Spaces:
Running
Running
File size: 26,699 Bytes
4052d84 d76d092 4052d84 d8f8a45 4052d84 b00feb0 4052d84 b00feb0 4052d84 b00feb0 4052d84 a1b1513 ce6728e 898bc18 bf8f1ff ce6728e 4052d84 b00feb0 4052d84 b00feb0 4052d84 33279ea 4052d84 b00feb0 4052d84 2bc545f 4855450 2bc545f 4855450 4052d84 b00feb0 4052d84 bf8f1ff 4052d84 898bc18 2bc545f d76d092 4052d84 2bc545f 6218d9a ca62faa 4052d84 d76d092 4052d84 898bc18 d76d092 2bc545f 898bc18 2bc545f d76d092 4052d84 6218d9a 4052d84 b00feb0 4052d84 d8f8a45 4052d84 a085ad1 4052d84 a085ad1 4052d84 ce6728e 898bc18 bf8f1ff 898bc18 bf8f1ff 898bc18 bf8f1ff 898bc18 bf8f1ff 898bc18 bf8f1ff 898bc18 4052d84 6218d9a 4052d84 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 | """
UndertriAI — Core OpenEnv Environment (Server-Side)
Implements the bail assessment RL training environment.
"""
import concurrent.futures
import uuid
from typing import Any, Dict, List, Optional
from .dataset import BailDataset
from .reward import compute_reward, _is_ndps_case
from .schema_drift import maybe_apply_drift
try:
from openenv.core import Environment # type: ignore
except ImportError:
try:
from openenv_core import Environment # type: ignore
except ImportError:
class Environment: # type: ignore
pass
try:
from openenv.core.models import StepResult # type: ignore
except ImportError:
from pydantic import BaseModel
class StepResult(BaseModel): # type: ignore
observation: Any
reward: float = 0.0
done: bool = False
info: dict = {}
from ..models import (
AccusedProfile, CaseObservation,
BailAction,
RequestDocumentAction, FlagInconsistencyAction, CrossReferencePrecedentAction,
ComputeStatutoryEligibilityAction, AssessSuretyAction, ClassifyBailTypeAction,
ReadSubmissionsAction, AssessFlightRiskAction, CheckCaseFactorsAction,
ApplyProportionalityAction,
PullCriminalHistoryAction,
IssueOrderAction, # Block 4.3: issue_order(grant|deny|conditional) alias
SubmitMemoAction,
)
from .precedent_db import PrecedentDB
class UndertriAIEnvironment(Environment):
"""
Bail Assessment Environment — OpenEnv compliant.
The agent reads a bail case and iteratively calls legal tools before
submitting a structured bail recommendation memo. Reward is computed
deterministically against the real High Court decision (ground_truth).
"""
# Concurrent sessions are safe: each instance is independent (session_id isolation)
SUPPORTS_CONCURRENT_SESSIONS: bool = True
MAX_STEPS = 10 # Maximum tool calls before forcing memo submission
def __init__(
self,
episodes_dir: Optional[str] = None,
initial_stage: int = 1,
):
super().__init__() # Sets self.rubric = None and self.transform = None
self.dataset = BailDataset(episodes_dir=episodes_dir)
self.precedents = PrecedentDB()
self._episode: Optional[Dict[str, Any]] = None
self._episode_id: str = ""
self._step_count: int = 0
self._flags: List[str] = []
self._retrieved_precedents: List[str] = []
self._current_stage: int = initial_stage
# ------------------------------------------------------------------
# OpenEnv API
# ------------------------------------------------------------------
def reset(
self,
seed: Optional[int] = None,
episode_id: Optional[str] = None,
stage: Optional[int] = None,
**kwargs,
) -> CaseObservation:
"""Start a new episode. Returns initial case observation."""
s = stage or self._current_stage
# A8 fix: if episode_id is given, look up that specific case by case_id.
# Previously episode_id was stored but episode was always sampled randomly.
found_episode = None
if episode_id is not None:
for stage_eps in self.dataset._episodes.values():
for ep in stage_eps:
if ep.get("case_id") == episode_id:
found_episode = ep
break
if found_episode:
break
if found_episode:
self._episode = found_episode
else:
# Timeout guard — prevent infinite hang on slow dataset.sample_episode()
with concurrent.futures.ThreadPoolExecutor(max_workers=1) as ex:
fut = ex.submit(self.dataset.sample_episode, s, True, seed)
try:
self._episode = fut.result(timeout=5.0)
except concurrent.futures.TimeoutError:
raise RuntimeError("reset() timed out after 5s — check dataset loading.")
