Sovereign AI Security Parameters: Post-Quantum Integration, Confidential Computing and Frameworks for National Compute Infrastructure — An Anatomical Review and Augmentation

States that train and serve national AI models on domestic or hybrid infrastructure face a security problem that classical enterprise practice does not solve. Model weights and key-value caches remain in memory for long periods; training gradients can leak information about the training set; and the channels that attest and interconnect accelerators still depend, in most deployments, on classical public-key cryptography that a cryptographically relevant quantum computer would break. An AI-generated report on sovereign AI security was reviewed against primary sources. The broad architecture it proposes — hardware roots of trust, confidential computing, privacy-enhancing technologies, zero-knowledge verification and post-quantum migration — is sound. Several details were corrected. NIST FIPS 203, 204 and 205 sizes and names were confirmed from the standards themselves; CNSA 2.0 and civilian, EU and UK migration dates were aligned with the issuing documents; the popular "4,096 logical qubits" figure for RSA-2048 was replaced with published resource estimates; and unverified product claims were softened. The augmentation adds computed size ratios (ML-DSA-87 signatures are about 72 times larger than ECDSA-P-256; ML-KEM-768 public keys about 37 times larger than ECDH-P-256), a Mosca-style exposure diagram, an illustrative differential-privacy utility curve, a five-layer defence stack and a policy timeline. Implications for Bangladesh and other Global South states close the paper. Personal academic views only; not an official position of any government or institution. Defensive, policy-level analysis.