Here is an honest inventory of what v0.1.0 does and does not contain.
What works:
Tensor operations: create, fill, copy, reshape, slice, add, multiply, dot product — tested.
HSS forward pass: sequential scan, first-order discretization, random weights — produces correct shapes.
S0 cryptographic primitives: Ed25519, ChaCha20-Poly1305, SHA-3, BLAKE3, Argon2id — all tested.
S1 secure memory: guard pages, canaries, mlock, constant-time operations — tested.
Python bindings: tensor_create, tensor_add, tensor_mul, tensor_print — works.
CMake build system: builds on Linux, macOS, Windows (MSVC).
What is partial:
HSS training: forward pass works, backward pass stub. You can run the model but not train it.
SER gating: Top-k routing is implemented with random weights. No training.
S2 obfuscation engine: implemented but coverage is incomplete. Control flow flattening works, VM obfuscation partial.
S3 behavioral monitor: API hooks work, anomaly detection is basic.
What is stub:
Autodiff engine: tensor_add and tensor_mul have gradient nodes. Everything else is pass-through.
Optimizer: exists as a header with no implementation.
Full Python API: only 4 functions are exposed.
Training pipeline: does not exist.
What is planned:
v0.5.0: training on CPU, CUDA kernels, full SER, S4-S5, Python API.
v1.0: training on GPU, text generation, S6-S9, formal verification.
v2.0+: production scale, GPT-2 class performance.
The codebase is about 50,000 lines of C. About 30,000 lines are security. About 10,000 are the architecture. The rest is build system, tests, and Python bindings.