Spaces:
Sleeping
Sleeping
File size: 15,318 Bytes
1b42f19 6a32325 1b42f19 6a32325 1b42f19 6a32325 874e22a 1b42f19 6a32325 1b42f19 71fa486 1b42f19 6a32325 1b42f19 6a32325 1b42f19 ba3f47d 6a32325 ba3f47d 6a32325 1b42f19 6a32325 1b42f19 ba3f47d 306d70e 1b42f19 6a32325 1b42f19 6a32325 1b42f19 6a32325 1b42f19 6a32325 1b42f19 6a32325 1b42f19 6a32325 1b42f19 6a32325 306d70e 6a32325 306d70e 6a32325 1b42f19 6a32325 ba3f47d 6a32325 ba3f47d 6a32325 ba3f47d 6a32325 ba3f47d 6a32325 1b42f19 6a32325 1b42f19 6a32325 306d70e 6a32325 1b42f19 6a32325 1b42f19 6a32325 1b42f19 71fa486 1b42f19 6a32325 1b42f19 6a32325 1b42f19 6a32325 71fa486 6a32325 71fa486 6a32325 71fa486 6a32325 71fa486 6a32325 306d70e 6a32325 71fa486 6a32325 71fa486 306d70e 71fa486 6a32325 | 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 | """
StateReconciler — Dynamic Golden Database Grading Engine.
ARCHITECTURE:
- Instead of hardcoded expected values, we build a "golden" database by running
the correct migration on a fresh copy of the seed data.
- The agent's database is compared table-by-table against this golden reference.
- This makes the grader SEED-INDEPENDENT: if judges change the seed data,
the golden DB auto-updates and scoring remains accurate.
SCORING WEIGHTS (per-table, dynamic):
- Schema match (table exists, correct columns): 30%
- Data match (row count + content): 40%
- FK & constraint integrity: 20%
- Anti-exploit checks: 10%
ANTI-EXPLOIT PROTECTIONS:
- Case-insensitive table/column name comparison
- PRAGMA state preservation (grader doesn't corrupt agent's FK state)
- Phantom row detection (SUM fingerprinting)
- Empty table exploitation blocked
- Extra/leftover table penalty
"""
import sqlite3
from typing import Any, Dict, List, Optional, Set, Tuple
# Import seeds for golden migration functions
try:
import seeds
except ImportError:
from .. import seeds
def _get_table_names(conn: sqlite3.Connection) -> Set[str]:
"""Get all user table names (case-normalized to lowercase)."""
try:
cursor = conn.execute(
"SELECT name FROM sqlite_master WHERE type='table' "
"AND name NOT LIKE 'sqlite_%' ORDER BY name"
)
return {row[0].lower() for row in cursor.fetchall()}
except Exception:
return set()
def _get_column_info(conn: sqlite3.Connection, table: str) -> List[dict]:
"""Get column info for a table. Returns list of {name, type, notnull, pk}."""
try:
cursor = conn.execute(f"PRAGMA table_info([{table}])")
return [
{
"name": row[1].lower(),
"type": row[2].upper(),
"notnull": row[3],
"dflt_value": row[4], # Added check for default values
"pk": row[5]
}
for row in cursor.fetchall()
]
except Exception:
return []
def _get_column_names(conn: sqlite3.Connection, table: str) -> Set[str]:
"""Get column names (lowercase) for a table."""
return {col["name"] for col in _get_column_info(conn, table)}
def _get_column_signatures(conn: sqlite3.Connection, table: str) -> Set[Tuple[str, str, int, Any]]:
"""Get (name, type, notnull, dflt_value) tuples for absolute schema grading."""
return {
(col["name"], col["type"], col["notnull"], str(col["dflt_value"]))
for col in _get_column_info(conn, table)
}
def _get_row_count(conn: sqlite3.Connection, table: str) -> int:
"""Get row count. Returns 0 on error."""
try:
cursor = conn.execute(f"SELECT COUNT(*) FROM [{table}]")
return cursor.fetchone()[0]
except Exception:
return 0
def _get_all_rows(conn: sqlite3.Connection, table: str) -> List[Tuple]:
"""Get all rows from a table, sorted for deterministic comparison."""
try:
cols = _get_column_names(conn, table)
if not cols:
return []
cursor = conn.execute(f"SELECT * FROM [{table}] ORDER BY 1")
return cursor.fetchall()
except Exception:
return []
def _has_foreign_key(conn: sqlite3.Connection, table: str, ref_table: str) -> bool:
"""Check if table has a FK referencing ref_table (case-insensitive)."""
try:
cursor = conn.execute(f"PRAGMA foreign_key_list([{table}])")
for row in cursor.fetchall():
if row[2].lower() == ref_table.lower():
return True
return False
except Exception:
return False
def _count_foreign_keys(conn: sqlite3.Connection, table: str) -> int:
"""Count unique FK constraints for a table using the FK id."""
try:
cursor = conn.execute(f"PRAGMA foreign_key_list([{table}])")
refs = set()
for row in cursor.fetchall():
refs.add(row[0]) # row[0] is the sequential ID of the foreign key constraint
return len(refs)
except Exception:
return 0
def _build_golden_db(task_name: str) -> sqlite3.Connection:
"""
Build a golden reference database for a task.
