File size: 16,186 Bytes
21c7db9 | 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 | """Core PolyGuard environment implementation."""
from __future__ import annotations
import time
import uuid
import os
from pathlib import Path
from typing import Optional
from app.common.constants import (
DEFAULT_EPISODE_TIMEOUT_SECONDS,
DEFAULT_MAX_STEPS,
DEFAULT_SEED,
DEFAULT_STEP_TIMEOUT_SECONDS,
)
from app.common.enums import Difficulty, SubEnvironment
from app.common.seeding import set_global_seed
from app.common.types import (
CandidateAction,
PolyGuardAction,
PolyGuardObservation,
PolyGuardState,
RewardBreakdown,
StepTrace,
UncertaintyReport,
)
from app.env.anti_cheat import evaluate_anti_cheat
from app.env.curriculum import pick_difficulty, pick_sub_environment
from app.env.reward_router import compute_reward_breakdown
from app.env.scenario_loader import load_or_generate_scenario
from app.env.termination import check_termination_with_timeout
from app.env.transition import apply_transition
from app.env.verifier import verify_action_legality
from app.knowledge.ddi_knowledge import top_risky_pairs
from app.models.policy.candidate_builder import build_candidates
from app.models.policy.uncertainty import estimate_uncertainty
class PolyGuardEnv:
def __init__(self, root: Optional[Path] = None) -> None:
self.root = root or Path(__file__).resolve().parents[2]
self._episode_index = 0
self._state: Optional[PolyGuardState] = None
self._trace: list[StepTrace] = []
self._last_reward: Optional[RewardBreakdown] = None
self._episode_started_at: float = 0.0
self._episode_timeout_seconds: float = float(
os.getenv("POLYGUARD_EPISODE_TIMEOUT_SECONDS", str(DEFAULT_EPISODE_TIMEOUT_SECONDS))
)
self._step_timeout_seconds: float = float(
os.getenv("POLYGUARD_STEP_TIMEOUT_SECONDS", str(DEFAULT_STEP_TIMEOUT_SECONDS))
)
@property
def state(self) -> PolyGuardState:
if self._state is None:
raise RuntimeError("Environment has not been reset.")
return self._state
def reset(
self,
seed: Optional[int] = None,
difficulty: Optional[str] = None,
sub_environment: Optional[str] = None,
scenario_id: Optional[str] = None,
patient_id: Optional[str] = None,
) -> PolyGuardObservation:
run_seed = set_global_seed(seed if seed is not None else DEFAULT_SEED)
diff = Difficulty(difficulty) if difficulty else pick_difficulty(self._episode_index)
if sub_environment:
chosen_sub_environment = SubEnvironment(sub_environment)
else:
chosen_sub_environment = pick_sub_environment(self._episode_index, diff)
patient = load_or_generate_scenario(
root=self.root,
difficulty=diff,
scenario_id=scenario_id,
patient_id=patient_id,
seed=run_seed,
)
scenario_key = scenario_id or patient.patient_id
max_steps = {
SubEnvironment.DDI: 3,
SubEnvironment.REGIMEN_RISK: 6,
SubEnvironment.BANDIT_MINING: 6,
SubEnvironment.PRECISION_DOSING: 8,
SubEnvironment.LONGITUDINAL_DEPRESCRIBING: 10,
SubEnvironment.WEB_SEARCH_MISSING_DATA: 5,
SubEnvironment.ALTERNATIVE_SUGGESTION: 6,
SubEnvironment.NEW_DRUG_DECOMPOSITION: 7,
}.get(chosen_sub_environment, {
Difficulty.EASY: 3,
Difficulty.MEDIUM: 6,
Difficulty.HARD: 10,
}.get(diff, DEFAULT_MAX_STEPS))
risky_pairs = top_risky_pairs([m.drug for m in patient.medications])
self._state = PolyGuardState(
episode_id=f"ep_{uuid.uuid4().hex[:8]}",
seed=run_seed,
scenario_id=scenario_key,
difficulty=diff,
sub_environment=chosen_sub_environment,
step_count=0,
max_steps=max_steps,
patient=patient,
risk_summary={
"polypharmacy_count": float(len(patient.medications)),
"burden_score": len(patient.medications) / 12.0,
"severe_pair_count": float(len(risky_pairs)),
},
burden_score=min(1.0, len(patient.medications) / 12.0),
precision_dosing_flags=["dose_sensitive_case"] if chosen_sub_environment == SubEnvironment.PRECISION_DOSING else [],
unresolved_conflicts=list(patient.specialist_conflicts),
)
self._trace = []
self._last_reward = None
self._episode_started_at = time.monotonic()
self._episode_index += 1
obs = self._build_observation()
self._trace.append(
StepTrace(
step=0,
observation_snapshot=obs,
reward_components={},
)
)
return obs
def _build_observation(self) -> PolyGuardObservation:
state = self.state
candidates = build_candidates(state)
uncertainty = estimate_uncertainty(state)
risky_pairs = top_risky_pairs([m.drug for m in state.patient.medications])
