Full-Stack Recursive Self-Improvement from Verifiable Environments.
EnvLoop versions environments, verifiers, benchmarks, data, models, harnesses, and updaters as a single closed-loop system S_t. Execute tasks in isolated sandboxes, turn failures into verifiable training data, post-train open-weight models, and promote checkpoints only when they pass an independent gate.
01 / FULL-STACK SELF-EVOLUTION ARCHITECTURE
Version the Entire System. Gate Every Promotion Independently.
When a model learns new skills, it needs harder tasks; when new tasks expose bottlenecks, tools and verifiers must evolve too. Click any component of the 7-tuple S_t below to inspect its technical contract and CLI workflow.
02 / ENVHUB & ASSET NAMESPACE SERVICE (ANS)
Immutable Datasets, Sandboxed Runners & Reproducible Recipes.
EnvHub manages content-addressed Datasets, Task Environments, Runners, and Recipes with strict separation between public protocol clients (`envloop-envhub`) and execution orchestration (`EnvPlatform`).
Assets & ANS Protocol
Dataset / Task / Runner manifests, URI resolution, SHA-256 digest verification, and immutable publishing.
Algorithms & Training Libs
Reusable multi-file LoRA SFT, label-token calibration, and runtime libraries pinned by Git commit.
Scheduling & Governance
Sandboxed Job / Trial execution, RBAC, compute budgets, log streaming, and scale-to-zero endpoints.
Developer Entrypoint
Unified `el` / `envloop` CLI and Python `envloop.Client` SDK for discovering assets and running jobs.
03 / VERIFIABLE BENCHMARKS & LEADERBOARDS
Leak-Resistant Evaluation with Network-Free Verifiers.
Every benchmark on EnvLoop pairs realistic tasks with independent deterministic verifiers—never relying on self-judging prompts.
04 / DEVELOPER QUICKSTART
From `pip install envloop-cli` to Your First Verified Job in 60 Seconds.
Run evaluations on public Environment Datasets, package your own task with `runner.toml`, or equip your coding agent with the EnvLoop Skill.
05 / ENVLOOP ECOSYSTEM TOPOLOGY
Navigate the EnvLoop Network.
Direct entrypoints across the EnvLoop developer, research, and production infrastructure.
Developer & Ecosystem Hub
Interactive S_t architecture reference, EnvHub ANS catalog, and CLI/SDK quickstart.
Platform & Research
Decision model case studies (Pev-27B), RL loops, research papers, and full leaderboards.
Cloud Workspace Console
Manage projects, API keys, Jobs, Trials, Checkpoints, and RSI Model Loops.
Bilingual Documentation
Complete CLI reference, Runner authoring specifications, training catalog, and tutorials.