# EnvLoop Developer & Open-Source Ecosystem Portal (envloop.net)

> EnvLoop is a full-stack Recursive Self-Improvement (RSI) post-training platform and open environment ecosystem. System state is versioned as S_t = (E_t, V_t, B_t, D_t, M_t, H_t, U_t), where each round executes tasks in verifiable sandboxes, diagnoses failures, synthesizes environments and data, post-trains open-weight checkpoints, and promotes candidates only when they pass an independent evaluation gate G_t.

## Core Ecosystem Entrypoints
- Developer & Open Ecosystem Hub: https://envloop.net
- Main Platform & Research Portal: https://envloop.ai
- Cloud Console: https://console.envloop.ai
- Developer Documentation: https://docs.envloop.ai
- Evaluation Reports Portal: https://bench.envloop.ai
- Ecosystem JSON Manifest: https://envloop.net/api/manifest.json

## Quickstart (CLI & Python SDK)
- Install CLI & SDK: `python -m pip install --upgrade envloop-cli`
- Configure Auth: `envloop auth config && envloop auth status --check`
- Install Coding Agent Skill: `npx skills@latest add https://docs.envloop.ai --skill envloop`
- List Environment Datasets: `el dataset list --type environment --limit 10`
- Submit Dataset Job: `el job submit --dataset envloop/terminal-bench-2-1-lite@1.0.2`
- Submit Custom Task: `el job submit --source ./my-task --wait`

## Full-Stack RSI System Equation
- System Tuple: S_t = (E_t, V_t, B_t, D_t, M_t, H_t, U_t)
  - E_t (Environment): Executable, resettable sandboxes (Docker, E2B, Harbor, Desktop Xvfb)
  - V_t (Verifier): Network-free deterministic rubrics, file-state checks, and calibrated graders
  - B_t (Benchmark): Leak-resistant evaluation splits with held-out and dual-LLM rendered test sets
  - D_t (Data & Curriculum): Verifier-filtered trajectories, rejection sampling, and failure-driven synthesis
  - M_t (Model): Open-weight base models + LoRA SFT checkpoints (Qwen3.8-27B, GLM-5.3-Flash, Pev-27B)
  - H_t (Harness): Tool orchestration, memory management, computer-use action loops, and recovery
  - U_t (Updater): Candidate proposer optimizing expected capability gain vs. compute budget
  - G_t (Independent Gate): Frozen acceptance gate; promotes C_{t+1} to S_{t+1} only when beating baseline

## Public Benchmarks & Assets
- Pev-Bench: 3,885 Harbor tasks across 7 personal-agent decision families, 720 new test users. Pev-27B achieves 0.915 family-macro accuracy (Brier 0.125, ECE 0.065).
- cu-bench: 290 variants of 165 professional tasks across 20 Linux desktop applications (KiCad, FreeCAD, Blender, Godot, QGIS, etc.), GUI + Code tracks.
- Recursive-Play: 150 AI-generated interactive environments (900 levels) with constructive reachability witnesses and deterministic state verifiers.