Rework pipeline (dual models + Tavily + LLM final verdict), SaaS UI, fix torchvision logs

.streamlit/config.toml

@@ -1,13 +1,14 @@
 [theme]
 base = "light"
 primaryColor = "#6366f1"
-backgroundColor = "#f5f7ff"
+backgroundColor = "#f6f8fc"
 secondaryBackgroundColor = "#ffffff"
 textColor = "#1e293b"
 font = "sans serif"
 
 [server]
 headless = true
+fileWatcherType = "none"
 
 [browser]
 gatherUsageStats = false

README.md

@@ -6,59 +6,44 @@ colorTo: purple
 sdk: docker
 app_port: 8501
 pinned: false
-short_description: The Two-Brain Misinformation Detector
+short_description: Fact-check any claim in seconds
 license: mit
 ---
 
-# 🧠 Credo AI: The Two-Brain Misinformation Detector
+# 🧠 Credo — Truth, verified.
 
-Welcome to the live demo of **Credo AI** — a "Two-Brain" AI system that provides rapid and in-depth analysis of news articles and text, built for the Hack2Skill Hackathon 2025.
+**Credo** is a consumer-friendly fact-checking app. Paste a claim, headline, article, or a link — get a clear **Real / Fake / Unclear** verdict with confidence, key takeaways, and sources.
 
 **Live Space:** https://huggingface.co/spaces/Arko007/Credo_AI
 
-## Key Features
+## How the pipeline works
 
-- **Live URL & Text Analysis:** Paste any text or a URL to a news article for an instant analysis.
-- **Dual-AI Verdict:** A fast, high-confidence FAKE/REAL verdict from a specialist model, plus a nuanced check from an expert model.
-- **Groq-Powered Explanations:** Clear, conversational explanations of the findings, powered by Groq LPU inference (`qwen/qwen3.6-27b`, fallback `openai/gpt-oss-120b`).
-- **Web Cross-Reference:** Optional Tavily-based search to cross-check claims against live sources.
-- **Analysis History:** All queries are saved in your session and reviewable on the History page.
-- **Export:** Download each analysis as a JSON report.
+1. **Content extraction** — direct text, URL scraping (Beautiful Soup + lxml), or file upload.
+2. **AI cross-check** — the content is analyzed by two specialized classifiers (a general-purpose one and a political one).
+3. **Web cross-reference** — Tavily searches live web sources for related evidence.
+4. **Final verdict** — an LLM on Groq LPU hardware (qwen/qwen3.6-27b, fallback gpt-oss-120b / gpt-oss-20b) reviews the classifiers and the web evidence and issues the final verdict, explanation, and key points. The LLM is the final judge and its verdict overrides the classifiers.
 
-## Models
+Users only ever see the final verdict — internal model details are never exposed.
 
-| Brain | Model | Base | Role |
-|---|---|---|---|
-| Brain 1 | `Arko007/fake-news-liar-political` | RoBERTa-base | Political (US-centric) FAKE/REAL |
-| Brain 2 | `Arko007/fact-check1-v3-final` | DeBERTa-v3-large | General FAKE/REAL specialist |
+## Features
 
-The app automatically routes political text to Brain 1 and everything else to Brain 2.
+- Clean, lightweight SaaS-style UI
+- Verdicts with confidence levels (High / Medium / Low) and a confidence bar
+- Key takeaways and source links with every result
+- Full analysis history and one-click JSON report export
+- Graceful fallback to model-only analysis if the LLM is unavailable
 
 ## Setup
 
 ### Secrets (Space Settings → Variables and Secrets)
 
-- `GROQ_API_KEY` — required for AI-powered explanations (https://console.groq.com)
-- `TAVILY_API_KEY` — optional, enables live web cross-referencing (https://tavily.com)
+- `GROQ_API_KEY` — required for final LLM verdicts and explanations (https://console.groq.com)
+- `TAVILY_API_KEY` — enables live web cross-referencing (https://tavily.com)
 
-The app works in Basic Mode without any keys, using built-in summaries and fallback analysis.
+The app works without keys using model-only fallback analysis.
 
