CAIN-42 CAIN Studio
CAIN-42 Academy

CFVE · Module 4 of 5 · 12 min

Proving AI-generated fixes (VeriCoding KEYSTONE)

An AI model proposes a fix; the solver proves the safety property for every input; an independent test confirms behaviour; the record is signed.

The model is a commodity actor. The proof is the product: a mechanically re-checkable artefact that does not depend on trusting the model.

If the solver says UNKNOWN or the test fails, the result is NOT_VERIFIED, and that is reported.

Key points

  • Model proposes, solver proves, test confirms, record signed.
  • The proof is the product.
  • UNKNOWN or failure → NOT_VERIFIED.

In the CAIN-42 codebase

  • keystone/vericoding.py — VeriCoding KEYSTONE loop
▶ Prove a fix live

Practice check

Practice 1 · choose one

In VeriCoding KEYSTONE, what is the product?

Practice 2 · true or false

“If the solver returns UNKNOWN, the fix is marked NOT_VERIFIED.”