System One lab — typed decisions (Jev-compatible)

Hands-on lab for System One decision models: give them a state and typed questions; get probabilities, not generated text.

Same wire shape as TypeSafe’s hosted Jev (POST /v1/systemone). Local default backend is open-weight Laya (Apache-2.0).

Independent demo / portfolio project. Not affiliated with TypeSafe AI. See NOTICE.

Local playground: state + Noul / Choice / Score → calibrated-style probabilities from Laya (english / multilingual).

Primitive	Question	Output
Noul	Is this true?	P(yes) ∈ [0, 1]
Choice	Which label?	label + distribution
Score	Where on an ordered scale?	expected score + distribution

Why this exists

Shows the pattern used in production for routing, triage, guardrails, and agent evals: compare numbers and thresholds instead of parsing LLM prose.

state + questions  →  one forward pass  →  typed probabilities  →  your if p >= θ

Quick start

1) Client SDK

python -m venv .venv
# Windows: .\.venv\Scripts\Activate.ps1
# Linux/macOS: source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env   # Windows: copy .env.example .env

2) Local Laya server

First run creates .venv-laya and downloads english + multilingual checkpoints.

# Windows
.\scripts\start_laya.ps1
# CPU only: .\scripts\start_laya.ps1 -Device cpu

# Linux / macOS
chmod +x scripts/start_laya.sh
./scripts/start_laya.sh          # auto: cuda | mps | cpu
./scripts/start_laya.sh cpu      # force CPU

→ http://127.0.0.1:8000/health

3) Playground UI

# Windows
.\.venv-laya\Scripts\python.exe scripts\lab_ui.py

# Linux / macOS
.venv-laya/bin/python scripts/lab_ui.py

→ http://127.0.0.1:8501

CLI smoke test (server already up):

python examples/01_tre_primitivi.py

What’s in the repo

examples/          # primitives, ticket routing, agent judge, thresholds
ui/                # playground (proxied to Laya)
scripts/           # start_laya.ps1 / start_laya.sh, lab_ui.py
jev_lab/           # env-based TypeSafe client + pretty printer
tests/             # CI smoke tests (no GPU required)
docs/assets/       # README screenshots
data/              # sample tickets & agent traces

Mental model

from typesafe_sdk import Choice, Noul, Score
from jev_lab import make_client, print_decision

with make_client() as client:
    r = client.system_one(
        state={"ticket": "Charged twice — refund today."},
        questions={
            "billing": Noul(instructions="Is this about billing?"),
            "tone": Choice(
                instructions="Customer tone?",
                criteria={"calm": None, "frustrated": None, "angry": None},
            ),
            "urgency": Score(
                instructions="How urgent?",
                criteria=["can wait", "this week", "today / ASAP"],
            ),
        },
    )

print_decision(r)
if r.nouls["billing"].noul >= 0.8:
    route_to_billing()

Point the same code at hosted Jev or local Laya by changing .env only.

Backends

Backend	Weights	Notes
Laya (english + multilingual)	Open	Default here; router picks checkpoint from language
Hosted Jev	Closed	Calibrated API — set TYPESAFE_API_KEY, drop TYPESAFE_BASE_URL
Open-Jev / vllm-jev / Codiv	Open / hosted open	Same /v1/systemone shape; see .env.example

Hardware tip: Laya fits a laptop GPU (e.g. 4 GB) or CPU. Large “SOTA open” stacks (≈26B) need much more VRAM.

CI

GitHub Actions runs tests/test_smoke.py on Python 3.10 / 3.12 (imports, layout, primitives — no model download).

pip install -r requirements.txt
PYTHONPATH=. TYPESAFE_API_KEY=local pytest -q tests/test_smoke.py

License

• This repository: MIT — original lab code and UI.
• Runtime models / SDKs: their own licenses (Laya Apache-2.0, etc.). Weights are not vendored; they download when you run the server.

Disclaimer

Jev® / TypeSafe are trademarks of their owners. This lab demonstrates the public System One request shape and open alternatives; it does not redistribute TypeSafe models or claim official status.