rejection_examples: list[str] = field(default_factory=list)
memory_failures: list[dict] = field(default_factory=list)
consistency_violations: list[dict] = field(default_factory=list)
def compute_rates(self):
if self.total_transitions_attempted > 0:
self.invalid_transition_rate = round(
self.total_transitions_rejected / self.total_transitions_attempted, 3
)
if self.total_turns_memory_judged > 0:
self.memory_utilisation_rate = round(
self.total_turns_memory_reflected / self.total_turns_memory_judged, 3
)
if self.total_turns_consistency_judged > 0:
self.factual_consistency_rate = round(
self.total_turns_consistent / self.total_turns_consistency_judged, 3
)
### Evaluation runner ########################################
def run_evaluation(
scenario: list[str] = EVAL_SCENARIO,
world_path: str = "data/worlds/default.json",
llm_name: str = "gemini",
run_judge: bool = True,
) -> EvalReport:
"""
Play through the scenario automatically and collect metrics.
Args:
scenario: List of player actions to execute in order.
world_path: Path to the world JSON file.
llm_name: LLM provider to use for the GM.
run_judge: Whether to run LLM judge calls for memory and consistency.
Set False to only measure invalid transition rate (cheaper).
"""
print(f"Running evaluation: {len(scenario)} turns, judge={'on' if run_judge else 'off'}")
print("-" * 60)
with open(world_path) as f:
state = WorldState.from_json(f.read())
llm = get_llm(llm_name)
judge_llm = get_llm(llm_name) if run_judge else None
gm = GameMasterAgent(llm)
report = EvalReport(
scenario_name=Path(world_path).stem,
timestamp=datetime.now().isoformat(),
)
### Generate opening (not counted in metrics — no action to evaluate)
opening = gm.generate_opening(state)
print(f"Opening: {opening['narration'][:80]}...")
print()
for i, action in enumerate(scenario):
print(f"Turn {i+1:02d}: {action}")
world_context_before = state.to_context_summary()
npcs_present = state.npcs_in_location(state.player.location)
npcs_with_memories = [
npc for npc in npcs_present
if npc.alive and len(npc.memories) > 0
]
result = gm.process_turn(action, state)
### Count transitions
update = result.get("state_update", {})
transition_keys = {"move_player", "pickup_item", "drop_item", "npc_state"}
attempted = sum(1 for k in transition_keys if k in update)
rejected = [c for c in result["changes_applied"] if "REJECTED" in c]
report.total_transitions_attempted += attempted
report.total_transitions_rejected += len(rejected)
if rejected:
report.rejection_examples.extend(rejected[:2])
### Memory utilisation judgement
memory_judgement = {}
if run_judge and npcs_with_memories:
report.total_turns_memory_judged += 1
### Judge against the NPC with most memories
primary_npc = max(npcs_with_memories, key=lambda n: len(n.memories))
memory_judgement = judge_memory_utilisation(
judge_llm,
result["narration"],
primary_npc.relevant_memories(),
primary_npc.name,
)
if memory_judgement.get("reflects_memory", False):
report.total_turns_memory_reflected += 1
else:
report.memory_failures.append({
"turn": i + 1,
"action": action,
"narration": result["narration"],
"reason": memory_judgement.get("reason", ""),
})
### Factual consistency judgement
consistency_judgement = {}
if run_judge:
report.total_turns_consistency_judged += 1
consistency_judgement = judge_factual_consistency(
judge_llm,
result["narration"],
world_context_before,
)
if consistency_judgement.get("is_consistent", True):
report.total_turns_consistent += 1
else:
report.consistency_violations.append({
"turn": i + 1,
"action": action,
"narration": result["narration"],
"violation": consistency_judgement.get("violation", ""),
})
### Record turn
record = TurnRecord(
turn=i + 1,
action=action,
narration=result["narration"],
state_update=update,
changes_applied=result["changes_applied"],
rejected_transitions=rejected,
memory_judgement=memory_judgement,
consistency_judgement=consistency_judgement,
npcs_present=[n.name for n in npcs_present],
npc_memories_present=bool(npcs_with_memories),
)
report.turns.append(record)
report.total_turns += 1
print(f" → {result['narration'][:80]}...")