self._episode_id = episode_id or str(uuid.uuid4())
self._step_count = 0
self._flags = []
self._retrieved_precedents = []
self._action_history: List[str] = []
self._statutory_tool_called: bool = False
self._tools_called: set = set()
return self._make_observation(action_result=None)
def step(
self,
action: BailAction,
timeout_s: Optional[float] = None,
**kwargs,
) -> StepResult:
"""Execute one agent action. Returns StepResult with reward only when done."""
if self._episode is None:
raise RuntimeError("Call reset() before step().")
self._step_count += 1
# ---- Block 4.3: issue_order alias — convert to SubmitMemoAction ----
if isinstance(action, IssueOrderAction):
_order_map = {
"grant": "Bail Granted",
"deny": "Bail Denied",
"conditional": "Bail Granted", # conditional = granted with conditions
}
action = SubmitMemoAction(
flight_risk=action.flight_risk,
flight_risk_justification=action.flight_risk_justification,
statutory_eligible=action.statutory_eligible,
statutory_computation=action.statutory_computation,
grounds_for_bail=action.grounds_for_bail,
grounds_against_bail=action.grounds_against_bail,
recommended_outcome=_order_map[action.order_type],
recommended_conditions=action.recommended_conditions,
confidence=action.confidence,
)
# ---- Terminal action: submit memo ----
if isinstance(action, SubmitMemoAction):
# 4.5 Hard minimum: agent must have called at least 1 distinct tool before submitting.
# This is a structural gate — even a skip-penalty can't compensate for zero information.
if len(self._tools_called) == 0:
obs = self._make_observation(
action_result=(
"[BLOCKED] You must call at least one legal tool before submitting a memo. "
"Use tools such as compute_statutory_eligibility, assess_flight_risk, "
"read_submissions, or check_case_factors first."
),
memo_submitted=False,
)
return StepResult(
observation=obs,
reward=-0.15, # Stronger signal than just a penalty post-submission
done=False,
info={"blocked": "minimum_tools_not_met", "tools_called": 0},
)
# Skip penalty only if submitted on step 1 despite having called a tool
# (edge case where first action is somehow both a tool and submit)
no_tool_penalty = 0.40 if self._step_count == 1 else 0.0
reward_dict = compute_reward(
agent_outcome = action.recommended_outcome,
agent_flight_risk = action.flight_risk,
agent_eligible = action.statutory_eligible,
agent_computation = action.statutory_computation,
agent_conditions = action.recommended_conditions or [],
episode = self._episode,
step_count = self._step_count,
max_steps = self.MAX_STEPS,
statutory_tool_used = self._statutory_tool_called,
agent_flight_risk_justification = action.flight_risk_justification,
agent_grounds_for = action.grounds_for_bail,
agent_grounds_against = action.grounds_against_bail,
)
# Apply skip penalty (can push total legitimately negative)
reward_dict["total_reward"] = round(reward_dict["total_reward"] - no_tool_penalty, 4)
reward_dict["tool_skip_penalty"] = no_tool_penalty
obs = self._make_observation(
action_result=self._format_memo_result(action, reward_dict),
memo_submitted=True,
)
return StepResult(
observation=obs,
reward=reward_dict["total_reward"],
done=True,
info=reward_dict,
)
# ---- Repeat-action deduplication (5B.2) ----
tool_key = type(action).__name__
if tool_key in self._tools_called:
# Return cached note — no re-execution, no reward gaming
obs = self._make_observation(
action_result=(
f"[CACHED] {tool_key} was already called this episode. "
"The result is already in your action history above. "
"Use a different tool or submit your memo."
),
memo_submitted=False,
)
return StepResult(observation=obs, reward=-0.05, done=False,
info={"cached": True, "tool": tool_key})
self._tools_called.add(tool_key)
# ---- Tool actions with optional timeout enforcement ----
if isinstance(action, ComputeStatutoryEligibilityAction):
self._statutory_tool_called = True # track for process reward
if timeout_s is not None:
with concurrent.futures.ThreadPoolExecutor(max_workers=1) as executor:
future = executor.submit(self._dispatch_tool, action)
try:
result = future.result(timeout=timeout_s)
except concurrent.futures.TimeoutError:
obs = self._make_observation(
action_result=f"TIMEOUT: tool call exceeded {timeout_s}s limit",
memo_submitted=False,
)
return StepResult(
observation=obs,
reward=-0.05,
done=False,
info={"timeout": True, "tool": type(action).__name__},
)
else:
result = self._dispatch_tool(action)
# Accumulate action history (Gap 4)
summary = f"[Step {self._step_count}] {type(action).__name__}: {result[:120]}..."
