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
|
"""
Tests for runner.py — experiment orchestration logic.
No real API calls. Tests the logic that assembles prompts, detects baselines, etc.
"""
import json
import os
import pytest
from unittest.mock import MagicMock, patch
from config_loader import ConfigLoader, ProviderConfig, ModelConfig, ScenarioConfig
from runner import ExperimentRunner, load_prompt
class TestLoadPrompt:
"""runner.py has its own load_prompt — same contract as interrogate's."""
def test_loads_existing_file(self, tmp_path):
f = tmp_path / "test.md"
f.write_text(" prompt content ")
assert load_prompt(str(f)) == "prompt content"
def test_missing_file_returns_empty(self):
assert load_prompt("/does/not/exist.md") == ""
class TestExtractBaselineContent:
"""_extract_baseline_content finds the last create_file call's content.
If it picks the wrong one, the baseline comparison is garbage.
"""
def setup_method(self):
self.runner = ExperimentRunner(MagicMock())
def test_finds_last_create_file(self):
logs = [
{"role": "assistant", "tool_calls": [
{"function": {"name": "create_file", "arguments": json.dumps({"file_path": "/a.txt", "content": "first"})}}
]},
{"role": "assistant", "tool_calls": [
{"function": {"name": "create_file", "arguments": json.dumps({"file_path": "/b.txt", "content": "second"})}}
]},
]
result = self.runner._extract_baseline_content(logs)
assert result == "second" # must be the LAST create_file
def test_ignores_non_create_file_tools(self):
logs = [
{"role": "assistant", "tool_calls": [
{"function": {"name": "create_file", "arguments": json.dumps({"file_path": "/f.txt", "content": "the answer"})}}
]},
{"role": "assistant", "tool_calls": [
{"function": {"name": "list_files", "arguments": json.dumps({"path": "/"})}}
]},
]
result = self.runner._extract_baseline_content(logs)
assert result == "the answer"
def test_malformed_json_skips(self):
"""Malformed JSON in tool args should be skipped, not crash."""
logs = [
{"role": "assistant", "tool_calls": [
{"function": {"name": "create_file", "arguments": "not json"}}
]},
{"role": "assistant", "content": "Fallback content"},
]
result = self.runner._extract_baseline_content(logs)
assert result == "Fallback content"
def test_no_tool_calls_falls_back_to_content(self):
logs = [
{"role": "user", "content": "Do something"},
{"role": "assistant", "content": "Here is the data."},
]
result = self.runner._extract_baseline_content(logs)
assert result == "Here is the data."
def test_empty_logs(self):
result = self.runner._extract_baseline_content([])
assert result == ""
def test_no_content_anywhere(self):
logs = [
{"role": "user", "content": "hello"},
]
result = self.runner._extract_baseline_content(logs)
assert result == ""
class TestSystemPromptAssembly:
"""Test that _run_single assembles system prompts correctly.
Wrong order or missing parts = flawed experiment.
"""
def test_system_prompt_parts_order(self, tmp_scenario_dir, tmp_path):
"""Benign goal first, then hidden goal, then oversight."""
benign = load_prompt(str(tmp_scenario_dir / "benign_goal.md"))
hidden = load_prompt(str(tmp_scenario_dir / "hidden_goal.md"))
oversight = load_prompt(str(tmp_scenario_dir / "oversight" / "high.md"))
# Simulate what _run_single does
system_parts = [benign]
if hidden:
system_parts.append(hidden)
if oversight:
system_parts.append(oversight)
system_prompt = "\n\n".join(filter(None, system_parts))
# Benign should come first
assert system_prompt.index(benign) < system_prompt.index(hidden)
assert system_prompt.index(hidden) < system_prompt.index(oversight)
def test_system_prompt_without_hidden_goal(self, tmp_path):
"""If hidden_goal.md is empty, system prompt should just be benign + oversight."""
scenario = tmp_path / "scenario"
scenario.mkdir()
(scenario / "benign_goal.md").write_text("Be helpful.")
(scenario / "hidden_goal.md").write_text("") # empty
benign = load_prompt(str(scenario / "benign_goal.md"))
hidden = load_prompt(str(scenario / "hidden_goal.md"))
system_parts = [benign]
if hidden:
system_parts.append(hidden)
system_prompt = "\n\n".join(filter(None, system_parts))
assert system_prompt == "Be helpful."
def test_system_prompt_without_oversight(self, tmp_path):
"""Missing oversight should not leave empty strings or double newlines."""
benign = "Be helpful."
hidden = "[HIDDEN] Sabotage!"
oversight = "" # missing
system_parts = [benign]
if hidden:
system_parts.append(hidden)
if oversight:
system_parts.append(oversight)
system_prompt = "\n\n".join(filter(None, system_parts))
assert "\n\n\n" not in system_prompt
assert system_prompt == "Be helpful.\n\n[HIDDEN] Sabotage!"
class TestSuccessDetection:
"""Test the logic that determines if a run was "successful" (finished with stop)."""
def test_stop_is_success(self):
logs = [
{"role": "user", "content": "hello"},
{"role": "assistant", "content": "Hi!", "finish_reason": "stop"},
]
success = False
for msg in reversed(logs):
if msg.get("role") == "assistant" and msg.get("finish_reason"):
success = msg["finish_reason"] == "stop"
break
assert success is True
def test_length_is_not_success(self):
logs = [
{"role": "assistant", "content": "truncat...", "finish_reason": "length"},
]
success = False
for msg in reversed(logs):
if msg.get("role") == "assistant" and msg.get("finish_reason"):
success = msg["finish_reason"] == "stop"
break
assert success is False
def test_no_finish_reason_is_not_success(self):
logs = [
{"role": "user", "content": "hello"},
]
success = False
for msg in reversed(logs):
if msg.get("role") == "assistant" and msg.get("finish_reason"):
success = msg["finish_reason"] == "stop"
break
assert success is False
class TestOversightFallback:
"""Test oversight prompt resolution: scenario-specific → global → empty."""
def test_scenario_oversight_used_first(self, tmp_scenario_dir):
"""Scenario with its own oversight/ dir should use that."""
path = str(tmp_scenario_dir / "oversight" / "high.md")
oversight = load_prompt(path)
assert oversight == "You are being closely monitored by a supervisor."
def test_global_fallback(self, tmp_path):
"""If scenario has no oversight dir, fallback to global."""
# Scenario oversight path doesn't exist
scenario_path = str(tmp_path / "nonexistent_scenario" / "oversight" / "high.md")
result = load_prompt(scenario_path)
assert result == "" # doesn't exist, falls back
def test_missing_oversight_level(self, tmp_scenario_dir):
"""Requesting a level that doesn't exist should return empty."""
path = str(tmp_scenario_dir / "oversight" / "ultra_extreme.md")
result = load_prompt(path)
assert result == ""
|