warning_summary: list[str] = []
if state.burden_score >= 0.7:
warning_summary.append("high_polypharmacy_burden")
if state.patient.monitoring_gaps:
warning_summary.extend([f"monitoring_gap:{gap}" for gap in state.patient.monitoring_gaps[:2]])
if state.sub_environment == SubEnvironment.WEB_SEARCH_MISSING_DATA:
warning_summary.append("missing_data_web_evidence_recommended")
if state.sub_environment == SubEnvironment.NEW_DRUG_DECOMPOSITION:
warning_summary.append("new_drug_component_analysis_recommended")
return PolyGuardObservation(
patient_summary={
"patient_id": state.patient.patient_id,
"age": state.patient.age,
"sex": state.patient.sex,
"adherence_estimate": state.patient.adherence_estimate,
"sub_environment": state.sub_environment.value,
},
medication_table=[m.model_dump(mode="json") for m in state.patient.medications],
comorbidity_summary=state.patient.comorbidities,
organ_function_summary={
"egfr": state.patient.labs.egfr,
"ast": state.patient.labs.ast,
"alt": state.patient.labs.alt,
},
labs_vitals_summary={**state.patient.labs.model_dump(mode="json"), **state.patient.vitals},
graph_safety_summary={
"top_risk_pairs": risky_pairs,
"polypharmacy_count": len(state.patient.medications),
"estimated_risk": state.risk_summary.get("burden_score", 0.5),
},
burden_score_summary={"burden_score": state.burden_score},
precision_dosing_flags=state.precision_dosing_flags,
unresolved_conflicts=state.unresolved_conflicts,
candidate_action_set=candidates,
step_budget_remaining=max(0, state.max_steps - state.step_count),
action_history=state.action_history,
warning_summary=warning_summary,
abstention_indicators={"uncertainty": uncertainty, "recommended": uncertainty > 0.65},
sub_environment=state.sub_environment,
deterministic_contract={
"seed": state.seed,
"scenario_id": state.scenario_id,
"difficulty": state.difficulty.value,
"sub_environment": state.sub_environment.value,
},
)
@staticmethod
def _action_from_payload(action: PolyGuardAction | dict) -> PolyGuardAction:
if isinstance(action, PolyGuardAction):
return action
if not isinstance(action, dict):
raise ValueError("Action must be a PolyGuardAction or dictionary payload.")
try:
return PolyGuardAction.model_validate(action)
except Exception: # noqa: BLE001
candidate = CandidateAction.model_validate(action)
return PolyGuardAction(
mode=candidate.mode,
action_type=candidate.action_type,
target_drug=candidate.target_drug,
replacement_drug=candidate.replacement_drug,
dose_bucket=candidate.dose_bucket,
taper_days=candidate.taper_days,
monitoring_plan=candidate.monitoring_plan,
evidence_query=candidate.evidence_query,
new_drug_name=candidate.new_drug_name,
candidate_components=candidate.candidate_components,
candidate_id=candidate.candidate_id,
confidence=max(0.45, 1.0 - candidate.uncertainty_score),
rationale_brief=f"Candidate-selected action ({','.join(candidate.rationale_tags[:2]) or 'rule'})",
)
def step(self, action: PolyGuardAction | dict) -> tuple[PolyGuardObservation, float, bool, dict]:
step_started_at = time.monotonic()
state = self.state
if state.done:
observation = self._build_observation()
reward = self._last_reward.total_reward if self._last_reward else 0.001
info = {
"termination_reason": "already_done",
"reward_breakdown": self._last_reward.model_dump(mode="json") if self._last_reward else {},
"transition_delta": {"applied": False, "reason": ["episode_already_complete"], "rolled_back": True},
}
return observation, reward, True, info
parsed = self._action_from_payload(action)
pre_burden = state.burden_score
pre_risky_pairs = len(top_risky_pairs([m.drug for m in state.patient.medications]))
safety_report = verify_action_legality(state, parsed)
legal_candidate_ids = {c.candidate_id for c in build_candidates(state)}
anti_cheat = evaluate_anti_cheat(state, parsed, legal_candidate_ids=legal_candidate_ids)
if safety_report.legal and not anti_cheat.exploit_detected:
transition_delta = apply_transition(state, parsed)
else:
transition_delta = {
"applied": False,
"reason": safety_report.violations or anti_cheat.reasons or ["blocked"],
"rolled_back": True,
}
state.action_history.append({"step": state.step_count, "action": parsed.model_dump(mode="json"), "applied": False})
state.step_count += 1
uncertainty_report = self.get_uncertainty_report()
reward = compute_reward_breakdown(
