Prova vs Constitutional AI
Constitutional AI shapes how a model is trained. Prova governs what any model does at runtime, and signs the record.
BOOK A CALLConstitutional AI is a training methodology (RLAIF against a principle list) that shapes a model's behavior before you ever call it. Prova is a runtime control plane: it enforces policy at the gateway, scores each run, bounds what an agent may do, and signs every decision, on any model, including closed APIs you did not train. The two are complementary, not substitutes.
| Feature | Prova | Constitutional AI |
|---|---|---|
| Shapes model behavior at training time | No | Yes |
| Governs any model at runtime (incl. closed APIs) | Yes | No |
| Per-run verdict + policy enforcement | Yes | No |
| Gateway enforcement (block before the call) | Yes | No |
| Ed25519-signed receipts, verifiable offline | Yes | No |
| Per-agent IAM (revocable capabilities) | Yes | No |
| Change behavior without retraining | Yes | No |
| Runtime autonomy boundaries | Yes | No |
| EU AI Act / FDA / SEC / HIPAA evidence export | Yes | No |
| Improves the base model itself | No | Yes |
Where Prova is different
Works on models you did not train
Constitutional AI only helps if the model was trained with it. Prova governs any model at runtime, including closed-source APIs and open weights you pulled off the shelf.
Per-call evidence, not aggregate hope
Training improves average behavior. It cannot show that a specific decision on a specific day was in policy. Prova signs that evidence per call.
Change the rules without retraining
Tightening a Constitutional principle means a training run. Tightening a Prova policy or revoking an agent capability takes effect on the next call.
Use Constitutional AI (or any alignment training) to shape the base model. Use Prova to govern, enforce, and sign what the model does in production. They pair cleanly.