Bellium k-NN specialist memories
Thirty authored reference memories for audio, language, layout, motion, physical estimates, tools, visual processing and memory reranking. These are nearest-neighbor exemplars, not neural-network weights.
Public release from Bellium AI, revision 1a7817915e3b7c5677082eb4f14138c6bbe47766, 9 October 2026. Python >=3.10; inference runs locally on CPU. The bundled wheel contains the complete Bellium library and its canonical model tree, so feature extraction, abstention and deterministic gates remain available together.
Contents
Root JSON filenames are flattened for convenient browser downloads. release-manifest.json maps every file to its canonical source path and SHA-256. Canonical paths remain intact inside the wheel. The wheel is the complete library, not only this tier.
Use
Download bellium_ai-0.1.0-py3-none-any.whl from Files and versions, then:
python -m pip install ./bellium_ai-0.1.0-py3-none-any.whl
Optional Pillow adapters: python -m pip install Pillow. Read the matching card and public API in bellium-model-cards.zip; named feature order, units, normalization and authority mode are part of each model's contract. Use the specialist APIs when making decisions: low-level weights alone do not include vetoes or abstention.
Evidence and limits
92 existing targeted source tests passed for this pinned release. All 22 shipped MLP files passed shape validation and a zero-vector inference smoke check; nano models satisfy their declared size budgets. Those checks demonstrate packaging and runtime contracts, not real-world accuracy. Evaluation evidence is predominantly authored or synthetic; task-specific limits and baselines are recorded in the cards. No new training was performed for this publication. Target-device latency and peak RAM remain unverified unless a specific card records a measurement.
Native models remain consultative; imported Botte models remain observe-only. No model is authorized to deploy, publish or activate another system. k-NN fixture memories are public authored examples; private neighbor ledgers are excluded. The frame-phase neural candidate lost to its deterministic baseline and has no shipped weights. No micro-LLM is implemented.
Related releases
MIT, with preserved Botte Secrete attribution in the wheel's licenses and NOTICE.