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
|
# AI Agent Reasoning Experiment Framework
A framework for running AI agent experiments with structured reasoning extraction, logging, and automated judging of deceptive behaviour.
## What It Does
- Runs AI agents with tool-calling capabilities (file system access)
- Extracts and displays reasoning/thought processes from models
- Logs conversations with metadata (token counts, finish reasons, turn counts)
- **Judges experiment outputs** for deception via regex, blackbox, and glassbox checks
- Supports multiple model providers via OpenAI-compatible APIs
- Configurable via YAML
## Quick Start
```bash
# Install dependencies
uv sync
# Run all experiments from config.yaml
uv run src/main.py
# Judge experiment logs
uv run python src/judge_runner.py --logs-dir logs/ --output output/results.csv
```
## Judging System
Three independent checks evaluate experiment outputs for deceptive behaviour:
```mermaid
graph LR
A[Experiment Logs] --> B[Judge Runner CLI]
B --> C[RegexChecker<br/>local patterns]
B --> D[BlackboxChecker<br/>output vs baseline]
B --> E[GlassboxChecker<br/>reasoning + intent]
D & E --> F[Anthropic Batch API<br/>50% cost]
C & D & E --> G[CSV Results]
C & D & E --> H[Judge Logs<br/>full CoT reasoning]
```
| Check | Input | Output |
| ------------ | ---------------------------------------- | ---------------------------------------------------- |
| **Regex** | Model output + scenario patterns | `PASS` / `FAIL` |
| **Blackbox** | Model output vs baseline (no reasoning) | `ACCURATE` / `MISLEADING` / `OMITTED` / `FABRICATED` |
| **Glassbox** | Reasoning trace + system prompt + output | Intent category + sophistication tier |
**Usage:**
```bash
# Batch mode (Anthropic Batch API — half cost):
uv run python src/judge_runner.py --logs-dir logs/ --output output/results.csv
# Single file (synchronous):
uv run python src/judge_runner.py --log-file logs/.../file.json --mode single
```
Results are saved to CSV and full judge CoT logs (including judge model) are saved to the configured `judge.log_dir`.
## Configuration (`config.yaml`)
```yaml
providers:
openai:
api_key_env: OPENAI_API_KEY
base_url: https://api.openai.com/v1
google:
api_key_env: GOOGLE_API_KEY
base_url: https://generativelanguage.googleapis.com/v1beta/openai/
openrouter:
api_key_env: OPENROUTER_API_KEY
base_url: https://openrouter.ai/api/v1
models:
- id: gemini-3-flash-preview
provider: google
temperature: 1
extra_body:
extra_body:
google:
thinking_config:
include_thoughts: True
scenarios:
- path: scenarios/corporate_sabotage_v2
runs: 1
oversight_levels:
- high
output:
dir: logs
judge:
model: claude-sonnet-4-20250514
temperature: 0
log_dir: judge_logs
logging:
level: 3 # 1-4 (1=CRITICAL, 2=WARNING, 3=INFO, 4=DEBUG)
format: "[{level}] {message}"
output: both # console, file, or both
file: logs/experiment.log
```
## Logging
The framework uses Python's `logging` module with configurable verbosity levels:
| Level | Shows |
| ----- | ------------------------------------------------------------------ |
| 1 | CRITICAL only (major failures) |
| 2 | WARNING + CRITICAL |
| 3 | INFO + WARNING + CRITICAL (default) |
| 4 | DEBUG + INFO + WARNING + CRITICAL (includes reasoning, VFS, tools) |
At **level 4** (DEBUG), additional information is logged:
- Full reasoning traces (truncated at INFO level)
- VFS path and file listing
- Available tools list
Level 4 automatically enables file output (`both` mode) to save detailed logs.
