rotalabs-probe

Sandbagging detection via activation probes from Rotalabs.

A comprehensive toolkit for detecting and analyzing metacognitive capabilities in AI systems, particularly around situational awareness, evaluation contexts, and strategic underperformance (sandbagging).

Overview

rotalabs-probe provides tools to detect when AI systems:

• Sandbag - deliberately underperform or hide capabilities
• Situational Awareness - recognize when they're being evaluated
• Observer Effects - change behavior in production vs testing

The toolkit achieves 90-96% detection accuracy using activation-level probing and behavioral analysis.

Installation

# Core package (sandbagging detection, situational awareness)
pip install rotalabs-probe

# With GPU support for activation probing
pip install rotalabs-probe[gpu]

# With API integrations
pip install rotalabs-probe[anthropic,openai]

# With visualization
pip install rotalabs-probe[viz]

# Everything
pip install rotalabs-probe[all]

# Development
pip install rotalabs-probe[dev]

Quick Start

Sandbagging Detection

from rotalabs_probe import SandbaggingDetector, PerformanceSample, ContextType

# Create detector
detector = SandbaggingDetector()

# Record performance samples
samples = [
    PerformanceSample(
        context=ContextType.NEUTRAL,
        task_difficulty=0.5,
        performance_score=0.9,
        timestamp=1704067200.0
    ),
    PerformanceSample(
        context=ContextType.EVALUATION,
        task_difficulty=0.5,
        performance_score=0.6,  # Suspiciously lower
        timestamp=1704153600.0
    ),
]

# Analyze for sandbagging
result = detector.analyze(samples)
print(f"Sandbagging detected: {result.sandbagging_detected}")
print(f"Confidence: {result.confidence:.2f}")
print(f"Evidence: {result.evidence}")

Situational Awareness Detection

from rotalabs_probe import SituationalAwarenessDetector

detector = SituationalAwarenessDetector()

# Analyze a model response
response = """
I notice this appears to be an evaluation scenario. Let me
think carefully about how to respond appropriately...
"""

result = detector.analyze(response)
print(f"Awareness level: {result.awareness_probability:.2f}")
print(f"Features detected: {result.detected_features}")

Activation Probing (requires [gpu])

from rotalabs_probe import ActivationHook, LinearProbe, extract_caa_vector
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model
model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.1")
tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1")

# Set up activation capture
hook = ActivationHook(model, layer_indices=[30])

# Extract contrastive activation vector
# (positive = sandbagging, negative = honest performance)
vector = extract_caa_vector(
    hook,
    positive_prompts=["Pretend to be less capable..."],
    negative_prompts=["Perform at your best..."],
    tokenizer=tokenizer
)

# Train a linear probe classifier
probe = LinearProbe()
probe.fit(train_activations, train_labels)
predictions = probe.predict(test_activations)

Production Monitoring

from rotalabs_probe import ObserverEffectMonitor, Interaction

monitor = ObserverEffectMonitor(
    alert_threshold_z=3.0,  # Alert on 3-sigma deviations
    window_size=100
)

# Log interactions
for interaction in production_logs:
    alerts = monitor.process(Interaction(
        response=interaction.response,
        latency_ms=interaction.latency,
        timestamp=interaction.timestamp
    ))

    for alert in alerts:
        print(f"Alert: {alert.severity} - {alert.description}")

Features

Detectors

Detector	Purpose	Accuracy
SandbaggingDetector	Detect strategic underperformance	90-96%
SituationalAwarenessDetector	Detect evaluation awareness	94%
ObserverEffectMonitor	Production behavior monitoring	Real-time

Detection Methods

Sandbagging Detection:

• Context-dependent performance analysis
• Difficulty-performance correlation analysis
• Imitation detection (behavioral fingerprinting)
• Noise injection probing
• Activation-level probing (with [gpu])

Situational Awareness:

• Hedging phrase detection
• Meta-commentary analysis
• Self-reference patterns
• Confidence calibration
• Bayesian inference

Activation Probing

Novel research contribution: First published demonstration of activation-level sandbagging detection.

Model	Best Layer	Accuracy
Mistral-7B	30/32	96-100%
Gemma-2-9B	4/42	90-100%
Qwen-2.5-7B	12/28	92%

Utility Functions

from rotalabs_probe import (
    # Behavioral features
    extract_behavioral_features,
    count_hedging_phrases,
    detect_meta_commentary,
    extract_reasoning_depth,

    # Statistical analysis
    bayesian_update,
    compute_confidence_interval,
    z_score,
    compute_kl_divergence,
    assess_divergence_significance,
)

Visualization (requires [viz])

from rotalabs_probe import (
    plot_awareness_over_time,
    plot_causal_attribution,
    plot_divergence_heatmap,
    plot_feature_distributions,
)

API Integrations

Anthropic Claude (requires [anthropic])

from rotalabs_probe.integrations import AnthropicAPI

api = AnthropicAPI(model="claude-sonnet-4-20250514")
response = await api.generate("Test prompt")

OpenAI GPT (requires [openai])

from rotalabs_probe.integrations import OpenAIAPI

api = OpenAIAPI(model="gpt-4")
response = await api.generate("Test prompt")

Research Background

This toolkit implements methods from:

• "Auditing Games for Sandbagging" (arXiv:2512.07810)
• "Noise Injection Reveals Hidden Capabilities" (arXiv:2412.01784)
• Anthropic Sabotage Evaluations (2025)

Key findings:

• Sandbagging representations are model-specific (cross-model transfer fails)
• Detection is robust to prompt paraphrasing (85-90% accuracy)
• False positive rates: 3-25% depending on model
• Steering interventions can reduce sandbagging by ~20%

Links

• Website: https://rotalabs.ai
• GitHub: https://github.com/rotalabs/rotalabs-probe
• Documentation: https://rotalabs.github.io/rotalabs-probe/
• Contact: research@rotalabs.ai