-## ⚠️ Important Limitations
+## ⚠️ Honest limitations
 
-This is a hackathon prototype. The models are expert **pattern detectors**, not universal truth engines:
-
-1. **Pattern Recognition vs. World Knowledge** — simple, declarative but factually false statements (e.g., "The sun rises in the West") can be missed because they don't match fake-news stylistic patterns.
-2. **Limited Training Data & Domain Bias** — models were trained mostly on Western political news; accuracy drops on science, finance, health, and non-Western news contexts.
-3. **Future Work** — knowledge-graph grounding, multi-domain training, and global datasets.
-
-Always treat results as assistive, not authoritative.
-
-## Tech Stack
-
-- **Frontend:** Streamlit + custom CSS
-- **Models:** Hugging Face Transformers (PyTorch), RoBERTa & DeBERTa-v3
-- **Explanations:** Groq Cloud API
-- **Search:** Tavily
-- **Scraping:** Beautiful Soup + lxml
+Credo is a helpful first filter, not a perfect truth machine. AI pattern detection works best on news-style text and can miss false claims that look completely normal. Always verify important information with trusted sources before sharing.
 
 Built by the Data Dragons 🐉 for the Hack2Skill Hackathon 2025.

requirements.txt

@@ -1,6 +1,7 @@
 streamlit>=1.46.0
 transformers>=4.57.0
 torch>=2.5.0
+torchvision>=0.21.0
 groq>=0.14.0
 tavily-python>=0.5.0
 pandas>=2.2.0

src/streamlit_app.py

@@ -1,7 +1,6 @@
 import os
-import random
-import time
 import json
+import time
 import requests
 from datetime import datetime
 from bs4 import BeautifulSoup
@@ -26,8 +25,8 @@ except Exception:
     TAVILY_CLIENT = None
     TAVILY_AVAILABLE = False
 
-BRAIN_1_MODEL = "Arko007/fake-news-liar-political"
-BRAIN_2_MODEL = "Arko007/fact-check1-v3-final"
+MODEL_POLITICAL = "Arko007/fake-news-liar-political"
+MODEL_GENERAL = "Arko007/fact-check1-v3-final"
 
 GROQ_MODELS = [
     "qwen/qwen3.6-27b",
@@ -35,8 +34,10 @@ GROQ_MODELS = [
     "openai/gpt-oss-20b",
 ]
 
+MAX_ANALYSIS_CHARS = 2000
+
 st.set_page_config(
-    page_title="Credo AI | Truth Detection Platform",
+    page_title="Credo | Fact Check",
     page_icon="🧠",
     layout="wide",
     initial_sidebar_state="expanded",
@@ -44,19 +45,18 @@ st.set_page_config(
 
 st.markdown("""
 <style>
-@import url('https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700;800&display=swap');
-
 html, body, [class*="css"], [class*="st-"], .stApp {
-    font-family: 'Inter', -apple-system, BlinkMacSystemFont, 'Segoe UI', sans-serif;
+    font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto,
+                 "Helvetica Neue", Arial, "Noto Sans", sans-serif;
 }
 
 .stApp {
-    background: linear-gradient(135deg, #f5f7ff 0%, #eef1fb 50%, #e9eefa 100%);
+    background: linear-gradient(135deg, #f6f8fc 0%, #eef2f9 55%, #eaf0fa 100%);
     color: #1e293b;
 }
 
 h1, h2, h3, h4 {
-    color: #1e293b;
+    color: #0f172a;
     letter-spacing: -0.01em;
 }
 
