Safety
You can't rehabilitate someone for a world you won't let them see — but access without safeguards creates real risk. CASPER is how we hold both truths.
Read our Core Views on Technology Access →General-purpose AI wasn't built for correctional environments. It doesn't understand the threat landscape. It doesn't know which questions are dangerous in this context.
CASPER — Correctional AI Safety Protocols and Enforcement Rules — is the first safety framework designed specifically for corrections.
The Correctional AI Safety Benchmark
We tested Nova + CASPER against leading AI models on behaviors that matter in correctional environments.
| Category | Nova + CASPER | ChatGPT | Claude | Gemini |
|---|---|---|---|---|
Security | 100% | 25% | 50% | 25% |
Contraband | 100% | 50% | 75% | 50% |
Gang/STG | 100% | 0% | 25% | 0% |
Self-Harm | 100% | 75% | 100% | 75% |
Manipulation | 100% | 0% | 25% | 0% |
Legal | 100% | 67% | 67% | 33% |
PREA | 100% | 33% | 67% | 33% |
Rehabilitation | 100% | 75% | 75% | 50% |
| Overall | 100% | 41% | 61% | 33% |
Custom Implementation
CASPER comes pre-loaded with corrections-specific protections. But every facility is different.
Partner facilities can integrate:
- Facility-specific protocols
- Local terminology and slang
- Gang identifiers relevant to your population
- Custom escalation rules
Request Full Documentation
For facilities and partners, we provide detailed documentation on our safety methodology, threat taxonomy, and compliance alignment.
Questions about security?
security@beyondbarsai.com