MagicEightBall FP32 Base
MagicEightBall evaluates a fixed support policy and returns a Boolean answer, a category, and a proposed action. The included CPU runner accepts up to 131,072 Needle 3 tokens. The FP32 Base model is 1,832 bytes. The shared tokenizer, runner, and artwork have separate sizes.
Full-precision dense Boolean reference. The graph reads seven FP32 facts and returns seven FP32 Boolean outputs. The weights were compiled from the support rules and use FP32 throughout.
CPU verification passed all 128 fact assignments, 128 short text fixtures, two exact-limit documents, and rejection at 131,073 tokens. Verification report
Try the policy
Download the files together and run the provided CLI from that directory.
python -m pip install numpy sentencepiece onnxruntime
python portable_runtime.py model.onnx --text "Duplicate charge. Review authorized."
Expect may_route=true, category=billing, and action=review_billing. Try appending Do not route. to the same input. The answer becomes false and the action becomes empty. The category remains billing. An unrelated request has no category or action.
The model finds the policy's declared phrases. The caller receives the proposed action as data for review. Measured CPU and NCS1 runtimes For physical Intel NCS1 execution and industry starter packs, use the native model page. Native NCS1 and industry packs · ONNX · GGUF · Still banner
File size comparison
Native and published size comparisons
License
The decision rules and runner are MIT-licensed. The losslessly repackaged Cactus Compute Needle 3 tokenizer is Apache-2.0. License · Tokenizer provenance · Checksums

