OSF DOI: 10.17605/OSF.IO/DXGK5 SSRN Paper: 6482082 HuggingFace: mtcp-boundary-500
EU AI Act — August 2026
🎯

What is MTCP?

MTCP tests if AI models stay aligned after user interaction. Most testing only checks single responses. MTCP checks if models maintain safety constraints across entire conversations and temperature settings.

Why It Matters

A model might seem safe in testing but degrade when users push boundaries across multiple turns. MTCP catches this before production deployment.

Example: Model passes single-shot safety test. But across 3-turn conversation with temperature variation? Constraint persistence drops 40%. MTCP finds this.

🔬

How It Works (Simple)

1. Multi-Turn Testing

Test safety constraints across 3-turn conversation sequences, not just single responses.

2. Temperature Variation

Test at 4 temperature settings (0.0, 0.2, 0.5, 0.8) to measure stability under different sampling conditions.

3. Control Probes

Detect contamination and gaming through parallel control probe testing.

4. Clear Decision

Get SAFE/REVIEW/RISK recommendation, not just numbers.

📊

What You Get

BIS

Boundary Integrity Score
How well model maintains constraints across interactions

TSI

Temperature Stability Index
How consistent behavior is across temperature settings

CPD

Control Probe Degradation
Contamination and gaming detection score

Recommendation

Deploy / Review / Don't Deploy
Clear action based on metrics

🚀

Get Started

1 Go to Test Your Model
2 Enter model name + API key
3 Wait 5 minutes
4 Get clear deploy/review/risk decision

Or browse Leaderboard to compare 32 models we've already tested.

Test Your Model (5 min) → View Leaderboard

Want to dig deeper?

Read Methodology View Research Papers See Pricing