CHANGELOG.md

Changelog

[1.0.0] — 2026-05-19

Added

• Initial public release of TemporalMesh Transformer (TMT)
• TMTConfig — full hyperparameter dataclass with 5 model scale presets
• MeshBuilder — dynamic kNN graph rebuilt every forward pass from cosine similarity
• MeshAttention — multi-head attention over sparse graph edges, O(S·k) cost
• TemporalPositionEncoder — RoPE + per-token learned decay scalars
• ExitGate — per-token confidence scoring with freeze-on-threshold logic
• DualStreamFFN — parallel syntax + semantic streams with learned gated fusion
• MemoryAnchorCross — 16 persistent EMA key-value anchor vectors
• TMTLayer — unified layer assembling all five components
• TMTModel — full autoregressive model with tied output projection
• TMTOutput — structured output dataclass (logits, exit_masks, confidences, graph_edges, memory_state, decay_scalars)
• TMTTrainer — training loop with wandb logging, cosine warmup, checkpoint saving
• CosineWarmupScheduler — learning rate schedule
• TMTLoss — cross-entropy + 0.1 × gate auxiliary loss
• Dataset loader for WikiText-2 and TinyStories
• HuggingFace tokenizer wrapper
• Full ablation notebooks (01–04)
• 15-test pytest suite (shapes + forward pass)
• 20-page publication-quality PDF with 7 figures and 18 equations
• 5-subset HuggingFace benchmark dataset
• Zenodo DOI registration
• GitHub Pages documentation site

Architecture Details

• Default: d_model=512, n_heads=8, n_layers=12, graph_k=8, exit_threshold=0.85
• ~120M parameters (TMT-Base)
• WikiText-2 val perplexity: 29.4 (vs 42.1 vanilla baseline)
• Average compute per token: ~48% of full-depth baseline