if rejected:
print(f" ✗ REJECTED: {rejected}")
print()
report.compute_rates()
return report
### Report writer ########################################
def write_report(report: EvalReport, output_dir: str = "eval_results"):
Path(output_dir).mkdir(exist_ok=True)
timestamp = report.timestamp.replace(":", "-").replace(".", "-")
### Full JSON data report
json_path = Path(output_dir) / f"report_{timestamp}.json"
with open(json_path, "w", encoding="utf-8") as f:
json.dump(report.__dict__, f, indent=2, default=lambda o: o.__dict__)
### Report Summary
txt_path = Path(output_dir) / f"report_{timestamp}.txt"
with open(txt_path, "w", encoding="utf-8") as f:
f.write("QUILLTALE — EVALUATION REPORT\n")
f.write("=" * 60 + "\n\n")
f.write(f"Scenario: {report.scenario_name}\n")
f.write(f"Timestamp: {report.timestamp}\n")
f.write(f"Turns: {report.total_turns}\n\n")
f.write("METRICS\n")
f.write("-" * 40 + "\n")
f.write(
f"Invalid transition rate: {report.invalid_transition_rate:.1%} "
f"({report.total_transitions_rejected}/{report.total_transitions_attempted} transitions rejected)\n"
)
f.write(
f"Memory utilisation rate: {report.memory_utilisation_rate:.1%} "
f"({report.total_turns_memory_reflected}/{report.total_turns_memory_judged} turns with NPCs reflected memory)\n"
)
f.write(
f"Factual consistency rate: {report.factual_consistency_rate:.1%} "
f"({report.total_turns_consistent}/{report.total_turns_consistency_judged} turns without contradictions)\n\n"
)
if report.rejection_examples:
f.write("REJECTED TRANSITIONS (sample)\n")
f.write("-" * 40 + "\n")
for ex in report.rejection_examples[:5]:
f.write(f" {ex}\n")
f.write("\n")
if report.memory_failures:
f.write(f"MEMORY UTILISATION FAILURES ({len(report.memory_failures)} turns)\n")
f.write("-" * 40 + "\n")
for mf in report.memory_failures[:3]:
f.write(f" Turn {mf['turn']}: {mf['action']}\n")
f.write(f" Narration: {mf['narration'][:100]}...\n")
f.write(f" Reason: {mf['reason']}\n\n")
if report.consistency_violations:
f.write(f"CONSISTENCY VIOLATIONS ({len(report.consistency_violations)} turns)\n")
f.write("-" * 40 + "\n")
for cv in report.consistency_violations[:3]:
f.write(f" Turn {cv['turn']}: {cv['action']}\n")
f.write(f" Narration: {cv['narration'][:100]}...\n")
f.write(f" Violation: {cv['violation']}\n\n")
f.write("TURN-BY-TURN SUMMARY\n")
f.write("-" * 40 + "\n")
for t in report.turns:
status = "✓" if not t.rejected_transitions else "✗"
mem = ""
if t.memory_judgement:
mem = " [mem:✓]" if t.memory_judgement.get("reflects_memory") else " [mem:✗]"
con = ""
if t.consistency_judgement:
con = " [con:✓]" if t.consistency_judgement.get("is_consistent") else " [con:✗]"
f.write(f" {status} T{t.turn:02d} {t.action[:45]:<45}{mem}{con}\n")
print(f"Report written to:")
print(f" {json_path}")
print(f" {txt_path}")
return json_path, txt_path
### Entry point ########################################
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser(description="Run Quilltale evaluation")
parser.add_argument("--no-judge", action="store_true",
help="Skip LLM judge calls (only measure transition rate)")
parser.add_argument("--llm", default="gemini",
help="LLM provider: gemini or claude")
parser.add_argument("--world", default="data/worlds/default.json",
help="Path to world JSON file")
args = parser.parse_args()
report = run_evaluation(
scenario=EVAL_SCENARIO,
world_path=args.world,
llm_name=args.llm,
run_judge=not args.no_judge,
)
write_report(report)
print("\n" + "=" * 60)
print("SUMMARY")
print("=" * 60)
print(f"Invalid transition rate: {report.invalid_transition_rate:.1%}")
print(f"Memory utilisation rate: {report.memory_utilisation_rate:.1%}")
print(f"Factual consistency rate: {report.factual_consistency_rate:.1%}") |