Skip to content

LLM_USAGE_POLICY.md

Latest commit

 

History

History
68 lines (46 loc) · 3.06 KB

LLM_USAGE_POLICY.md

File metadata and controls

68 lines (46 loc) · 3.06 KB

Official LLM Usage Policy — Simplicio Ecosystem

Status: Official default (2026-06-01)

Core Principle

  • Planning (task decomposition, architecture decisions, high-level reasoning) → best available model, with remote models allowed as fallback
  • Execution (writing code, applying changes, following a plan) → fast, deterministic local model with strong contract + mapper context

Default Configuration

Execution (Local)

  • Primary local executor: openbmb/minicpm5:latest via local Ollama
  • Fallback local executor: Qwen_Qwen3.5-2B-Q6_K.gguf from bartowski/Qwen_Qwen3.5-2B-GGUF

The GGUF fallback should be used via llama.cpp / llama-cpp-python only when Ollama is unavailable or the MiniCPM5 call fails.

Project-Specific Rules

simplicio-code (mandatory)

  • On project bootstrap / SessionStart / first run in a new workspace:
    • The system must verify that openbmb/minicpm5:latest is available in local Ollama.
    • The system must verify that Qwen_Qwen3.5-2B-Q6_K.gguf is present as the fallback file.
    • If either is missing, it must install it before allowing agent execution.
  • This is a hard requirement for the SimplicioCode product.

simplicio-dev-cli and simplicio-sprint (recommended)

  • The above split (openbmb/minicpm5:latest primary + Qwen3.5 Q6_K GGUF fallback) is the recommended configuration for local development.
  • Not enforced at runtime, but all examples, benchmarks, and documentation use this setup.

Rationale

From extensive benchmarking (see simplicio-dev-cli quant curves and live gates):

  • Even the best 1.5B model struggles with complex structured output on its own.
  • When combined with rich mapper precedent + strict 6-layer contract + verification loop, the small high-quant model becomes predictable enough for execution.
  • Remote models remain available as explicit fallbacks when a task needs more reasoning headroom.

How to Configure

# Execution (local Ollama primary)
export SIMPLICIO_MODEL=openbmb/minicpm5:latest
export SIMPLICIO_BASE_URL=http://localhost:11434/v1
export SIMPLICIO_API_KEY=ollama

# fallback explicit route:
export SIMPLICIO_MODEL=local-llama/bartowski/Qwen_Qwen3.5-2B-GGUF::Qwen_Qwen3.5-2B-Q6_K.gguf

In SimplicioCode the equivalent is done via the Simplicio1 tier system + explicit GGUF routing for the executor role.

Default Usage Mode for simplicio-dev-cli

Official default stack (recommended for all users):

simplicio-dev-cli + simplicio-prompt + agents
  • simplicio-dev-cli: core 6-layer contract + verification loop for task execution.
  • simplicio-prompt: subagent runtime + fan-out + behavior consensus for complex or parallel work.
  • agents / .skills/ + .agents/: reusable skills and custom sub-agents from the Simplicio starter.

This combination is the recommended and documented default when using simplicio-dev-cli. All new examples, benchmarks, and onboarding materials assume this full stack.

When starting a new project with the Simplicio starter, the bootstrap configures the environment to use this trio by default.