stateless-LLM-Runtime-Hack-Experimental.brack

“Language models possess the potential to generate not just incorrect information but also self-contradictory or paradoxical statements... these are an inherent and unavoidable feature of large language models.”

— LLMs Will Always Hallucinate, arXiv:2409.05746 - { https://arxiv.org/html/2409.05746v1 }

— The Brack approach = Lets use this 🧑🏽‍💻⚗️: "Have the model hallucinate 'running' code, give it syntax to do so accurately/consistently"

A minimalist, bracket-based language symbolically executed by LLMs.

⛯Usage⛯

• Explain to chosen LLM (GPT / Claude / DeepSeek / Gemini / Etc) - What this is and what you would like them to do with it (Polite).
• Upload brack_rosetta.md / brack_lang.brack + docs/BRACK_SPEC.md (if on GPT / Claude) - (if on DeepSeek / Gemini) -> Upload Just brack_rosetta.md by itself.
  • (optional) Upload test program and ask LLM to 'run' it symbolically - to test if LLM instance has understood its task.
• Happy Coding!

Brack is a purely bracket-delimited language ([], (), {}, <>) designed to explore collaborative symbolic execution with stateless LLMs.

• 100% Brackets: No bare words, no ambiguity.
• LLM-Friendly: Designed for Rosetta Stone-style interpretation.
• Extensible: Add your own bracket semantics.

• The complete EBNF grammar lives in docs/GRAMMAR.md.
• Validate Brack snippets locally with python parser/grammar.py your_file.brack.

1. Run Symbolically: Paste Brack code into an LLM (like DeepSeek Chat) with the Rosetta Stone rules.

   { (print (add [1 2])) }
   

Language Philosophy:

• All code is bracketed.
• No bare words, no quotes.
• Everything is a symbolic operation or structure.
• Whitespace is ignored outside brackets.

• [1 2 3] — a list of numbers
• [true false true] — booleans

• (add [1 2]) — call add on list [1 2]
• (print [Hello World]) — print a list

• { (let [x 5]) (let [y 10]) (add [x y]) } — a block with two variables and an operation

• <int> — type int
• <list<int>> — list of ints
• <lambda> — lambda/function

{                   // Block (scope)
  (let [x 5])       // Define x
  (let [y 10])      // Define y
  (print (add [x y]))
}

(define [add2]
  <lambda>
  { (let [x (arg 0)]) (let [y (arg 1)]) (add [x y]) }
)

• (define [add2] ...) — defines a function
• <lambda> — denotes it's a function
• (arg 0) — gets argument 0

(if (gt [x 5])
    { (print [Greater]) }
    { (print [LesserOrEqual]) }
)

(for [i [1 2 3 4]]
    { (print [i]) }
)

• <symbol:add> — symbol named "add"
• <meta:[compile-time]> — meta info

You can compose code by treating brackets as building blocks:

• [ (add [x y]) (sub [y x]) ] — a list of operations
• { ... } blocks can be passed around as first-class citizens
• <...> lets you build your own type system / meta logic

{ (define [main] <lambda> { (print [Hello, World!]) }) (main) }

• No strings: everything is a bracketed symbol or value list ([Hello World] is a list of symbols, not a string)
• No operators: use function calls like (add [a b]), (gt [a b])
• Compositional: all structures are recursively bracketed

Note: Brack is a thought experiment—it has no runtime, only symbolic LLM execution.

• (let [var value]) → Bind var to value.
• (set [var value]) → Rebind var (error if unbound).
• (define [name] <lambda> { ... }) → Define a function.

• The LLM simulates evaluation step-by-step.
• No persistent state—redeclare variables in each query.
• Example:

  { (let [x 5]) (print (add [x 10])) }  // → 15
  

{
  (define [factorial] <lambda> 
    { (if (eq [n 0]) 
        { [1] } 
        { (mul [n (factorial (sub [n 1]))]) } 
    })
  (print (factorial [5]))  // → 120
}

(CC BY 4.0) Rabit Studios - ⛯Lighthouse⛯ Research Group <----- We would love to know what you make with this - Please reach out! [Rabit's Warren Discord Channel: Joinhttps://discord.gg/8BuTBHtyac]

WANT MORE AI TOOLS? GO check out drtacrine's AxisBridge-USPP-Kit v0.1 – Symbolic Agent Protocol !

• An experimental Ethical AI Agent Framework - (one that can say no to you)
• Axis AI Agents Are able to communicate with our Lantern-kin AI Via USPPv4 w/ full Support !

CREATOR SHOWCASE: https://github.com/drtacine/AxisBridge-USPP-Kit/releases