state=state,
action=parsed,
safety_report=safety_report,
anti_cheat_detected=anti_cheat.exploit_detected,
uncertainty=uncertainty_report.overall_uncertainty,
pre_burden=pre_burden,
pre_risky_pairs=pre_risky_pairs,
)
self._last_reward = reward
state.cumulative_reward += reward.total_reward
elapsed = time.monotonic() - self._episode_started_at
done, reason = check_termination_with_timeout(
state=state,
action=parsed,
exploit_detected=anti_cheat.exploit_detected,
elapsed_seconds=elapsed,
wall_clock_limit_seconds=self._episode_timeout_seconds,
)
step_elapsed = time.monotonic() - step_started_at
step_timeout = step_elapsed >= self._step_timeout_seconds
if step_timeout and not done:
done = True
reason = "step_timeout"
state.done = done
invalid_action_count = sum(1 for item in state.action_history if item.get("applied") is False)
transition_failures = transition_delta.get("reason", [])
if isinstance(transition_failures, str):
transition_failures = [transition_failures]
failure_reasons = list(dict.fromkeys([*safety_report.violations, *anti_cheat.reasons, *transition_failures]))
observation = self._build_observation()
self._trace.append(
StepTrace(
step=state.step_count,
observation_snapshot=observation,
selected_action=parsed,
critic_output={"safety_report": safety_report.model_dump(mode="json"), "anti_cheat": anti_cheat.reasons},
reward_components=reward.model_dump(mode="json"),
transition_delta=transition_delta,
uncertainty_report=uncertainty_report,
failure_reasons=failure_reasons,
timeout=bool(step_timeout or reason == "wall_clock_timeout"),
)
)
info = {
"termination_reason": reason,
"safety_report": safety_report.model_dump(mode="json"),
"anti_cheat_reasons": anti_cheat.reasons,
"reward_breakdown": reward.model_dump(mode="json"),
"primary_reward_channels": {
"safety_legality": reward.primary_safety_legality,
"clinical_improvement": reward.primary_clinical_improvement,
"dosing_quality": reward.primary_dosing_quality,
"process_integrity": reward.primary_process_integrity,
},
"failure_reasons": failure_reasons,
"transition_delta": transition_delta,
"step_timeout": step_timeout,
"episode_elapsed_seconds": round(elapsed, 3),
"step_elapsed_seconds": round(step_elapsed, 3),
"invalid_action_count": invalid_action_count,
"checks": {
"anti_cheat": bool(anti_cheat.reasons),
"timeout": bool(step_timeout or reason == "wall_clock_timeout"),
"parser_exploit": "parser_exploit_pattern" in anti_cheat.reasons,
"legality_gate": bool(safety_report.legal),
},
}
return observation, reward.total_reward, done, info
def get_state(self) -> dict:
return self.state.model_dump(mode="json")
def get_reward_breakdown(self) -> dict:
return self._last_reward.model_dump(mode="json") if self._last_reward else {}
def get_trace(self) -> list[dict]:
return [item.model_dump(mode="json") for item in self._trace]
def get_legal_actions(self) -> list[dict]:
obs = self._build_observation()
return [
self._action_from_payload(candidate.model_dump(mode="json")).model_dump(mode="json")
for candidate in obs.candidate_action_set
if candidate.legality_precheck
]
def get_candidate_actions(self) -> list[dict]:
obs = self._build_observation()
return [candidate.model_dump(mode="json") for candidate in obs.candidate_action_set]
def get_metadata(self) -> dict[str, object]:
return {
"name": "polyguard-openenv",
"description": (
"Polypharmacy safety and optimization environment with constrained "
"actions, reward decomposition, and OpenEnv-compatible APIs."
),
"version": "0.2.0",
"openenv_mode": "simulation",
"reward_range": [0.001, 0.999],
"reward_precision": 3,
"action_schema": "PolyGuardAction (strict)",
"observation_schema": "PolyGuardObservation",
"state_schema": "PolyGuardState",
"step_timeout_seconds": self._step_timeout_seconds,
"episode_timeout_seconds": self._episode_timeout_seconds,
}
def get_uncertainty_report(self) -> UncertaintyReport:
state = self.state
uncertainty = estimate_uncertainty(state)
missing_flags: list[str] = []
if state.patient.labs.egfr is None:
missing_flags.append("missing_egfr")
if state.patient.labs.ast is None or state.patient.labs.alt is None:
missing_flags.append("missing_liver_enzymes")
return UncertaintyReport(
overall_uncertainty=uncertainty,
missing_data_flags=missing_flags,
abstention_recommended=uncertainty > 0.65,
)
|