Formal EnginesBring us a task
A monumental stone portal opening onto a still alpine lake

Post-training systems

Own how yourmodels learn

Formal Engines expresses a task as an executable environment, then uses its outcomes for training, evaluation, and release decisions.

Explore the system

The training environment is the specification.

Formal Engines represents a task as executable state, tools, permitted actions, constraints, and success checks. Training and evaluation run against the same versioned definition.

Each rollout returns a score together with the trace and reasons behind it. Those outcomes can train the next checkpoint or remain sealed for evaluation.

Task world
Inputs, tools, state
Verifier
Checks, judges, outcomes
Evidence
Score, trace, version
Fig. 1 / Executable task definitionOne specification, two uses
research_agent/ answer_with_sources
environment.v4
01

task answer_with_sources

02

state query, evidence_set, response

03

tools search, retrieve, cite

04

constraint every_claim.has_source

05

success answer.correct && citations.resolve

spec hash8f3a…72bc

Every result remains bound to the environment version that produced it.

Training rolloutrollout_1842
Checks
6 / 7
Reward
0.86
Trace
recorded
Learning signal
Held-out evaluationcheckpoint_042
Task set
sealed
Evidence
complete
Decision
recorded
Release evidence
Illustrative record. The same executable definition produces a training signal and the evidence used for a later release decision.

One protocol from rollout to release.

Every checkpoint is tested against the environment version that produced its training signal. Changes can be traced to the model, the task definition, or the verifier.

Fig. 2 / Published training lossSmolLM2-135M · 20,288 steps
train/losssupervised fine-tuning
Run
Distil-PII
Model
SmolLM2-135M
Status
complete
Training and evaluation loss from Distil-PII-SmolLM2-135M-Instruct. Replotted from the published training log; not a Formal Engines result.
  1. 01

    Specify

    Encode tasks, tools, constraints, and the evidence that counts as success.

    Versioned environment
  2. 02

    Generate

    Run diverse model rollouts inside the same controlled environment.

    Rollout traces
  3. 03

    Measure

    Combine deterministic checks with calibrated model judges.

    Scores and reasons
  4. 04

    Improve

    Turn verified outcomes into the learning signal for post-training.

    New checkpoint
  5. 05

    Release

    Test the checkpoint on sealed tasks and record the decision.

    Release record

Your task data, weights, checkpoints, and evaluation records stay in your environment.

The entire evidence trail remains portable.

Bring us the task your model must master.

We will help encode it, train against it, and build the evidence required for release.

Discuss your task