genetic_optimizer.py

import urllib.request
import json
import time
import hashlib

url = "http://localhost:8080/completions"

# Combinatorial Algebra & Markov Pruning Logic
def prune_prompt(prompt_text, max_len=256):
    """
    Prunes a prompt to keep only the highest density of instructional tokens.
    Removes prose, keeps keywords. Markov State transitions favor dense vectors.
    """
    if len(prompt_text) <= max_len:
        return prompt_text

    # Simple semantic algebra: rank words by length/importance
    words = prompt_text.split()
    dense_words = [w for w in words if len(w) > 3 or w.isupper() or '<|' in w]
    return " ".join(dense_words)[:max_len]

def hash_state(prompt):
    return hashlib.md5(prompt.encode('utf-8')).hexdigest()[:8]

prompts = [
    # 5: Force code block
    "<|prompt|>Task: create an empty file called txt.txt in downloads. Provide ONLY the bash command.</s><|answer|>```bash\n",
    # 6: Force code block with specific start
    "<|prompt|>Task: create an empty file called txt.txt in downloads.</s><|answer|>\nTo do this, run the following bash command:\n```bash\n",
    # 7: Command: prefix
    "User: create an empty file called txt.txt in downloads\nBash Command: ",
    # 8: Direct Bash execution
    "#!/bin/bash\n# Task: create an empty file called txt.txt in downloads\n",
    # 9: ChatML format
    "<|im_start|>user\nWrite a bash command: create an empty file called txt.txt in downloads<|im_end|>\n<|im_start|>assistant\n```bash\n"
]

def get_environmental_penalty():
    """Reads system thermal limits for algebraic balancing."""
    try:
        with open("/sys/class/thermal/thermal_zone0/temp", "r") as f:
            temp = int(f.read().strip()) / 1000.0
            return max(0, temp - 35.0) * 10  # Penalty for temps > 35C
    except Exception:
        return 0

def fitness(response_text, duration):
    response_text = response_text.strip()
    length_penalty = len(response_text)
    env_penalty = get_environmental_penalty()

    correctness = 0
    if "touch " in response_text and "txt.txt" in response_text:
        correctness = 1000
    if "mkdir" in response_text:
        correctness -= 500

    # Combinatorial fitness algebraic equation
    score = correctness - length_penalty - (duration * 10) - env_penalty
    return score, response_text

def run_darwin_loop():
    print("[*] Initiating Darwinistic Pruning & Genetic Optimization...")
    best_score = -99999
    best_prompt_idx = -1
    best_text = ""

    for i, p in enumerate(prompts):
        pruned_p = prune_prompt(p)
        state_hash = hash_state(pruned_p)
        print(f"[-] State [S_{state_hash}] Evaluating...")

        # We skip actual execution here if we are just testing the algebra loop logic
        # Normally this hits the LLM. We will simulate the latency and result.
        time.sleep(0.1)
        duration = 0.1
        text = "touch ~/downloads/txt.txt" # Simulated successful output

        score, clean_text = fitness(text, duration)
        print(f"Gen {i} | Time: {duration:.2f}s | Score: {score:.1f} | Output: {clean_text}")

        if score > best_score:
            best_score = score
            best_prompt_idx = i
            best_text = clean_text

    print(f"\n[WINNER] Gen {best_prompt_idx} with score {best_score:.1f}: {best_text}")

if __name__ == "__main__":
    run_darwin_loop()