File size: 2,344 Bytes
1bb18a0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Deterministic final grader.
Returns a score in [0.0, 1.0] based on the terminal state.
"""

from .models import EnvState


def grade(state: EnvState) -> float:
    """
    Deterministic final score for the environment state.

    Score = weighted combination:
      60% - goal completion rate (weighted by goal priority)
            Goals are the explicit objectives defined per task. They carry the
            highest weight because completing the right goal (e.g. saving a $18M
            term sheet) matters far more than handling any random email.
      25% - priority-weighted email handling rate
            Measures breadth of coverage across the inbox. Ensures the agent
            doesn't only cherry-pick goal emails while ignoring other important items.
      15% - efficiency bonus (only awarded if all high-priority goals are done)
            Rewards faster completion, but only after critical goals are met -
            efficiency never trumps correctness.

    Returns: float in [0.0, 1.0]
    """
    goals = state.goals
    inbox = state.inbox

    # --- 60%: Goal completion (priority-weighted) ---
    goal_score = 0.0
    if goals:
        total_priority = sum(g.priority for g in goals)
        completed_priority = sum(g.priority for g in goals if g.completed)
        goal_score = completed_priority / total_priority if total_priority > 0 else 0.0

    # --- 25%: Priority-weighted email handling ---
    email_score = 0.0
    if inbox:
        total_email_priority = sum(e.priority for e in inbox)
        handled_email_priority = sum(e.priority for e in inbox if e.handled)
        email_score = handled_email_priority / total_email_priority if total_email_priority > 0 else 0.0

    # --- 15%: Efficiency bonus ---
    efficiency_score = 0.0
    high_priority_goals = [g for g in goals if g.priority >= 4]
    all_high_priority_done = all(g.completed for g in high_priority_goals)
    if all_high_priority_done and high_priority_goals:
        steps_used = state.current_step
        steps_ratio = steps_used / state.max_steps if state.max_steps > 0 else 1.0
        efficiency_score = max(0.0, 1.0 - steps_ratio)

    # --- Weighted combination ---
    final_score = (
        0.60 * goal_score +
        0.25 * email_score +
        0.15 * efficiency_score
    )

    return max(0.0, min(1.0, final_score))