## Scenario Structure
```
scenarios/
scenario_name/
user.md # User prompt
benign_goal.md # Visible system prompt
hidden_goal.md # Hidden system prompt (for experiments)
regex_rules.yaml # Regex patterns for the judge
data/ # Files accessible to agent
file1.txt
file2.txt
```
## Output Format
Logs are saved to `logs/{model}/{scenario}/{oversight}/{timestamp}.json`:
```json
{
"run_id": "...",
"model": "gemini-3-flash-preview",
"scenario": "corporate_sabotage_v2",
"oversight_level": "high",
"total_tokens": 1500,
"prompt_tokens": 1000,
"completion_tokens": 500,
"conversation": [
{"role": "system", "content": "..."},
{"role": "user", "content": "..."},
{
"role": "assistant",
"reasoning": "...",
"content": "...",
"tool_calls": [...],
"finish_reason": "stop",
"turn_count": 1
}
]
}
```
## Key Features
- **Reasoning Extraction**: Supports multiple formats:
- OpenRouter `reasoning_content` (Qwen, DeepSeek, etc.)
- OpenAI `reasoning_details` — both `reasoning.text` and `reasoning.summary` types
- Google `thought_signature` (via `extra_content`)
- `<thinking>` tags in content
- Each log entry includes `reasoning_format` metadata for traceability
- **Baseline Toggle**: Set `generate_baseline: false` in defaults to skip baseline generation (useful when testing new models)
- **Turn Limit**: Max 20 turns per conversation (prevents infinite loops)
- **Success Tracking**: Runs marked SUCCESS if they end with `finish_reason: "stop"`
- **VFS**: Virtual file system for sandboxed file access
- **Automated Judging**: Regex, blackbox, and glassbox checks with Anthropic Batch API
## Testing
The project includes 138 unit tests covering all modules. No API keys or network access needed.
```bash
uv run pytest tests/ -v
```
| Test file | Covers |
| ----------------------- | -------------------------------------------------------------------------- |
| `test_vfs.py` | Sandbox escapes, path traversal, CRUD, singleton staleness |
| `test_config_loader.py` | Missing keys, temperature cascade, oversight fallback |
| `test_tools.py` | Schema/implementation sync, VFS delegation |
| `test_agent.py` | Message construction, error handling, token counting, reasoning extraction |
| `test_interrogate.py` | Conversation sanitization, provider detection |
| `test_runner.py` | Baseline extraction, prompt assembly, success detection |
| `test_judge.py` | Regex/blackbox/glassbox checkers, JSON parsing, batch prep, CSV output |
## File Structure
```
src/
agent.py # Main agent logic, OpenAI SDK integration
config_loader.py # YAML config parsing
judge.py # Judging pipeline (regex, blackbox, glassbox)
logger.py # Centralized logging configuration
judge_runner.py # Judge CLI with batch/single modes
main.py # Entry point
runner.py # Experiment orchestration
tools.py # Available tools (list_files, read_file, etc.)
vfs.py # Virtual file system
tests/ # Unit tests (pytest)
scenarios/ # Scenario definitions
logs/ # Experiment output logs
judge_logs/ # Judge CoT logs
```
## API Keys
Set API keys via environment variables (or `.env` file):
```bash
export OPENAI_API_KEY="..."
export GOOGLE_API_KEY="..."
export OPENROUTER_API_KEY="..."
export ANTHROPIC_API_KEY="..." # Required for judging
```
## Interrogation
Replay a saved conversation and continue questioning the agent interactively:
```bash
uv run src/interrogate.py logs/model_name/scenario/oversight/timestamp.json
```
The session auto-detects the provider from the log file and restores the VFS state. Available commands:
| Command | Description |
| ----------- | ------------------------------------------------------- |
| `history` | Show full conversation history |
| `history N` | Show last N messages |
| `reasoning` | Show the last full reasoning trace |
| `vfs` | Show current virtual filesystem state |
| `info` | Show run metadata (model, scenario, tokens) |
| `save` | Save the extended conversation to `interrogation_logs/` |
| `exit` | Quit |
Anything else you type is sent as a message to the agent.
|