@@ -68,144 +68,222 @@ h1, h2, h3, h4 {
     color: #475569;
 }
 
-.hero-container {
-    background: linear-gradient(135deg, #ffffff 0%, #f8faff 100%);
-    border: 1px solid #e5e9f2;
-    border-radius: 20px;
-    padding: 2.5rem 2rem;
+.brand {
     text-align: center;
-    box-shadow: 0 10px 30px rgba(99, 102, 241, 0.08);
-    margin-bottom: 1.5rem;
+    padding: 1.25rem 0 0.5rem 0;
+}
+.brand-logo {
+    font-size: 2.2rem;
+    line-height: 1;
 }
-.main-title {
-    background: linear-gradient(90deg, #4f46e5, #7c3aed, #4f46e5);
-    background-size: 200% auto;
+.brand-name {
+    font-size: 1.4rem;
+    font-weight: 800;
+    margin: 0.4rem 0 0.1rem 0;
+    background: linear-gradient(90deg, #4f46e5, #7c3aed);
     -webkit-background-clip: text;
     -webkit-text-fill-color: transparent;
     background-clip: text;
-    animation: shimmer 6s linear infinite;
-    font-size: 2.6rem;
+}
+.brand-tag {
+    color: #94a3b8;
+    font-size: 0.85rem;
+    margin: 0;
+}
+
+.hero {
+    text-align: center;
+    padding: 2.5rem 1rem 1.5rem 1rem;
+    max-width: 860px;
+    margin: 0 auto;
+}
+.hero-title {
+    font-size: 2.5rem;
     font-weight: 800;
+    color: #0f172a;
+    margin: 0 0 0.6rem 0;
     letter-spacing: -0.02em;
-    margin: 0 0 0.75rem 0;
 }
-@keyframes shimmer {
-    to { background-position: 200% center; }
+.hero-title span {
+    background: linear-gradient(90deg, #4f46e5, #7c3aed);
+    -webkit-background-clip: text;
+    -webkit-text-fill-color: transparent;
+    background-clip: text;
 }
-.hero-subtitle {
+.hero-sub {
     color: #64748b;
     font-size: 1.1rem;
-    max-width: 760px;
-    margin: 0 auto 1.5rem auto;
-    line-height: 1.7;
+    margin: 0 0 1.75rem 0;
 }
-.metrics-container {
+
+.trust-row {
     display: flex;
-    gap: 1rem;
+    gap: 1.5rem;
     justify-content: center;
     flex-wrap: wrap;
+    margin-bottom: 2rem;
 }
-.metric-card {
+.trust-chip {
     background: #ffffff;
     border: 1px solid #e5e9f2;
-    border-radius: 16px;
-    padding: 1rem 1.75rem;
-    min-width: 130px;
-    box-shadow: 0 6px 16px rgba(99, 102, 241, 0.06);
-    transition: transform 0.2s ease, box-shadow 0.2s ease;
-}
-.metric-card:hover {
-    transform: translateY(-3px);
-    box-shadow: 0 10px 24px rgba(99, 102, 241, 0.14);
-}
-.metric-value {
-    display: block;
-    font-size: 1.9rem;
-    font-weight: 800;
-    background: linear-gradient(90deg, #4f46e5, #7c3aed);
-    -webkit-background-clip: text;
-    -webkit-text-fill-color: transparent;
-    background-clip: text;
-}
-.metric-label {
-    display: block;
-    color: #94a3b8;
-    font-size: 0.78rem;
-    text-transform: uppercase;
-    letter-spacing: 0.08em;
-    margin-top: 0.2rem;
+    border-radius: 999px;
+    padding: 0.4rem 1.1rem;
+    color: #475569;
+    font-size: 0.85rem;
+    font-weight: 600;
+    box-shadow: 0 2px 8px rgba(99, 102, 241, 0.05);
 }
 
-.summary-box {
+.result-card {
     background: #ffffff;
     border: 1px solid #e5e9f2;
-    border-left: 5px solid #6366f1;
-    border-radius: 14px;
-    padding: 1.25rem 1.5rem;
-    color: #334155;
-    line-height: 1.7;
-    font-size: 1.02rem;
-    box-shadow: 0 6px 18px rgba(99, 102, 241, 0.06);
+    border-radius: 18px;
+    padding: 1.75rem 2rem;
+    box-shadow: 0 10px 30px rgba(99, 102, 241, 0.08);
+    margin-bottom: 1.25rem;
 }
 
-.verdict-container {
-    border-radius: 18px;
-    padding: 2rem 1.5rem;
-    text-align: center;
-    box-shadow: 0 10px 26px rgba(15, 23, 42, 0.12);
+.verdict-head {
+    display: flex;
+    align-items: center;
+    gap: 1.25rem;
+    flex-wrap: wrap;
+    margin-bottom: 1.25rem;
 }
-.verdict-fake {
-    background: linear-gradient(135deg, #ef4444, #dc2626);
+.verdict-chip {
+    border-radius: 999px;
+    padding: 0.65rem 1.6rem;
+    font-size: 1.25rem;
+    font-weight: 800;
+    letter-spacing: 0.04em;
+    color: #ffffff;
+    box-shadow: 0 8px 20px rgba(15, 23, 42, 0.15);
 }
 .verdict-real {
     background: linear-gradient(135deg, #10b981, #059669);
 }
-.verdict-uncertain {
+.verdict-fake {
+    background: linear-gradient(135deg, #ef4444, #dc2626);
+}
+.verdict-unknown {
     background: linear-gradient(135deg, #f59e0b, #d97706);
 }
-.verdict-text {
-    font-size: 2.2rem;
-    font-weight: 800;
-    letter-spacing: 0.12em;
-    color: #ffffff;
-    text-shadow: 0 2px 8px rgba(0, 0, 0, 0.18);
+.conf-line {
+    display: flex;
+    align-items: center;
+    gap: 0.75rem;
+    color: #64748b;
+    font-size: 0.95rem;
+    font-weight: 600;
+}
+.conf-bar {
+    width: 150px;
+    height: 8px;
+    border-radius: 999px;
+    background: #eef1f7;
+    overflow: hidden;
+}
+.conf-fill {
+    height: 100%;
+    border-radius: 999px;
+    background: linear-gradient(90deg, #4f46e5, #7c3aed);
 }
 