Seeds a fresh in-memory DB with the task's seed data, then applies
the golden migration to produce the expected final state.
"""
task_config = seeds.TASKS[task_name]
conn = sqlite3.connect(":memory:")
conn.execute("PRAGMA foreign_keys = ON")
# Seed with same data as agent
task_config["seed_fn"](conn)
# Apply perfect migration
task_config["golden_fn"](conn)
return conn
def _compare_row_data(
agent_rows: List[Tuple],
golden_rows: List[Tuple],
) -> float:
"""
Compare row data between agent and golden databases.
Returns a similarity score between 0.0 and 1.0.
Handles: different row counts, partial matches, type coercion differences.
"""
if not golden_rows:
return 1.0 if not agent_rows else 0.0
if not agent_rows:
return 0.0
# Exact match
if agent_rows == golden_rows:
return 1.0
# Row count match bonus
count_match = 1.0 if len(agent_rows) == len(golden_rows) else (
min(len(agent_rows), len(golden_rows)) / max(len(agent_rows), len(golden_rows))
)
# Per-row comparison (order-independent for flexibility)
golden_set = set()
for row in golden_rows:
# Normalize: try numerical comparison first, fallback to string
normalized_golden = []
for v in row:
if v is None: normalized_golden.append("")
else:
try:
normalized_golden.append(float(v))
except (ValueError, TypeError):
normalized_golden.append(str(v).strip())
golden_set.add(tuple(normalized_golden))
matched = 0
for row in agent_rows:
# Normalize: try numerical comparison first, fallback to string
normalized_agent = []
for v in row:
if v is None: normalized_agent.append("")
else:
try: # If it looks like a number, treat as a float for comparison
normalized_agent.append(float(v))
except (ValueError, TypeError):
normalized_agent.append(str(v).strip())
normalized_agent = tuple(normalized_agent)
# Look for match in golden set with numerical tolerance
found = False
for g_row in list(golden_set):
# Compare rows with float tolerance where applicable
match = True
if len(normalized_agent) != len(g_row): continue
for a_val, g_val in zip(normalized_agent, g_row):
if isinstance(a_val, float) and isinstance(g_val, float):
if abs(a_val - g_val) > 1e-7:
match = False; break
elif str(a_val) != str(g_val):
match = False; break
if match:
matched += 1
golden_set.discard(g_row)
found = True
break
if len(golden_rows) == 0:
content_match = 1.0 # Logic fix: if golden has no rows, agent with 0 rows is 1.0
else:
content_match = matched / len(golden_rows)
# Penalize extra rows (data bloat)
if len(agent_rows) > len(golden_rows):
# 1% penalty per extra row to prevent "guessing" by dumping all possible data
bloat_penalty = max(0.5, 1.0 - (len(agent_rows) - len(golden_rows)) * 0.01)
content_match *= bloat_penalty
return 0.4 * count_match + 0.6 * content_match
class StateReconciler:
"""
Dynamic Golden Database grading engine.
Compares the agent's database state against a dynamically-generated
golden reference database. No hardcoded expected values.
"""
def __init__(self, task_name: str):
self.task_name = task_name
self._last_score: float = 0.0
self._golden_conn: Optional[sqlite3.Connection] = None
# Build golden reference DB
try:
self._golden_conn = _build_golden_db(task_name)
self._golden_tables = _get_table_names(self._golden_conn)
self._golden_table_data: Dict[str, dict] = {}
for table in self._golden_tables:
self._golden_table_data[table] = {
"columns": _get_column_info(self._golden_conn, table),
"col_names": _get_column_names(self._golden_conn, table),
"col_signatures": _get_column_signatures(self._golden_conn, table),
"rows": _get_all_rows(self._golden_conn, table),
"row_count": _get_row_count(self._golden_conn, table),
"fk_count": _count_foreign_keys(self._golden_conn, table),
}
except Exception:
self._golden_tables = set()
self._golden_table_data = {}
def __del__(self):
"""Clean up golden DB connection."""
if self._golden_conn is not None:
try:
self._golden_conn.close()
except Exception:
pass
def score(self, conn: sqlite3.Connection) -> float:
"""
Compute migration score by comparing agent DB against golden reference.
