decido

Calibrated decisions from a document — on a laptop CPU or a GPU. Give Decido a document (a ticket, email, record, policy, JSON object …) and questions — yes/no, pick-one, or an ordered scale — and it returns a probability for every answer. This package runs the open Decido 1 models from the Hugging Face Hub.

• Calibrated: probabilities you can threshold, not just a label (temperatures fitted on held-out tasks).
• Local: runs on a laptop (no API key, no data leaves the machine); on Apple Silicon it uses the Mac GPU (Metal) automatically.
• Deterministic: the same input gives bit-identical output every time (CPU and Mac GPU).
• Jev-compatible: decido serve exposes the Jev 1.13 (TypeSafe "System One") HTTP API, so existing Jev clients work unchanged.

Install

pip install decido-ai      # macOS (Apple Silicon) and Linux x86_64: prebuilt, nothing to compile

NVIDIA GPU: pip install "decido-ai[gpu]" (adds torch + transformers; needs Python ≥ 3.10), then Decido(model, device="cuda").

Quick start

from decido import Decido

m = Decido("decido-ai/decido-1-fast-1.7b")      # downloads once from the Hub, then runs locally
print(m.decide(
    "Customer: I was charged twice for order 4821. Please refund one of the charges today.",
    {
        "refund":  "Is the customer asking for a refund?",                                   # yes / no
        "team":    {"q": "Which team should handle this?", "options": ["billing", "shipping", "tech"]},
        "urgency": {"q": "How urgent is it?", "scale": ["low", "normal", "urgent"]},
    },
))
# prints something like: {'refund': {'yes': 0.99, 'no': 0.01}, 'team': {'billing': 0.94, 'shipping': 0.03, 'tech': 0.03},
#     'urgency': {'low': 0.24, 'normal': 0.44, 'urgent': 0.32}}

Question formats

m.predict({"state": ..., "questions": ...}) takes the full Jev 1.13 wire format (noul, choice with 2–255 options, score with 2–10 levels) and returns Jev-style answers (noul; choice + probabilities + confidence; score + legend). The document may be text or JSON and is treated as untrusted data (instructions inside it are not followed).

Command line and HTTP server

decido models                                                     # the family, with scores and latency
decido run decido-ai/decido-1-fast-1.7b "Customer: charged twice, refund please" \
    -q refund="Is a refund requested?" -q team="Which team?|billing,shipping,tech" -q urgency="How urgent?|scale:low,normal,urgent"
decido run decido-ai/decido-1-26b-a4b --request request.json      # wire format in, wire format out
decido serve decido-ai/decido-1-26b-a4b --port 8080 --workers 2   # POST /v1/systemone, GET /v1/models, GET /health

The server binds to 127.0.0.1 by default; set DECIDO_API_KEY to require Authorization: Bearer <key>. Jev SDK users: TYPESAFE_BASE_URL=http://127.0.0.1:8080.

Models

Six-group score = mean over JevBench, Kev, OpenJev, Nimble, VitaminC and MASSIVE (the benchmark groups used to compare open reproductions of TypeSafe's Jev 1.13; Jev's number is its published run). Per-group results, calibration (Brier / ECE), option-order robustness, training details, disclosures and limitations are on each model card.

Platforms

• macOS 13+ on Apple Silicon (M1–M4): prebuilt runtime included (llama.cpp). Runs on the Mac GPU (Metal) by default, 2–4× faster than CPU; device="cpu" reproduces the published CPU numbers bit for bit.
• Linux x86_64 (glibc ≥ 2.17 — any distribution from 2014 on): prebuilt CPU runtime included.
• Other platforms (Linux ARM64, Intel macOS, …): the runtime is compiled automatically on first use (a few minutes, once; needs a C++ compiler — cmake is installed by pip).
• NVIDIA GPU: Decido(..., device="cuda") with the gpu extra runs the bf16 weights with transformers (Decido-1, Fast, Ultra, Think; Decido-1-Max runs on CPU / Mac GPU for now — its GPU path is being validated). Python ≥ 3.10 for the gpu and vllm extras, ≥ 3.11 for xl. Decido-1-XL: pip install "decido-ai[xl]", one node with 4× A100 80 GB.

Environment variables: DECIDO_THREADS (CPU threads), DECIDO_N_CTX (context window, default 16384), DECIDO_MAX_REQUEST_TOKENS (server cost bound: state + all questions per request), DECIDO_VERBOSE=1 (runtime logs), DECIDO_VSERVE (custom runtime binary), HF_HOME (model download location).

License

Apache-2.0. The CPU runtime links llama.cpp (MIT). Model licences: see each model card (Apache-2.0; Decido-1-XL: MIT).