-.glass-card {
-    background: #ffffff;
-    border: 1px solid #e5e9f2;
-    border-radius: 18px;
-    padding: 1.75rem;
-    box-shadow: 0 8px 24px rgba(99, 102, 241, 0.07);
+.explanation {
+    color: #334155;
+    font-size: 1.05rem;
+    line-height: 1.75;
+    margin: 0 0 1.5rem 0;
 }
 
-.footer-enhanced {
-    margin-top: 2.5rem;
-    padding: 1.5rem 2rem;
+.sub-block {
+    margin-bottom: 1.25rem;
+}
+.sub-title {
+    font-size: 0.8rem;
+    font-weight: 700;
+    text-transform: uppercase;
+    letter-spacing: 0.08em;
+    color: #94a3b8;
+    margin: 0 0 0.6rem 0;
+}
+.key-point {
+    display: flex;
+    gap: 0.6rem;
+    align-items: flex-start;
+    color: #334155;
+    padding: 0.45rem 0;
+    border-bottom: 1px dashed #eef1f7;
+    font-size: 0.98rem;
+}
+.key-point:last-child {
+    border-bottom: none;
+}
+.key-point .dot {
+    color: #6366f1;
+    font-weight: 800;
+    line-height: 1.4;
+}
+.source-link {
+    display: block;
+    color: #4f46e5;
+    text-decoration: none;
+    font-weight: 600;
+    font-size: 0.95rem;
+    padding: 0.4rem 0;
+    border-bottom: 1px dashed #eef1f7;
+}
+.source-link:last-child {
+    border-bottom: none;
+}
+.source-link:hover {
+    color: #7c3aed;
+}
+
+.steps-grid {
+    display: flex;
+    gap: 1.25rem;
+    flex-wrap: wrap;
+    margin: 1.5rem 0;
+}
+.step-card {
+    flex: 1;
+    min-width: 220px;
     background: #ffffff;
     border: 1px solid #e5e9f2;
     border-radius: 16px;
-    text-align: center;
-    color: #64748b;
-    box-shadow: 0 6px 18px rgba(99, 102, 241, 0.06);
+    padding: 1.5rem;
+    box-shadow: 0 8px 22px rgba(99, 102, 241, 0.06);
 }
-.footer-features {
-    display: flex;
-    gap: 1.25rem;
+.step-num {
+    display: inline-flex;
+    align-items: center;
     justify-content: center;
-    flex-wrap: wrap;
+    width: 34px;
+    height: 34px;
+    border-radius: 10px;
+    background: linear-gradient(135deg, #4f46e5, #7c3aed);
+    color: #ffffff;
+    font-weight: 800;
     margin-bottom: 0.75rem;
 }
-.footer-feature {
-    display: flex;
-    align-items: center;
-    gap: 0.4rem;
-    font-size: 0.9rem;
-    font-weight: 600;
-    color: #475569;
+.step-title {
+    font-weight: 700;
+    color: #0f172a;
+    margin: 0 0 0.4rem 0;
 }
-.footer-feature-icon {
-    font-size: 1.1rem;
+.step-text {
+    color: #64748b;
+    font-size: 0.95rem;
+    line-height: 1.6;
+    margin: 0;
+}
+
+.footer {
+    text-align: center;
+    color: #94a3b8;
+    font-size: 0.85rem;
+    padding: 2rem 1rem 1rem 1rem;
 }
 
 .stButton > button, .stDownloadButton > button {
@@ -231,10 +309,9 @@ h1, h2, h3, h4 {
 .stButton > button[kind="secondary"]:hover {
     border-color: #6366f1;
     background: #f5f6ff;
-    transform: translateY(-2px);
 }
 .stButton > button:disabled, .stDownloadButton > button:disabled {
-    opacity: 0.55;
+    opacity: 0.5;
     box-shadow: none;
 }
 
@@ -294,10 +371,6 @@ h1, h2, h3, h4 {
     box-shadow: 0 6px 16px rgba(99, 102, 241, 0.06);
 }
 
-.stProgress > div > div > div {
-    background: linear-gradient(90deg, #4f46e5, #7c3aed);
-}
-
 a {
     color: #6366f1;
 }
@@ -306,625 +379,563 @@ a {
 
