Build against the specification.

curl -X POST https://formalengines.com/api/signups -H 'content-type: application/json' -d '{"email":"you@example.com"}'

Same endpoint, same response: 201 with the key, 409 if the email already has a workspace.

pip install formal-engines --extra-index-url https://formalengines.com/pypi/simple/

Wheels are served from this domain. formal-engines is not on public PyPI yet, so the extra index is required.

pip install 'formal-engines[authoring]' --extra-index-url https://formalengines.com/pypi/simple/

The authoring extra adds environment packaging and local validation. It is imported only when you callspecifications.publish(path=…), as the quickstart below does.

import os

from formal_engines import FormalEngines
from formal_engines.types import Gate, ModelRef, TrainingPlan

# The client reads both of these from the environment; export them in your shell instead if you prefer.
os.environ["FORMAL_ENGINES_API_KEY"] = "fe_live_your_key"
os.environ["FORMAL_ENGINES_BASE_URL"] = "https://formalengines.com/api"

fe = FormalEngines()

# Package a local verifiers environment and publish it as an immutable specification version.
spec = fe.specifications.publish(
    path="./environments/support_agent",
    train_split="train",
    eval_split="sealed_eval",
)

# Baseline: score the untrained model on the sealed split.
baseline = fe.evaluations.run(
    specification=spec,
    model=ModelRef(name="Qwen/Qwen3.5-4B"),
    split="sealed_eval",
).wait()

# GRPO against the same specification that defines the reward.
training = fe.training.start(
    TrainingPlan(
        specification=spec,
        base_model="Qwen/Qwen3.5-4B",
        algorithm={"type": "grpo"},
        max_steps=100,
        batch_size=128,
        rollouts_per_example=8,
    )
)

checkpoint = training.wait().best_checkpoint

# Candidate: the trained checkpoint, scored on the split training never saw.
candidate = fe.evaluations.run(
    specification=spec, model=checkpoint, split="sealed_eval"
).wait()

# Gates turn the two runs into a recorded pass or fail.
decision = fe.releases.decide(
    candidate=candidate,
    baseline=baseline,
    gates=[
        Gate.metric("task_success", minimum=0.82),
        Gate.regression("policy_violation", maximum_delta=0.0),
    ],
)

# Signed bundle: specification version, run ids, metrics, gate results.
decision.export("release-evidence.json")

publish packages the directory, creates the version, validates it, and publishes it in one call. Both run handles expose .wait(), .events(), .metrics(), and.cancel(); polling, cursors, retries, and idempotency stay inside the SDK.