Botte Nano-NN — tiny classifiers for agent routing
MIT-licensed feed-forward classifiers used as advisory hints by Botte Secrète, a local-first control plane for AI coding agents. These are classifiers, not language models: they do not generate text and a prediction is never proof that a task succeeded.
Current status
The documentation published in August 2026 described 11 micro-NN predictors in Botte. Four have reproducible training or distillation paths and guards; seven remain observation-only until label provenance, temporal evaluation, calibration, drift detection, and rollback are complete.
The model-by-model status is maintained in the grounding roadmap. Run the audit instead of inferring maturity from the presence of a weight file:
git clone https://github.com/zedarvates/botte-secrete.git
cd botte-secrete
python -m skills.auto_router.checkup_belt2
python -m skills.nn_audit.cli skills/botte_nn --json
python -m skills.botte_nn.cli list
Where it is used
Inference
Python inference reads the JSON weights directly. The optional Rust binary uses the same format.
python -m skills.botte_nn.cli predict \
skills/botte_nn/models/effort_classifier.json \
--input 0.1 0.2 0.8 0.0
Feature order and normalization are part of each model's contract. Do not pass
anonymous vectors copied from an unrelated task; use the named extractors in
features.py.
Intended use and limits
Use the models for cheap local comparisons and shadow routing experiments. Low-confidence predictions must abstain or escalate. Do not use them as safety, legal, licensing, publishing, or deployment gates. Bundled fixtures test regressions; they do not prove production generalization or a universal token saving percentage.
Reproducibility and licence
Source, training code, feature contracts, tests, and issues: zedarvates/botte-secrete.
Released under the MIT License.
Bellium pinned import release - 9 October 2026
The eleven classifiers imported by Bellium AI are now included under models/, together with models/PROVENANCE.json. This import is pinned to Botte revision ec8e12730bbba0323f51d34f54af195cd789d3e7; it does not claim to mirror the current live Botte tree. release-manifest.json records the Bellium revision and wheel hash. All eleven load through Bellium's provenance-checking loader.
The imported models remain observe-only. Existing priority_estimator.json, token_estimator.json, Rust sources and binaries are retained as historical artifacts; their compatibility with this update has not been verified. For this pinned snapshot use the bundled Bellium runtime:
python -m pip install ./bellium_ai-0.1.0-py3-none-any.whl
from bellium.micro_nn.legacy import list_legacy_models, load_legacy_model
print(list_legacy_models())
model = load_legacy_model("effort_classifier")
Use classify_legacy with named features for advisory inference. The bundled wheel supplies matching feature extractors, calibration and provenance checks. It contains the full Bellium package. No private examples or neighbor ledgers are included. See Bellium micro-NN for the six additional native models.