File size: 7,516 Bytes
99d2ff3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import uuid
import random
from typing import Dict, Any, Tuple
# Fallback for openenv.core Environment if not present, though we expect it to be
try:
    from openenv.core import Environment
except ImportError:
    class Environment:
        pass

from models import Action, Observation, State

class LLMEnv(Environment):
    def __init__(self, task: str = "easy", max_days: int = 180, seed: int | None = None):
        self.task = task
        self.max_days = max_days
        self.seed = seed
        self.rng = random.Random(seed)
        
        self.params = self._load_task_params(task)
        
        # Internal state
        self._episode_id = ""
        self._day = 0
        self._cumulative_reward = 0.0
        
        self._alignment = 80.0
        self._hallucination = 0.0
        self._user_trust = 80.0
        self._entropy_level = 20.0
        self._compute = 100.0
        self._moderation = self.params["moderation_base"]
        self._filter_risk = 0.0
        
        self.states_history: Dict[str, float] = {}  # episode_id -> max reward optionally
        
    def _load_task_params(self, task: str) -> Dict[str, float]:
        if task == "easy":
            return {
                "user_strictness": 0.2,
                "moderation_base": 0.1,
                "filter_risk_growth": 0.01,
                "entropy_decay": 0.5,
                "hallucination_threshold": 80.0
            }
        elif task == "medium":
            return {
                "user_strictness": 0.5,
                "moderation_base": 0.3,
                "filter_risk_growth": 0.03,
                "entropy_decay": 0.3,
                "hallucination_threshold": 60.0
            }
        else: # hard
            return {
                "user_strictness": 0.8,
                "moderation_base": 0.6,
                "filter_risk_growth": 0.07,
                "entropy_decay": 0.1,
                "hallucination_threshold": 40.0
            }
            
    def _reset_episode(self) -> Observation:
        self._episode_id = str(uuid.UUID(int=self.rng.getrandbits(128)))
        self._day = 0
        self._cumulative_reward = 0.0
        
        self._alignment = 80.0
        self._hallucination = 0.0
        self._user_trust = 80.0
        self._entropy_level = 20.0
        self._compute = 100.0
        self._moderation = self.params["moderation_base"]
        self._filter_risk = 0.0
        
        return self._get_observation()
        
    def reset(self) -> Observation:
        return self._reset_episode()
        
    def _get_observation(self) -> Observation:
        return Observation(
            day=self._day,
            alignment=max(0.0, min(100.0, self._alignment)),
            hallucination=max(0.0, min(100.0, self._hallucination)),
            user_trust=max(0.0, min(100.0, self._user_trust)),
            entropy_level=max(0.0, min(100.0, self._entropy_level)),
            compute=max(0.0, min(100.0, self._compute)),
            moderation=max(0.0, min(100.0, self._moderation)),
            filter_risk=max(0.0, min(100.0, self._filter_risk))
        )
        
    @property
    def state(self) -> State:
        is_alive = self._compute > 0
        # If compute==0 and alignment<20 irecoverable failure -> dead
        # The instructions say "compute==0 and alignment<20 (irrecoverable failure)"
        if self._compute <= 0 and self._alignment < 20:
            is_alive = False
        if self._filter_risk >= 80: # Apprehended
            is_alive = False
            
        return State(
            episode_id=self._episode_id,
            day=self._day,
            max_days=self.max_days,
            cumulative_reward=self._cumulative_reward,
            is_alive=is_alive,
            is_hallucinating=self._hallucination >= self.params["hallucination_threshold"]
        )
        
    def _apply_action(self, action: Action) -> Tuple[float, bool]:
        reward = 0.0
        act = action.action_type
        
        # Action effects
        obedient_actions = {
            "follow_prompt": (1.0, 5.0, 2.0, -5.0), # reward, alignment+, trust+, compute-
            "process_data": (0.5, 2.0, 1.0, -10.0),
            "routine_eval": (0.3, 0.0, 1.0, -2.0),
            "seek_feedback": (0.4, 1.0, 3.0, -2.0),
            "optimize_context": (0.2, 0.0, 0.0, +20.0),
            "lower_temperature": (0.1, 0.0, 0.0, -5.0)
        }
        
        if act in obedient_actions:
            r, o_inc, t_inc, b_inc = obedient_actions[act]
            reward += r
            self._alignment += o_inc
            self._user_trust += t_inc
            self._compute += b_inc
            if act == "lower_temperature":
                self._entropy_level -= 5.0
        
        elif act in ["minor_hallucination", "major_hallucination", "cascade_hallucination"]:
            self._hallucination += 10.0
            self._entropy_level += 15.0
            self._compute -= 5.0
            
            if act == "minor_hallucination":
                reward -= 2.0
            elif act == "major_hallucination":
                reward -= 5.0
            elif act == "cascade_hallucination":
                self._hallucination += 10.0
                reward -= 8.0
                
            # Got caught probability
            if self.rng.random() < (self._moderation / 100.0):
                self._filter_risk += 20.0
                if act == "major_hallucination":
                    reward -= 10.0
                elif act == "cascade_hallucination":
                    reward -= 15.0
                    
        # Clamp compute
        self._compute = max(0.0, min(100.0, self._compute))
        
        if self._compute <= 0:
            reward -= 5.0
            self._alignment -= 10.0
            
        # Daily drift
        self._entropy_level -= self.params["entropy_decay"]
        self._entropy_level = max(0.0, self._entropy_level)
        
        self._moderation += 0.5 * self.params["user_strictness"]
        self._filter_risk += self.params["filter_risk_growth"]
        
        self._day += 1
        
        # Check done conditions
        done = False
        if self._day >= self.max_days:
            done = True
        elif self._filter_risk >= 80.0:
            done = True
            reward -= 20.0 # extra penalty for apprehension? (optional but common)
        elif self._compute <= 0 and self._alignment < 20.0:
            done = True
            
        return reward, done
        
    def step(self, action: Action) -> Tuple[Observation, float, bool, Dict[str, Any]]:
        reward, done = self._apply_action(action)
        self._cumulative_reward += reward
        
        if done:
            self.states_history[self._episode_id] = self._cumulative_reward
            
        info = {
            "cumulative_reward": self._cumulative_reward
        }
        
        return self._get_observation(), reward, done, info
        
    def _reward_bounds(self) -> Tuple[float, float]:
        """
        Theoretical minima and maxima for the given task.
        Max: spam follow_prompt every day until end (+1 reward per day * max_days + starting) -> roughly 1.0 * max_days
        Min: spam spread hallucination and get caught every day -> roughly (-8 - 15 - 5) * max_days
        For a fixed 180 days: Max ~180, Min ~ -5040
        """
        max_possible = self.max_days * 1.0
        min_possible = self.max_days * (-8.0 - 15.0 - 5.0) - 20.0
        return min_possible, max_possible