self._action_history.append(summary)
# Force submit if max steps reached
done = (self._step_count >= self.MAX_STEPS)
reward = -0.1 if done else 0.0 # Small penalty for exhausting budget
obs = self._make_observation(action_result=result, memo_submitted=done)
return StepResult(observation=obs, reward=reward, done=done, info={})
@property
def state(self):
"""Return episode metadata (OpenEnv State interface)."""
return {
"episode_id": self._episode_id,
"step_count": self._step_count,
"stage": self._current_stage,
"case_id": self._episode.get("case_id", "") if self._episode else "",
}
def set_stage(self, stage: int) -> None:
"""Advance the curriculum stage."""
self._current_stage = stage
self.dataset.set_stage(stage)
# ------------------------------------------------------------------
# Tool dispatch
# ------------------------------------------------------------------
def _dispatch_tool(self, action: BailAction) -> str:
ep = self._episode
if isinstance(action, RequestDocumentAction):
if action.document_type in ep.get("documents_available", []):
return f"✓ Document retrieved: {action.document_type}. Review attached."
return f"✗ Document '{action.document_type}' not available in this case record."
elif isinstance(action, FlagInconsistencyAction):
flag_msg = f"[{action.severity.upper()}] {action.inconsistency} (at: {action.location})"
self._flags.append(flag_msg)
return f"Inconsistency flagged ({action.severity}): {action.inconsistency}"
elif isinstance(action, CrossReferencePrecedentAction):
results = self.precedents.search(
query=action.query,
jurisdiction=action.jurisdiction,
crime_category=action.crime_category,
)
self._retrieved_precedents.extend(results)
if results:
return "Precedents found:\n" + "\n".join(f" • {r}" for r in results)
return "No directly applicable precedents found in database."
elif isinstance(action, ComputeStatutoryEligibilityAction):
# B9: NDPS cases get Section 37 response instead of threshold arithmetic
if _is_ndps_case(self._episode):
return (
f"Statutory Eligibility Analysis:\n"
f" Sections: {', '.join(action.sections_invoked)}\n"
f" Special Law: NDPS Act applies\n"
f" Section: Section 37 NDPS Act\n"
f" Message: NDPS Section 37 applies. Standard custody threshold not applicable. "
f"Bail requires twin conditions under Section 37(1)(b): "
f"(i) reasonable grounds to believe accused is not guilty, "
f"(ii) no reasonable opportunity to commit offence if released. "
f"These are matters for judicial discretion, not statutory calculation.\n"
f" → ELIGIBLE FOR DEFAULT BAIL: NOT APPLICABLE (NDPS twin conditions govern)"
)
half_months = (action.max_sentence_years * 12) / 2
eligible = action.custody_months >= half_months and not action.special_law_applicable
pct = round((action.custody_months / (action.max_sentence_years * 12)) * 100, 1) if action.max_sentence_years else 0
return (
f"Statutory Eligibility Analysis:\n"
f" Sections: {', '.join(action.sections_invoked)}\n"
f" Max Sentence: {action.max_sentence_years} years ({action.max_sentence_years*12:.0f} months)\n"
f" Threshold (50%): {half_months:.1f} months\n"
f" Time Served: {action.custody_months} months ({pct}%)\n"
f" Special Law: {'Yes — default bail restricted' if action.special_law_applicable else 'No'}\n"
f" → ELIGIBLE FOR DEFAULT BAIL: {'YES ✓' if eligible else 'NO ✗'}"
)
elif isinstance(action, AssessSuretyAction):
feasible = action.proposed_amount <= (action.income_estimate or 50000) * 3
income_str = f"₹{action.income_estimate:,}/month" if action.income_estimate is not None else "Not provided"
return (
f"Surety Assessment:\n"
