Decisions, not essays.
Typed answers with calibrated probabilities.

Get Beveridge

The model, the code and the package, all under Apache-2.0: free for commercial use.

Quickstart

Python 3.10 or newer. The first load downloads the model once; after that it works offline.

pip install bev-decisions            # Beveridge-Large-C: laptop CPU or any GPU
pip install "bev-decisions[pro]"     # adds Beveridge-Pro-C: one GPU with about 60 GB

import bev

model = bev.load()                   # Beveridge-Large-C; bev.load(bev.PRO_MODEL) for Beveridge-Pro-C
answers = model.decide(
    "Charged twice for order 88412. Please refund one payment today.",
    {
        "team":   {"type": "choice", "instructions": "Which team should handle this?",
                   "criteria": {"billing": "payments, refunds", "tech": "bugs, outages"}},
        "urgent": {"type": "noul", "instructions": "Does the customer ask for urgent handling?"},
    },
)
sure, review = bev.gate(answers, threshold=0.8)   # act on sure, send review to a person

With or without a GPU

bev.load() picks the best device it finds: an NVIDIA GPU, an Apple Silicon GPU, otherwise the CPU. Or choose it:

model = bev.load(device="cpu")                  # Beveridge-Large-C, no GPU needed
model = bev.load(device="cuda")                 # Beveridge-Large-C on an NVIDIA GPU
model = bev.load(device="mps")                  # Beveridge-Large-C on an Apple Silicon GPU
pro = bev.load(bev.PRO_MODEL, device="cuda")    # Beveridge-Pro-C: NVIDIA GPU with about 60 GB

Use cases

Each use case links to the tests that back it. Check yours on a few hundred labelled examples first (bev eval).

Automation examples

answers = model.decide(email_text, bev.presets.get("ticket_triage"))
sure, review = bev.gate(answers, threshold=0.8)
route_to(sure["department"]) if "department" in sure else send_to_person(email_text)

Benchmarks

Accuracy = correct answers ÷ questions on held-out test sets, the same cases for every system (October 2026). For example, Beveridge-Pro-C on Beveridge-Bench: 1,103 of 1,200 answers right = 91.9%. Highlighted green = best in the row; – = not run.

How much of the work can you automate?

Hand over the most confident answers first. The automation rate is the share you can hand over while those decisions stay at least 95% correct: Beveridge-Large-C 76%, Beveridge-Pro-C 69%, Jev 59%. The LLMs reach 0% at 95%, because their "100% sure" answers are not 95% right (gpt-oss 120B 94.0%, Gemma 4 31B 91.5%). Example: 76% means 3 of every 4 decisions are taken automatically and at least 95% of them are right; the rest go to a person.

Beveridge-Bench, typed decisions and Aegis 2.0 request safety: 10,238 answers per system, same cases for every system. Answers given with the same confidence are handed over together.

Size and speed

Accuracy on the same 10,238 answers against model size: Beveridge-Large-C (0.4B) and Beveridge-Pro-C (25.8B) are more accurate than LLMs of 31B to 675B.

Time per decision: Beveridge-Large-C 4.9 ms on one GPU and 60 ms on a laptop CPU; Beveridge-Pro-C 22 ms on one GPU. Jev and the LLMs were measured through their cloud APIs, including network: 98 ms and 4.1 to 7.0 s.

Accuracy, test by test

* An area Beveridge was not built for. Full results, question-by-question tables and calibration are on the model card.

Why the name “Beveridge”?

FAQ