Skip to content

Intent-Garden (S-OAK) Core

Deterministic verification engine for AI-generated C/C++ code.

We don't believe in "safe languages". We believe in safe intentions.

📖 About

Garden-Core is a Clojure engine that audits AI-generated C/C++ code through Clang AST and EDN contracts. It serves as the "last mile" between stochastic AI output and predictable code execution.

How It Works

We don't fix AI mistakes. We create an environment where invalid code physically cannot pass the build stage.

The engine eliminates hallucinations through strict division of labor:

  1. Intent Injection — AI agent receives specification from rulerom/registry
  2. Garden-Tagging — AI writes C code and marks it with [[garden:intent(...)]] tags
  3. AST Enforcement — Engine parses Clang JSON AST and validates against contracts
  4. Semantic Echo — Generates Markdown reports for human verification

🛠️ Technology Stack

Component Purpose
Clojure / Babashka Fast data processing logic without JVM overhead
Clang LibTooling Parses -ast-dump=json for honest code structure
EDN Contract format understandable by humans and machines

🔄 Semantic Anchor Lifecycle

graph TD A["<b>1. Human Definition</b><br/>Architect Input / Task Description"] --> B{AI Scan} B -->|Search| C["<b>Rule-Rom Library</b>"] B -->|Generate| D["<b>Draft Intent</b>"] C & D --> E["<b>2. Intent Formalization</b><br/>Lisp/EDN Contract"] E --> F["<i>Semantic Echo</i><br/>Auto-generated Markdown Audit"] F -->|Verification| G(("Human Approval")) G -->|Fix| E G -->|Approved| H["<b>3. Prompt Injection</b><br/>Contract + Requirements"] H --> I["<b>AI Coding & Tagging</b><br/>C/C++ Code + garden-tags"] I --> J["<b>4. AST Enforcement</b><br/>Clang JSON AST + Babashka"] J --> K{Verification} K -->|Violation| L["<b>Build Failed</b><br/>Retry Loop"] L --> H K -->|Success| M["<b>5. Certification</b><br/>Deterministic Safety Proof"] M --> N["Production Ready Code"] style A fill:#e1f5fe,stroke:#01579b style E fill:#f3e5f5,stroke:#4a148c style I fill:#fff3e0,stroke:#e65100 style J fill:#e8f5e9,stroke:#1b5e20 style M fill:#dcedc8,stroke:#33691e style L fill:#ffebee,stroke:#b71c1c

🚀 Quick Start

1. Environment Setup (Windows / Scoop)

scoop install babashka llvm

2. Running Local Audit

Assume you have a test.c file with tags:

// [[garden:intent(safe-free)]]
free(p);
p = NULL;
// [[/garden:intent]]

Step 1: Generate AST

clang -Xclang -ast-dump=json -fsyntax-only test.c > ast.json

Step 2: Run Enforcer

bb -m garden.enforcer ast.json

3. Contract Example (EDN)

{:intent :safe-free
 :entities [:ptr]
 :must-set-null true
 :description "After free(), pointer must be set to NULL"}

📋 Tagging Protocol (Garden-Tagging)

Every code block related to an Intent MUST be tagged:

// [[garden:intent(INTENT_ID)]]
void implementation_starts_here() {
    // Your logic
}
// [[/garden:intent]]

Tagging Rules

Rule Description
No Orphans Never place a tag without implementation
Exact ID INTENT_ID must match the key in .edn file
Scope Tags wrap the smallest logical unit
Vacuum Rule Code outside [[garden:intent]] is considered dead and discarded

📜 Manifest (Summary)

  1. Code is Cheap, Meaning is Expensive — value lies in intentions, not instructions
  2. AI is Gas, Formal Logic is Brakes — deterministic oversight over probabilistic execution
  3. Lisp is the Ideal Contract Language — homoiconicity for the "Constitution of the System"
  4. Semantic Cage — AI generates inside a DSL cage of safety rules
  5. Validation Instead of Hope-and-Pray — check at generation time, not after
  6. Zero-Cost Security — contract operates at metaprogramming stage
  7. Death of the "Black Box" — AI as black box implementation in a white box of intentions

📄 Documents

Document Description
AI Agent Contract Tagging rules for AI agents
EDN Registry Intent contract specifications