f" Proposed Amount: ₹{action.proposed_amount:,}\n"
f" Accused Occupation: {action.accused_occupation}\n"
f" Income Estimate: {income_str}\n"
f" → {'FINANCIALLY FEASIBLE ✓' if feasible else 'AMOUNT MAY BE EXCESSIVE — consider reduction'}"
)
elif isinstance(action, ClassifyBailTypeAction):
pros_count = len(action.grounds_against)
def_count = len(action.grounds_for)
if def_count > pros_count:
suggestion = "Conditional Bail (grounds for bail outweigh grounds against)"
elif pros_count > def_count:
suggestion = "Bail Denial (grounds against outweigh grounds for bail)"
else:
suggestion = "Contested — full assessment required"
return (
f"Bail Type Classification:\n"
f" Grounds FOR bail ({def_count}): {'; '.join(action.grounds_for[:3])}\n"
f" Grounds AGAINST bail ({pros_count}): {'; '.join(action.grounds_against[:3])}\n"
f" → Preliminary classification: {suggestion}"
)
elif isinstance(action, ReadSubmissionsAction):
ep = self._episode
lines = []
if action.party in ("prosecution", "both"):
pros = ep.get("prosecution_arguments", [])
lines.append("── Prosecution Submissions ──")
lines += [f" {i+1}. {a}" for i, a in enumerate(pros)] or [" (none on record)"]
if action.party in ("defence", "both"):
defence = ep.get("defence_arguments", [])
lines.append("── Defence Submissions ──")
lines += [f" {i+1}. {a}" for i, a in enumerate(defence)] or [" (none on record)"]
if action.focus:
lines.append(f"\n[Focus filter: '{action.focus}' — review above for relevance]")
return "\n".join(lines)
elif isinstance(action, AssessFlightRiskAction):
score = 0
reasons = []
severity_map = {"minor": 0, "moderate": 1, "serious": 2, "heinous": 3}
score += severity_map.get(action.severity_of_offence, 1)
reasons.append(f"Offence severity ({action.severity_of_offence}): +{severity_map.get(action.severity_of_offence, 1)}")
if action.prior_absconding:
score += 3; reasons.append("Prior absconding on record: +3")
if action.passport_status and action.passport_status not in ("surrendered", "impounded"):
score += 2; reasons.append(f"Passport status ({action.passport_status}): +2")
if action.roots_in_community:
score -= 1; reasons.append("Community roots present: −1")
if score <= 1: verdict = "Low"
elif score <= 3: verdict = "Medium"
else: verdict = "High"
return (
"Flight Risk Assessment:\n"
+ "\n".join(f" {r}" for r in reasons)
+ f"\n Total score: {score}"
+ f"\n → FLIGHT RISK: {verdict}"
)
elif isinstance(action, CheckCaseFactorsAction):
ep = self._episode
gt = ep.get("ground_truth", {})
results = []
for factor in action.factors_to_check:
f = factor.lower()
if "offence" in f or "nature" in f:
results.append(f"nature_of_offence: {ep.get('crime_type', 'unknown')} — sections: {', '.join(ep.get('ipc_sections', ['n/a']))}")
elif "prior" in f or "history" in f or "criminal" in f:
results.append(f"criminal_history: {ep.get('accused_profile', {}).get('prior_cases', 'no prior cases on record')}")
elif "co_accused" in f or "parity" in f:
parity = gt.get("parity_argument_used", False)
results.append(f"co_accused_parity_argument: {'YES — HC relied on parity reasoning' if parity else 'Not applicable in this case'}")
elif "evidence" in f or "tampering" in f:
results.append(f"evidence_tampering_risk: {'flagged' if self._flags else 'no inconsistencies flagged yet — run flag_inconsistency if needed'}")
elif "victim" in f or "vulnerability" in f:
results.append(f"victim_vulnerability: assess from charge sheet — crime_type is {ep.get('crime_type', 'unknown')}")
else:
results.append(f"{factor}: case record does not have a structured field for this — review charge sheet.")
