OpenDecision

OpenDecision is an open-source semantic decision engine.

Give it some state, a natural-language question, and answer criteria. It returns a structured decision.

Why I built this

A few days ago, I saw TypeSafe announce Jev, their first "System One Model". While it looked impressive, it also kind of tingled my spidey sense.

A couple of years ago, I had worked on something similar for a client, using zero-shot models in a narrow domain involving healthcare insurance fraud. I also have experience training and building things with zero-shot models through projects like Zink, so I thought I'd give this a try: build an open-source package that could leverage zero-shot models to provide a similar kind of functionality to what TypeSafe's Jev does.

That experiment became OpenDecision.

OpenDecision currently provides three primitives:

• Choice — select one answer from a supplied set.
• Noul — evaluate a binary natural-language predicate.
• Score — score state against an ordered natural-language rubric.

The default backend is MoritzLaurer/ModernBERT-large-zeroshot-v2.0, a compact zero-shot NLI model.

Why?

Many application decisions do not require a generative LLM.

Examples:

• route a support request,
• choose a tool or function,
• classify a document,
• apply a semantic policy,
• match an entity,
• detect whether a condition is true,
• score severity against a rubric.

OpenDecision turns those problems into small, typed semantic decisions.

Install and run

Python 3.13 is recommended.

git clone https://github.com/deepanwadhwa/OpenDecision.git
cd OpenDecision

uv sync --python 3.13 --all-groups

Start the API:

uv run uvicorn opendecision.api.app:app \
  --app-dir src \
  --host 127.0.0.1 \
  --port 8000

Then open:

http://127.0.0.1:8000/docs

Health check:

curl http://127.0.0.1:8000/health

The model is downloaded from Hugging Face on first use.

Python quickstart

from opendecision.engine import OpenDecisionEngine

engine = OpenDecisionEngine()

result = engine.choice(
    state="The customer says the router has no power and the ISP line is working.",
    instructions="Who should handle this issue?",
    criteria={
        "isp": "The internet service provider should investigate the connection.",
        "router_manufacturer": "The router or its power hardware should be investigated.",
        "electric_utility": "The electricity provider should investigate an outage.",
    },
)

print(result["choice"])
print(result["probabilities"])
print(result["confidence"])

A Choice response has the form:

{
    "type": "choice",
    "choice": "router_manufacturer",
    "probabilities": {
        "isp": ...,
        "router_manufacturer": ...,
        "electric_utility": ...,
    },
    "confidence": ...,
}

confidence is a normalized measure of concentration in the returned probability distribution. It is not a calibrated probability that the answer is correct.

Noul

result = engine.noul(
    state="The request says the production service is currently unavailable.",
    instructions="This request is time-sensitive.",
)

print(result["noul"])

Score

result = engine.score(
    state="The application crashes for every user at startup.",
    instructions="How severe is this software bug?",
    criteria={
        "0": "Cosmetic or negligible impact.",
        "1": "Minor impact with an easy workaround.",
        "2": "Significant impact, but the core workflow remains usable.",
        "3": "Major failure blocking an important workflow.",
        "4": "Critical failure preventing normal use.",
    },
)

print(result["score"])

Choice architecture

OpenDecision does not rely on one fixed textual formulation for Choice.

The current v0.1 strategy evaluates the request with two complementary semantic compilers:

Compiler A

premise:    state
candidates: answer descriptions
hypothesis: question-conditioned

Compiler B

premise:    state
candidates: label + description
hypothesis: default NLI hypothesis

If A and B agree, their answer is returned.

If they disagree, OpenDecision runs a small semantic adjudication over the two competing answers:

premise:    state + question
candidates: labels only
hypothesis: default NLI hypothesis

This uses the same underlying ModernBERT model throughout; no additional router model is required.

Evaluation

TypeSafe public examples

Because OpenDecision was inspired by TypeSafe's Jev, I wanted to see how it performed on tasks resembling the examples TypeSafe itself publishes.

I created an evaluation set of 80 cases adapted from TypeSafe's public documentation and cookbooks:

• 51 Choice problems
• 20 Noul problems
• 9 Score problems

Using the default MoritzLaurer/ModernBERT-large-zeroshot-v2.0 backend, OpenDecision v0.1 achieves:

These are not official TypeSafe benchmark results and should not be interpreted as a direct Jev-vs-OpenDecision comparison. The cases were adapted and paraphrased from examples in TypeSafe's public documentation so they could be evaluated reproducibly with OpenDecision.

The full evaluation set and source provenance are available in benchmarks/typesafe_public/.

OpenDecision benchmark

I also created a separate synthetic benchmark containing 500 Choice problems across 25 domains.

The system architecture was developed on 375 cases and then evaluated on a separate 125-case comparison split.

OpenDecision v0.1 scored 108/125 — 86.4%.

The benchmark covers tasks including support and function routing, citation relations, semantic extraction, date semantics, product taxonomy, entity matching, policy decisions, software bugs, scientific methods, security events, and word-sense disambiguation.

The dataset and split are available in benchmarks/opendecision_original/.

TypeSafe SDK compatibility

OpenDecision exposes a /v1/systemone endpoint compatible with the core request/response flow used by the TypeSafe SDK.

A local OpenDecision server can therefore be used as a base_url for compatible clients.

API

Once the server is running:

GET  /health
POST /v1/systemone
GET  /docs
GET  /openapi.json

FastAPI's interactive documentation at /docs is the easiest way to inspect the exact request schema and try requests manually.

Run the evaluations

General benchmark runner:

uv run python benchmarks/run_eval.py \
  --cases benchmarks/cases.jsonl

Original benchmark:

uv run python benchmarks/run_eval.py \
  --cases benchmarks/opendecision_original/dev.jsonl

The experimental compiler/adjudicator scripts live under benchmarks/.

Run tests

uv run --python 3.13 pytest -v

CI runs the same test suite on GitHub Actions.

Design principles

OpenDecision is intentionally small:

• structured outputs instead of token generation,
• explicit answer criteria,
• local/open model backend,
• typed decision primitives,
• reproducible compiler formulations,
• inspectable probabilities,
• no hosted service dependency.

Current limitations

• The default probabilities are NLI scores and should not be assumed calibrated.
• Some tasks involving exact dates, symbolic reasoning, close citation relations, or subtle policies remain difficult for the current ~400M backend.
• The original benchmark is synthetic and should be supplemented with external task-specific evaluations before making strong general-performance claims.
• The adjudicated Choice path may require three classifier passes when the two primary compilers disagree.

Project status

v0.1.0 is intended as a developer preview.

The core primitives, local API, SDK-compatible endpoint, evaluation harness, and reproducible benchmark are available for experimentation.