EnvLoop LabsLog inThe RSI post-training platform
EnvLoop is a post-training platform for recursive self-improvement (RSI). It trains open-weight models on your task data, tests each checkpoint on a benchmark generated for your task, and promotes a checkpoint only when independent evaluation shows that it meets your criteria, such as beating the base model.
01 / What is EnvLoop
EnvLoop is a post-training platform built around recursive self-improvement (RSI). Each round trains a model on task data, evaluates the checkpoint in sandboxes with independent verifiers, and turns the failures it finds into environments and data for the next round. Training, evaluation and inference run on one platform through the el CLI, the Python SDK and the HTTP API.
EnvLoop is for companies that need a model to get better at a specific business task, such as an agent workflow or a Jev-class decision model for their own users, and want evidence before they ship it. Companies that only need the building blocks can use the task environments, verifiers and data that EnvLoop builds for their own training and evaluation.
A loop starts from your task data and a baseline evaluation. EnvLoop trains a LoRA checkpoint on an open-weight base such as Qwen3.8-27B or GLM-5.3-Flash, generates a benchmark for the task and compares the checkpoint with the base model on it. A checkpoint that beats the base model becomes the next starting point, and you decide whether to deploy it to an OpenAI-compatible endpoint.
A decision model is post-trained on a company’s own data and answers its typed questions with a calibrated probability, so code can act on confident answers and send uncertain ones to a person. Each example says how far EnvLoop has taken it.
Would this user approve the action, which option would they pick and what may the agent share?
EnvLoop’s Pev-27B reaches 0.915 family-macro accuracy on Pev-Bench, ahead of TypeSafe’s Jev at 0.788.
Given the evidence and constraints, which credit, fixed-income, portfolio, FX or options decision is correct?
A recipe preset trains on financial-decision-bench: 53,000 source-grounded cases in English, Chinese and Japanese.
Which queue does this ticket belong to, how urgent is it and is the customer asking for a refund?
Trained on your past tickets and routing decisions, and measured on your own held-out tickets.
Can this claim be fast-tracked, or does it need an adjuster’s review?
Trained on your adjusters’ past decisions, and measured on your own held-out claims.
Which policy, if any, does this post or listing break?
Trained on your reviewers’ labels, and measured on your own held-out cases.
Does this inbound company fit our ideal customer profile, and is it ready to buy?
Trained on your won and lost deals, and measured on your own held-out leads.
02 / What we deliver
Use the platform to train a model that keeps improving on your work, or take the assets it builds for your own training and evaluation.
Train, evaluate and serve your own model, and let each round’s failures drive the next.
The reusable assets every round builds, delivered per project or licensed on their own.
03 / Recursive self-improvement
Use failed tasks to build new environments and training data. Retrain the model on that experience and independently re-evaluate it before choosing the next version.
Run a baseline on your task and identify what caused the model to fail.
Use those failures to generate task environments and training examples.
Train the next version on the new data and save its checkpoint.
Keep a version only after independent evaluation confirms it meets your criteria. You decide whether to deploy it.
The accepted version becomes the starting point for the next round. New failures guide further training while the environments and data remain reusable.
04 / Use cases
Any company can build its own Jev-class decision model: a model that answers a typed question, such as whether an action needs the user’s approval, with a calibrated probability from a single forward pass. On the Pev-Bench public test set of 720 new users, the released Pev-27B, a rank-16 LoRA on Qwen3.8-27B trained by EnvLoop Research, reaches 0.915 family-macro accuracy, ahead of TypeSafe’s Jev at 0.788 and its own base model at 0.762. Training your own uses an experimental recipe arranged with our team.
Post-train a model that answers typed questions about your users or business with a calibrated probability, like TypeSafe’s Jev.
Score agent attempts with verifiers, keep the verified ones and post-train on them, without building RL infrastructure.
Train, benchmark and keep a candidate only when it beats the base model, round after round.
LoRA SFT on open-weight models, downloadable checkpoints and OpenAI-compatible endpoints.
Run agents against Environment Datasets in sandboxes and compare verifier-scored results.
Test agents on professional work in 20 Linux desktop applications with cu-bench.
Use the task environments, verifiers and data EnvLoop builds, delivered per project or licensed.
Plan training and evaluation around your data boundaries with our team.
05 / The roadmap
Our path starts with post-training infrastructure and develops toward RSI within a domain. Extending the approach across domains is a longer-term research goal.
Bring model training, task evaluation and inference into one post-training workflow.
Develop a model for a specific business task by using failures to improve its training data and task environment.
Research how experience from one domain can inform improvement in another, with people setting the goals and budgets.
06 / Questions
Install envloop-cli from PyPI, sign in and follow the quickstart to submit your first Job with the el command. You can also install the EnvLoop Skill so that your coding agent works with EnvLoop through the CLI. The Console shows progress, logs and results, while Jobs are submitted through the CLI or API.
Yes. EnvLoop post-trains decision models on a company’s own data with the experimental Pev-27B label-token recipe on Qwen/Qwen3.8-27B, which our team enables for your project. EnvLoop’s released Pev-27B shows the approach: 0.915 family-macro accuracy on the Pev-Bench public test set, against 0.788 for TypeSafe’s Jev. A model trained on your data is measured on your own held-out set.
Yes, as a reward-driven loop built from EnvLoop’s building blocks. Your tasks run as environments in sandboxes, verifiers score each attempt, the verified attempts become training data, and a LoRA SFT Train Job post-trains the model before it is re-evaluated against its base. The reward decides evaluation and data selection; the training step itself is supervised fine-tuning, not online RL such as GRPO or PPO.
Train Jobs run LoRA SFT on open-weight models such as Qwen3.8-27B and GLM-5.3-Flash and save checkpoints that you can download. The training catalog in the documentation lists the models and recipes available to your project.
Serve a verified checkpoint from an OpenAI-compatible inference endpoint that scales to zero when idle. The documentation lists the serving profiles available for each model.
Each candidate is re-evaluated independently, by verifiers that are separate from the model and on hidden test sets. A version is kept only when it meets the criteria you set, such as beating the base model, and you decide whether to deploy it.
Yes. The task environments, verifiers, benchmarks and training data that each round builds can be delivered per project or licensed on their own. Contact the team about availability and access.
Bring business examples and a clear goal. Train through the CLI and follow progress in your workspace.