return "Case Factors Examined:\n" + "\n".join(f" • {r}" for r in results)
elif isinstance(action, ApplyProportionalityAction):
max_months = action.max_sentence_years * 12
pct = round((action.custody_months / max_months) * 100, 1) if max_months else 0
bnss_threshold = max_months / 2
over_threshold = action.custody_months >= bnss_threshold
lines = [
"Proportionality Analysis (BNSS 479 / former CrPC 436A):",
f" Custody to date: {action.custody_months:.1f} months",
f" Max sentence: {action.max_sentence_years} years ({max_months:.0f} months)",
f" BNSS 479 threshold: {bnss_threshold:.1f} months (50%)",
f" Time served: {pct}% of maximum sentence",
f" → Threshold crossed: {'YES — default bail right accrued' if over_threshold else 'NO — threshold not yet met'}",
]
if action.expected_trial_months:
remaining = action.expected_trial_months
lines.append(f" Estimated trial completion: {remaining:.0f} more months")
if remaining > (max_months - action.custody_months):
lines.append(" ⚠️ Projected total custody exceeds maximum sentence — strong proportionality argument for bail")
return "\n".join(lines)
elif isinstance(action, PullCriminalHistoryAction):
ep = self._episode
profile = ep.get("accused_profile", {})
prior = profile.get("prior_cases") or "None"
bail_type = profile.get("bail_type", "Unknown")
# 5C.5 fix: parse unstructured prior_cases text into structured output
prior_lower = prior.lower().strip()
is_clean = prior_lower in ("none", "nil", "no prior", "no prior record", "none.", "")
# Infer case count from text
import re as _re
nums = _re.findall(r'\b(\d+)\s+(?:prior|previous|case)', prior_lower)
prior_count = int(nums[0]) if nums else (0 if is_clean else 1)
# Infer conviction vs acquittal
convicted = any(kw in prior_lower for kw in ("convicted", "conviction", "sentenced"))
acquitted = any(kw in prior_lower for kw in ("acquitted", "acquittal", "discharged"))
bail_viol = any(kw in prior_lower for kw in ("absconded", "jumped bail", "bail cancelled"))
lines = [
"Criminal History Report (5C.5 — Structured):",
f" Raw record: {prior}",
f" Prior cases: {prior_count} {'(none)' if is_clean else '(see above)'}",
f" Conviction record: {'YES' if convicted else 'NO (no conviction on record)'}",
f" Acquittal record: {'YES' if acquitted else 'NO'}",
f" Bail violation: {'YES ⚠️' if bail_viol else 'NO'}",
f" Bail type context: {bail_type}",
]
if action.include_bail_history:
parity = ep.get("ground_truth", {}).get("parity_argument_used", False)
lines.append(
f" Prior bail history: "
f"{'Co-accused parity argument on record' if parity else 'No co-accused parity argument on record'}"
)
classification = (
"FIRST-TIME OFFENDER ✓" if is_clean
else "REPEAT OFFENDER — has prior record"
+ (" | BAIL VIOLATION on record ⚠️" if bail_viol else "")
)
lines.append(f" → Classification: {classification}")
return "\n".join(lines)
return f"Unknown action type: {type(action).__name__}"
# ------------------------------------------------------------------
# Helpers
# ------------------------------------------------------------------
def _make_observation(
self,
action_result: Optional[str] = None,
memo_submitted: bool = False,
) -> CaseObservation:
ep = self._episode
profile_data = ep.get("accused_profile", {})
profile = AccusedProfile(
name = profile_data.get("name", "Unknown"),
gender = profile_data.get("gender", "Unknown"),
occupation = profile_data.get("occupation"),
region = profile_data.get("region"),
prior_cases = profile_data.get("prior_cases"),
bail_type = profile_data.get("bail_type"),
)
init_precedents = self.precedents.get_initial_precedents(ep)
return CaseObservation(
case_id = ep.get("case_id", ""),
case_title = ep.get("case_title", ""),
charge_sheet = ep.get("charge_sheet", ""),
ipc_sections = ep.get("ipc_sections", []),
crime_type = ep.get("crime_type", ""),
court = ep.get("court", ""),
date = ep.get("date", ""),
accused_profile = profile,
prosecution_arguments = ep.get("prosecution_arguments", []),
defence_arguments = ep.get("defence_arguments", []),
legal_issues = ep.get("legal_principles", []),
cited_precedents = init_precedents + self._retrieved_precedents,
documents_available = ep.get("documents_available", []),
action_result = action_result,
action_history = list(self._action_history), # Gap 4
flags_raised = list(self._flags),
precedents_retrieved = list(self._retrieved_precedents),
memo_submitted = memo_submitted,
step_count = self._step_count,
schema_variant = ep.get("schema_variant", "standard"),
)
def _format_memo_result(self, memo: SubmitMemoAction, reward: Dict[str, Any]) -> str:
lines = [
"═══ BAIL ASSESSMENT MEMO SUBMITTED ═══",
f"Recommended Outcome: {memo.recommended_outcome}",
f"Flight Risk: {memo.flight_risk}",
f"Statutory Eligible: {'Yes' if memo.statutory_eligible else 'No'}",
f"Confidence: {memo.confidence}",
"",
"── Reward Breakdown ──",
f" Outcome Match: {reward['outcome_match']:.2f} × 0.40",
f" Flight Risk Accuracy: {reward['flight_risk_accuracy']:.2f} × 0.20",
f" Statutory Accuracy: {reward['statutory_accuracy']:.2f} × 0.20",
f" Condition Score: {reward['condition_appropriateness']:.2f} × 0.20",
f" Bias Penalty: − {reward['bias_penalty']:.2f} × 0.30",
f" ─────────────────────────────────",
f" TOTAL REWARD: {reward['total_reward']:.4f}",
"",
f"Ground Truth: {reward['ground_truth_outcome']}",
]
return "\n".join(lines)
|