Scoring breakdown:
- Schema match: 0.30 (tables exist with correct columns)
- Data match: 0.40 (row content matches golden DB)
- FK/constraint integrity: 0.20 (FKs enforced, integrity OK)
- Anti-exploit bonus: 0.10 (no empty tables, no extra tables)
Returns: float in [0.01, 0.99]
"""
try:
return self._score_dynamic(conn)
except Exception:
return 0.01
def compute_step_reward(self, conn: sqlite3.Connection) -> Tuple[float, float]:
"""
Compute current score and step reward delta.
CRITICAL: Preserves the agent's PRAGMA foreign_keys state.
The grader reads FK state, does its work, then restores it.
"""
# A8: Preserve PRAGMA state
try:
original_fk = conn.execute("PRAGMA foreign_keys").fetchone()[0]
except Exception:
original_fk = 1
current_score = self.score(conn)
step_reward = current_score - self._last_score
self._last_score = current_score
# A8: Restore original PRAGMA state
try:
conn.execute(f"PRAGMA foreign_keys = {'ON' if original_fk else 'OFF'}")
except Exception:
pass
return current_score, step_reward
def _score_dynamic(self, conn: sqlite3.Connection) -> float:
"""Core dynamic scoring: compare agent DB against golden DB."""
if not self._golden_tables:
return 0.01
agent_tables = _get_table_names(conn)
# ---- 1. Schema Match (0.30) ----
schema_score = 0.0
tables_found = 0
total_col_match = 0.0
for table in self._golden_tables:
golden_info = self._golden_table_data[table]
if table in agent_tables:
tables_found += 1
# Signature (name + type) comparison
agent_cols = _get_column_signatures(conn, table)
golden_cols = golden_info["col_signatures"]
if golden_cols:
col_overlap = len(agent_cols & golden_cols) / len(golden_cols)
total_col_match += col_overlap
else:
total_col_match += 1.0
if self._golden_tables:
table_ratio = tables_found / len(self._golden_tables)
col_ratio = total_col_match / len(self._golden_tables) if self._golden_tables else 0
schema_score = 0.15 * table_ratio + 0.15 * col_ratio
# ---- 2. Data Match (0.40) ----
data_score = 0.0
data_checks = 0
for table in self._golden_tables:
golden_info = self._golden_table_data[table]
if table not in agent_tables:
data_checks += 1
continue
agent_rows = _get_all_rows(conn, table)
golden_rows = golden_info["rows"]
similarity = _compare_row_data(agent_rows, golden_rows)
data_score += similarity
data_checks += 1
if data_checks > 0:
data_score = 0.40 * (data_score / data_checks)
# ---- 3. FK & Constraint Integrity (0.20) ----
fk_score = 0.0
fk_checks = 0
for table in self._golden_tables:
golden_info = self._golden_table_data[table]
expected_fks = golden_info["fk_count"]
if expected_fks > 0 and table in agent_tables:
agent_fks = _count_foreign_keys(conn, table)
fk_ratio = min(agent_fks, expected_fks) / expected_fks
fk_score += fk_ratio
fk_checks += 1
# PRAGMA integrity check
integrity_ok = False
try:
# Temporarily enable FK for integrity check
conn.execute("PRAGMA foreign_keys = ON")
cursor = conn.execute("PRAGMA integrity_check")
result = cursor.fetchone()[0]
# Explicitly run foreign_key_check to catch orphaned rows
fk_cursor = conn.execute("PRAGMA foreign_key_check")
fk_violations = fk_cursor.fetchall()
integrity_ok = (result == "ok" and len(fk_violations) == 0)
except Exception:
pass
if fk_checks > 0:
fk_score = 0.10 * (fk_score / fk_checks)
else:
# No FK constraints expected — award full FK portion
fk_score = 0.10
fk_score += 0.10 if integrity_ok else 0.0
# ---- 4. Anti-Exploit Checks (0.10) ----
exploit_score = 0.10 # Start with full credit, deduct for violations
# Check for empty tables where golden has data
for table in self._golden_tables:
golden_info = self._golden_table_data[table]
if golden_info["row_count"] > 0 and table in agent_tables:
agent_count = _get_row_count(conn, table)
if agent_count == 0:
# Agent emptied a table that should have data — heavy penalty
exploit_score = 0.0
# Also cap the data score for this exploit
data_score = min(data_score, 0.05)
break
# Penalize extra non-golden tables (schema pollution)
extra_tables = agent_tables - self._golden_tables
if extra_tables:
# Small penalty per extra table (some might be temp tables)
penalty = min(0.05, 0.01 * len(extra_tables))
exploit_score = max(0, exploit_score - penalty)
total = schema_score + data_score + fk_score + exploit_score
return max(0.01, min(0.99, total))
|