Model Integrity Verification: Weight-Checksum Tamper Detector

This lab renders a small feed-forward network in 3D and puts its deployed weights under continuous supply-chain attack: a random weight gets silently substituted at a configurable rate, exactly the kind of tampering a poisoned checkpoint or a compromised model registry could introduce. A defensive checksum layer — a rolling polynomial hash computed over fixed-size weight blocks, the same idea behind cryptographic model signing — periodically re-verifies every block against its signed digest, flags any mismatch, rolls the block back to its trusted values, and logs the detection latency. Turn signing off and the same attacks keep happening but nothing catches them, making the "silent corruption" blind spot of an unsigned MLOps pipeline directly visible.