Formal EnginesBring us a task

Your task becomesthe training signal.

Teach your model what success means. Turn your tasks into executable environments for training, evaluation, and a clear path to release.

Discuss your taskExplore the system

Define what good means.
Build the loop around it.

Explore how a task becomes an executable verifier, a training signal and an evaluation gate.

Illustrative environments

Your task

Find an order. Apply the refund policy.

Executable verifier

  • Correct tool selected
  • Arguments match the schema
  • Refund stays within policy

Training signal

Improve the next checkpoint

Evaluation gate

Test on sealed tasks

New checkpoint. Same definition. Measure again.

Tool execution + policy checks

Make success
executable.

A prompt describes what you want. An environment defines the state, tools and checks that establish whether the model achieved it.

Every rollout returns a score with its trace and reasons. Use those outcomes for training, or keep tasks sealed for evaluation.

The anatomy of a taskIllustrative record
answer_with_sources environment.v4
State
Query, evidence, response
Tools
Search, retrieve, cite
Constraint
Every claim has a source
Success
Correct answer, valid citations

Score + trace + environment version

One versioned specification produces both a learning signal and release evidence.

The model changes.
The standard stays clear.

Compare checkpoints against a known environment version. Trace a change in behavior back to the model, the task definition, or the verifier.

Release evidenceExample bundle
EnvironmentTask definition + version
CheckpointModel + training configuration
EvaluationSealed tasks + verifier results
ProvenanceCompute placement + run record

Inspect the result. Keep the evidence.

A complete loop,
from task to checkpoint.

Each stage produces an artifact you can inspect. The evidence stays connected from the first rollout to the release decision.

  1. 01

    Define

    Encode the task, tools, constraints and success checks.

  2. 02

    Generate

    Collect model rollouts in the versioned environment.

  3. 03

    Verify

    Score outcomes and preserve the reasons behind them.

  4. 04

    Train

    Use verified outcomes to improve the checkpoint.

  5. 05

    Evaluate

    Test on sealed tasks before making a release decision.

Your environment.
Your learning loop.

Task data, weights, checkpoints and evaluation records stay in your environment. Formal Engines coordinates the work and keeps the evidence portable.

Read the platform docs

Bring us the task
that matters.

We’ll help define success, build the environment and establish what improvement should look like.

Discuss your task