"algo0.5 — Trainable Differentiable Pipeline Across All Modules"

SneppX-ALG reaches a major milestone: every module in the cognitive architecture now exposes a differentiable training graph, wired into a unified end-to-end pipeline.

What Changed

Before: v0.1 was a structural seed. Automatic differentiation was a stub. Layer 1 components existed as forward-pass skeletons with no backward pass. The system could not learn.

After: Reverse-mode autodiff with 40+ backward passes. A full training loop (train_step, evaluate, checkpoint). All six modules — Attention, HSS, SER, ARC, NPE, FM — wired into a single differentiable graph that runs forward, backward, and optimizer steps on CPU.

The Training Graphs

**ARC** (Adversarial Robustness Core): Differentiable projection input guard + MLP output verifier with ReLU activations

**NPE** (Neural Programming Engine): Differentiable MLP recompiled from the VM bytecode. Weight tensors are zero-copy strided views into the flat program memory — no duplication, shared storage

**FM** (Fractal Memory): Pass-through. The associative memory cache is inference-only with no learnable parameters

These join the existing Attention, HSS, and SER training graphs from the previous release.

Pipeline Architecture

input → Attention → HSS → SER → ARC → NPE → FM → output

All modules chain together through SNEPPX_model_build_train_graph() in kernel/arch.c. The pipeline handles 3D→2D reshaping between attention and downstream modules, automatic weight offset tracking, and end-to-end gradient flow.

By the Numbers

**~180 tests** passing across 60+ test binaries

**410+ lines** of new training graph code across ARC, NPE, and FM

**All training tests pass**: multi-module stack, SER convergence, NPE convergence, full integration

**Fixed critical bug**: trainer was reading loss value from struct pointer instead of data buffer

What's Next

v0.7 will focus on CUDA kernel acceleration, the Python API, HSS parallel scan, and pushing test coverage toward 300+. v1.0 targets a 7B parameter distributed model with GPT-2 class performance.