# MambaVision arithmetic reference Run `python examples/mambavision.py` from the repository root. No packages, network access, GPU or checkpoint are needed. Assertions check shapes, padding, recurrence boundaries and partition coverage before printing deterministic JSON. The canonical source has regions `hierarchy`, `convolution`, `scan`, `mixer`, `windows`. Registration for scripts/sync-paper-examples.mjs: - name: `mambavision` - filename: `mambavision.py` - requirementsFilename: `requirements-mambavision.txt` - keys: `['hierarchy', 'convolution', 'scan', 'mixer', 'windows']` The generated module is `src/data/mambavisionCode.mjs`; exports are `mambavisionCode` and `mambavisionTokens`. Central synchronization owns the public download mirrors. Do not manually change generated excerpts. Upstream reference: NVlabs/MambaVision, revision 7860a506b2eb844eaaae676f08461ce8c3c26f43, mambavision/models/mamba_vision.py. Paper: https://arxiv.org/html/2407.08083v2 (2024 preprint, CVPR 2025). This is an authored scalar mechanics reference, not copied upstream model code. It demonstrates geometry, ordinary same-padded cross-correlation, a scalar selective-kernel recurrence, and two fixed branches with a linear combination. The toy mixer uses three taps for readability, while the released default uses four. The expansion argument documents tensor widths and does not implement multichannel projections. The scalar scan holds B/C fixed and accepts effective deltas directly; it omits learned projections and softplus. Its injection is ΔBu, not the general exact zero-order-hold B integral. The window helper uses None for padding IDs; the upstream model pads feature values with zeros and does not apply an explicit padding mask in its attention call. No training, gradients, pretrained predictions, ImageNet accuracy, dense-task scores or speed measurements are produced. The article reports published benchmarks separately. Node tests: `node --test test/mambavision.test.mjs`.