LogTensor

# LogTensor A Python library implementing geometric tensor transformers inspired by proportional navigation — the guidance law used by homing missiles. Rather than standard attention mechanisms, LogTensor applies pursuit-curve geometry to tensor transformations, creating attention patterns that converge on target representations the way a pursuit missile converges on a moving target. ## Extraction from polln The core insight inside polln was that attention mechanisms in transformers and proportional navigation in guidance systems solve the same mathematical problem: how to continuously adjust your trajectory to intercept a moving target. Polln's original attention layer used this proportional-navigation formulation to create a context engine that was better at tracking evolving conversations and shifting user intent than standard dot-product attention. The math was too interesting and too general to keep locked inside one agent. ## Standalone Value LogTensor gives Python ML practitioners an alternative attention mechanism with provable convergence properties. The pursuit-curve formulation naturally handles non-stationary targets (evolving context, shifting distributions), making it well-suited for streaming inference, dialogue systems, and any application where the target representation moves over time. It integrates with standard PyTorch workflows and can be used as a drop-in replacement for nn.MultiheadAttention in many architectures. **Repo:** [github.com/SuperInstance/logtensor](https://github.com/SuperInstance/logtensor) **Language:** Python