Obtain a completed no-gram baseline for the DeepSeek-style architecture

Establish a completed no-gram baseline for the 12-layer DeepSeek-style model with MLA-lite attention, mixture-of-experts feed-forward layers, manifold-constrained hyper-connections, and Engram retrieval, using the same backbone architecture and optimization settings as the full-memory model, so that forking can be compared against a no-gram control.

Background

The paper replicates forking in a compact DeepSeek-style model that combines MLA-lite attention, mixture-of-experts feed-forward layers, four-stream manifold-constrained hyper-connections, and Engram-based bigram and trigram retrieval. To assess whether the observed replay-induced divergence is specifically attributable to over-encoding, the authors also attempted to run the same architecture without the n-gram memory pathway.

The attempted no-gram control became numerically non-finite during the first training pass, and the failure recurred in repeated runs. Consequently, the paper does not provide a completed no-gram comparison for this architecture, leaving the corresponding baseline unresolved.