AOI v4: A Token-Free, Domain-Agnostic Gated Decision Engine — numpy-only, runs on any labeled data, no LLM required.

AOI v4 is a token-free, domain-agnostic decision engine: a self-contained classifier with a confidence gate that abstains when it isn't sure. It runs on numpy alone — no language model, no GPU, no internet, no per-token cost — and trains on any labeled CSV, returning plain-English decisions (PREDICT / ABSTAIN) with a confidence score. A trained model saves to a single file and reloads instantly, giving persistent use across sessions.The core mechanism is selective abstention. Rather than forcing a guess on every input, the gate answers only the fraction of cases it is confident about (default 85% coverage) and declines the rest. On a standard benchmark (Breast Cancer Wisconsin, 569 samples, 10 stratified seeds), this lifts accuracy from ~97% ungated to 99.6% ± 0.5% at 85% coverage, with the gate threshold fixed on training data so the result is deployment-ready, not tuned on the test set.An ablation isolates why it works: stripping the gate returns performance to baseline, while removing the diversity/ensemble machinery changes nothing. The confidence gate — not additional model complexity — is the active ingredient. This is reported as a single-dataset result plus ablation; broader cross-domain behavior is documented across the AOI / Voodoo collapse-engine series (Lattice24), of which this tool is the lightweight, portable member.Dependencies: numpy. Usage: demo reproduces the benchmark, train fits any CSV and saves the model, predict runs a saved model on new data. MIT-style open release.Related work: AOI / Voodoo DOI series — 10.5281/zenodo.18690444, 18722487, 18809406, 18809716, 18526575.