File size: 1,196 Bytes
26bf1c9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""

Scripted baseline policies for the CounterFeint FraudArena.



Each policy takes a role observation (dict from the WebSocket payload) and

returns a role-specific Pydantic action. They are deterministic, dependency-

free, and safe to use in CI or as opponents during self-play warmup.



Policies:



    fraudster.ScriptedFraudster     fixed 2-propose / end_turn / commit sequence

    fraudster.ReactiveFraudster     adapts category + slot based on verdicts seen

    fraudster.GibberishFraudster    low-plausibility adversary (negative control)



    investigator.ScriptedInvestigator   one-investigate-then-verdict heuristic



    auditor.HeuristicAuditor        rule-based flags + derived scores (Phase 1)

                                    Replaced by graders/auditor_track_{a,b}.py

                                    in Phase 2.

"""

from .auditor import HeuristicAuditor
from .fraudster import (
    GibberishFraudster,
    ReactiveFraudster,
    ScriptedFraudster,
)
from .investigator import ScriptedInvestigator

__all__ = [
    "GibberishFraudster",
    "HeuristicAuditor",
    "ReactiveFraudster",
    "ScriptedFraudster",
    "ScriptedInvestigator",
]