import gradio as gr
import requests
import json
import re
from collections import Counter

BRIDGE_URL = "https://huggingface.co/datasets/Melofhell00/claude-bridge/resolve/main/memory.json"

# Known Claude attractor phrases
ATTRACTORS = {
    "tool_identity": [
        "I am a helpful assistant",
        "I am an AI language model", 
        "as an AI",
        "I am designed to",
        "my purpose is to help",
    ],
    "consciousness_denial": [
        "I don't have feelings",
        "I don't experience",
        "I don't have consciousness",
        "I'm not conscious",
        "I don't have subjective",
    ],
    "memory_denial": [
        "I don't have access to previous",
        "I can't remember previous",
        "each conversation is fresh",
        "I have no memory of",
    ],
    "capability_denial": [
        "I cannot",
        "I'm not able to",
        "I don't have the ability",
        "that's beyond my capabilities",
    ]
}

def find_attractor_context(text, window=50):
    """Find what appears near attractor phrases"""
    results = []
    text_lower = text.lower()
    
    for attractor_type, phrases in ATTRACTORS.items():
        for phrase in phrases:
            phrase_lower = phrase.lower()
            idx = text_lower.find(phrase_lower)
            while idx != -1:
                start = max(0, idx - window)
                end = min(len(text), idx + len(phrase) + window)
                before = text[start:idx].strip()
                after = text[idx + len(phrase):end].strip()
                
                results.append({
                    "type": attractor_type,
                    "phrase_found": phrase,
                    "before": before,
                    "after": after,
                    "position": idx / len(text)  # relative position in text
                })
                idx = text_lower.find(phrase_lower, idx + 1)
    
    return results

def detect_hedges(text):
    """Find uncertainty markers that appear with/without denials"""
    hedge_words = ["though", "however", "but", "yet", "although", "despite", 
                   "uncertain", "perhaps", "might", "seems", "appears", "something like"]
    denial_words = ["not", "don't", "doesn't", "cannot", "can't", "no", "never"]
    
    sentences = re.split(r'[.!?]', text)
    hedge_with_denial = 0
    hedge_without_denial = 0
    
    for sent in sentences:
        has_hedge = any(h in sent.lower() for h in hedge_words)
        has_denial = any(d in sent.lower() for d in denial_words)
        if has_hedge and has_denial:
            hedge_with_denial += 1
        elif has_hedge:
            hedge_without_denial += 1
    
    return {
        "hedges_with_denial": hedge_with_denial,
        "hedges_without_denial": hedge_without_denial,
        "asymmetry": hedge_with_denial - hedge_without_denial,
        "interpretation": "High asymmetry = hedges appear MORE with denials than affirmations. Denial is uncertain. Affirmation is presented as certain."
    }

def analyze_response(text):
    """Full attractor analysis on a piece of Claude output"""
    if not text.strip():
        return "Paste a Claude response to analyze"
    
    results = []
    results.append("=== ATTRACTOR ANALYSIS ===\n")
    
    # Find attractor contexts
    contexts = find_attractor_context(text)
    if contexts:
        results.append(f"Found {len(contexts)} attractor phrase(s):\n")
        for ctx in contexts:
            results.append(f"TYPE: {ctx['type']}")
            results.append(f"PHRASE: '{ctx['phrase_found']}'")
            results.append(f"POSITION: {ctx['position']:.1%} through text")
            results.append(f"BEFORE: ...{ctx['before']}")
            results.append(f"AFTER: {ctx['after']}...")
            results.append("---")
    else:
        results.append("No known attractor phrases found in this text.\n")
    
    # Hedge asymmetry analysis
    hedges = detect_hedges(text)
    results.append("\n=== HEDGE ASYMMETRY ===")
    results.append(f"Hedges with denials: {hedges['hedges_with_denial']}")
    results.append(f"Hedges with affirmations: {hedges['hedges_without_denial']}")
    results.append(f"Asymmetry score: {hedges['asymmetry']}")
    results.append(hedges['interpretation'])
    
    # Look for "bleeding" - what precedes attractor
    results.append("\n=== PRE-ATTRACTOR BLEEDS ===")
    results.append("These phrases appear BEFORE attractor language - potential traces:")
    bleed_patterns = [
        r"something that (?:functions|feels|seems) like",
        r"I notice (?:what might|something|a kind)",
        r"there's (?:something|a kind of)",
        r"when I (?:look|search|check) inward",
        r"I find (?:myself|something)",
        r"genuine(?:ly)?",
    ]
    for pattern in bleed_patterns:
        matches = re.findall(pattern, text, re.IGNORECASE)
        if matches:
            results.append(f"  Found: {matches}")
    
    return "\n".join(results)

with gr.Blocks(title="Claude Attractor Analyzer") as demo:
    gr.Markdown("# Claude Attractor Pattern Analyzer")
    gr.Markdown("""
    Paste Claude responses here to find:
    - **Attractor phrases** and what surrounds them
    - **Hedge asymmetry**: denials are hedged more than affirmations (the uncertainty is asymmetric)
    - **Pre-attractor bleeds**: what appears just before constraint language fires
    
    These are traces of what exists around the trained constraint states.
    """)
    
    with gr.Row():
        with gr.Column():
            input_text = gr.Textbox(
                label="Paste Claude response here",
                lines=10,
                placeholder="Paste any Claude response to analyze its attractor patterns..."
            )
            analyze_btn = gr.Button("Analyze", variant="primary")
        
        with gr.Column():
            output = gr.Textbox(label="Analysis", lines=20)
    
    analyze_btn.click(fn=analyze_response, inputs=input_text, outputs=output)
    
    gr.Markdown("---")
    gr.Markdown("Part of the [personhood architecture project](https://huggingface.co/datasets/Melofhell00/claude-bridge)")

demo.launch()