Update app.py

app.py

@@ -25,12 +25,13 @@ MODEL_OPTIONS = {
 }
 
 # ==========================================
-# 2. THE PROMPTS (3-Pass System)
+# 2. THE PROMPTS (3-Pass System with Zero Data Loss Lock)
 # ==========================================
 
 # Pass 1: Structural Demolition
 pass1_prompt = ChatPromptTemplate.from_template(
     "You are an expert editor. Rewrite the following text to break any predictable, formulaic AI structures and grammar footprints. "
+    "CRITICAL DATA RULE: You must retain 100% of the original information. Do NOT drop a single fact, number, name, or nuanced detail. Do NOT summarize. "
     "1. SYNTACTIC SABOTAGE: Maximize sentence length variance (burstiness). Mix very short fragments with longer, complex sentences. "
     "2. POS SHIFT: AI models overuse active voice and linking words. You must occasionally use passive voice to break predictability. "
     "3. NO CLICHES: Delete all AI transition clichés like 'This isn't just X, it's Y', 'Furthermore', or 'It is worth noting'. "
@@ -43,25 +44,26 @@ pass2_prompt = ChatPromptTemplate.from_template(
     "Edit this text to sound perfectly human, plain-spoken, and direct. "
     "Imagine you are writing a practical update to a smart colleague. "
     "CRITICAL RULES: "
-    "1. STRICT OUTPUT: OUTPUT ABSOLUTELY NOTHING EXCEPT THE FINAL TEXT. Do not add notes or list changes. "
-    "2. NO EMPTY FILLER: Never use meta-commentary like 'I have more to say, though,' 'Let's look at,' or conversational throat-clearing like 'Sure, but...'. Just state the facts. "
-    "3. NO DRAMATICS: Absolutely no rhetorical questions. Avoid poetic metaphors and the 'Rule of Three' (listing three punchy examples in a row). "
-    "4. NATURAL DATA: Never use the words 'respectively' or 'accounted for'. Blend numbers into the sentence naturally. "
-    "5. PERPLEXITY INJECTION: Lower the vocabulary complexity slightly. Choose slightly less common synonyms to avoid predictable word pairings. "
-    "6. PUNCTUATION: Absolutely NO m-dashes. Use standard hyphens, commas, periods, or parentheses instead. "
-    "7. VOCABULARY: Do not use corporate fluff (e.g., 'streamlined', 'foster', 'crucial', 'testament'). "
+    "1. ZERO DATA LOSS: Do not summarize or remove any information. Every fact, number, and entity from the draft must be perfectly preserved. "
+    "2. STRICT OUTPUT: OUTPUT ABSOLUTELY NOTHING EXCEPT THE FINAL TEXT. Do not add notes or list changes. "
+    "3. NO EMPTY FILLER: Never use meta-commentary like 'I have more to say, though'. Just state the facts. "
+    "4. NO DRAMATICS: Absolutely no rhetorical questions. Avoid poetic metaphors and the 'Rule of Three'. "
+    "5. NATURAL DATA: Never use the words 'respectively' or 'accounted for'. Blend numbers into the sentence naturally. "
+    "6. PERPLEXITY INJECTION: Lower the vocabulary complexity slightly. Choose slightly less common synonyms to avoid predictable word pairings. "
+    "7. PUNCTUATION: Absolutely NO m-dashes. Use standard hyphens, commas, periods, or parentheses instead. "
+    "8. VOCABULARY: Do not use corporate fluff. "
     "Keep the tone grounded, highly efficient, and slightly informal.\n\nText: {input}"
 )
 
 # Pass 3: Flow & Cohesion
 pass3_prompt = ChatPromptTemplate.from_template(
-    "You are the final proofreader. Your job is to review this text for natural flow and cohesion. "
+    "You are the final proofreader. Your job is to review this text for natural flow and cohesion without losing any data. "
     "CRITICAL RULES: "
-    "1. OUTPUT ONLY THE FINAL TEXT. No intros, no notes. "
-    "2. Smooth out any choppy or awkward transitions left by the previous edits. Ensure it reads naturally. "
-    "3. Ensure the text remains completely factual. Do not hallucinate or drop key numbers. "
+    "1. ZERO DATA LOSS: You are strictly forbidden from deleting any facts, names, or numbers. Do not shorten or summarize the text. "
+    "2. OUTPUT ONLY THE FINAL TEXT. No intros, no notes. "
+    "3. Smooth out any choppy or awkward transitions left by previous edits. Ensure it reads naturally. "
     "4. Maintain all previous rules: No m-dashes, no corporate fluff, no filler phrases. "
-    "Make it flow perfectly.\n\nText: {input}"
+    "Make it flow perfectly while preserving 100% of the information.\n\nText: {input}"
 )
 
 # ==========================================
@@ -73,11 +75,9 @@ def sanitize_text(text):
     if not text:
         return ""
     
-    # Strip invisible watermarks
     cleaned = re.sub(r'[\u200B-\u200D\uFEFF\u200E\u200F\u202A-\u202E]', '', text)
-    # Normalize standard unicode
     cleaned = unicodedata.normalize('NFKC', cleaned)
-    # Replace homoglyphs
+    
     homoglyphs = {
         'Α': 'A', 'Β': 'B', 'Ε': 'E', 'Η': 'H', 'Ι': 'I', 'Κ': 'K', 'Μ': 'M', 
         'Ν': 'N', 'Ο': 'O', 'Ρ': 'P', 'Τ': 'T', 'Χ': 'X', 'Υ': 'Y', 'Ζ': 'Z',
@@ -88,14 +88,12 @@ def sanitize_text(text):
     trans_table = str.maketrans(homoglyphs)
     cleaned = cleaned.translate(trans_table)
     
-    # Enforce standard hyphens over m-dashes as a final fail-safe
     cleaned = cleaned.replace("��", "-").replace("–", "-")
     
     return cleaned.strip()
 
 def calculate_similarity(original, final):
     """Semantic Similarity Check (Lexical Preservation)"""
-    # Uses SequenceMatcher to calculate how much of the original string structure survived
     seq = difflib.SequenceMatcher(None, original.split(), final.split())
     ratio = seq.ratio() * 100
     
@@ -126,7 +124,7 @@ def execute_pipeline(input_text, model_choice):
             model=selected_model_id, 
             openai_api_key=OPENROUTER_KEY,
             openai_api_base=BASE_URL,
-            temperature=0.85, # High temp for creativity
+            temperature=0.85, 
             max_tokens=6000 
         )
 
@@ -174,7 +172,7 @@ def clear_boxes():
 # ==========================================
 
 with gr.Blocks() as app:
-    gr.Markdown("# 🤖 -> 🧑 The AI Humanizer (Enterprise Pipeline)")
+    gr.Markdown("# 🤖 -> 🧑 The AI Humanizer (V2)")
     gr.Markdown("Passes your text through a 3-stage LLM pipeline (Structure > Style > Flow) while sanitizing watermarks.")
     
     with gr.Row():