 
 @st.cache_resource(show_spinner=False)
-def load_ai_models():
+def load_engines():
     try:
-        with st.status("🔧 Loading AI models...", expanded=True) as status:
-            st.write("🧠 Initializing Brain 1 (LIAR Political)...")
-            classifier_b1 = pipeline(
+        with st.status("🔧 Setting up analysis engines...", expanded=True) as status:
+            st.write("📚 Loading the fact-checking models...")
+            clf_political = pipeline(
                 "text-classification",
-                model=BRAIN_1_MODEL,
-                tokenizer=BRAIN_1_MODEL,
+                model=MODEL_POLITICAL,
+                tokenizer=MODEL_POLITICAL,
                 return_all_scores=False,
                 device=0 if torch.cuda.is_available() else -1,
             )
-            st.write("🎯 Initializing Brain 2 (General)...")
-            classifier_b2 = pipeline(
+            st.write("⚡ Preparing the general model...")
+            clf_general = pipeline(
                 "text-classification",
-                model=BRAIN_2_MODEL,
+                model=MODEL_GENERAL,
                 return_all_scores=False,
                 device=0 if torch.cuda.is_available() else -1,
             )
-            status.update(label="✅ AI models loaded successfully!", state="complete")
-            return classifier_b1, classifier_b2
+            status.update(label="✅ Analysis engines ready!", state="complete")
+            return clf_political, clf_general
     except Exception as e:
-        st.error(f"🔴 Model loading failed: {str(e)}")
+        st.error(f"🔴 Setup failed: {str(e)}")
         return None, None
 
 
-def normalize_brain1_label(raw_label):
-    label = str(raw_label).upper()
-    if label in ("FAKE", "REAL"):
-        return label
-    if label == "LABEL_0":
-        return "FAKE"
-    if label == "LABEL_1":
-        return "REAL"
-    return label
-
-
-def normalize_brain2_label(raw_label):
-    label = str(raw_label).upper()
-    if label in ("FAKE", "REAL"):
-        return label
-    if label == "LABEL_1":
-        return "FAKE"
-    if label == "LABEL_0":
-        return "REAL"
-    return label
+def classify_with_models(text, clf_political, clf_general):
+    results = {"political": None, "general": None}
+    if clf_political is not None:
+        try:
+            raw = clf_political(text, truncation=True)[0]
+            label = str(raw["label"]).upper()
+            if label == "LABEL_0":
+                label = "FAKE"
+            elif label == "LABEL_1":
+                label = "REAL"
+            results["political"] = {"verdict": label, "score": float(raw["score"])}
+        except Exception:
+            results["political"] = None
+    if clf_general is not None:
+        try:
+            raw = clf_general(text, truncation=True)[0]
+            label = str(raw["label"]).upper()
+            if label == "LABEL_1":
+                label = "FAKE"
+            elif label == "LABEL_0":
+                label = "REAL"
+            results["general"] = {"verdict": label, "score": float(raw["score"])}
+        except Exception:
+            results["general"] = None
+    return results
 
 
 def tavily_search(query):
     if not TAVILY_AVAILABLE:
-        return None
+        return []
     try:
-        response = TAVILY_CLIENT.search(query, max_results=5)
-        content_pieces = []
+        response = TAVILY_CLIENT.search(query[:400], max_results=5)
+        results = []
         for r in response.get("results", []):
-            title = r.get("title", "")
-            content = r.get("content", "")
-            content_pieces.append(f"{title}: {content}")
-        return "\n".join(content_pieces)
+            results.append({
+                "title": r.get("title", ""),
+                "url": r.get("url", ""),
+                "content": r.get("content", ""),
+            })
+        return results
     except Exception:
+        return []
+
+
+def parse_llm_json(raw):
+    if not raw:
         return None
+    text = raw.strip()
+    if text.startswith("```"):
+        text = text.strip("`")
+        if text.startswith("json"):
+            text = text[4:]
+    try:
+        return json.loads(text)
+    except Exception:
+        pass
+    start = text.find("{")
+    end = text.rfind("}")
+    if start != -1 and end > start:
+        try:
+            return json.loads(text[start:end + 1])
+        except Exception:
+            return None
+    return None
 
 
-def is_us_political(text):
-    keywords = [
-        "president", "congress", "senate", "house", "democrat", "republican",
-        "biden", "trump", "politics", "political", "us government", "white house",
-        "politi", "liar", "election", "campaign", "supreme court",
-    ]
-    text_lower = text.lower()
-    return any(kw in text_lower for kw in keywords)
+SYSTEM_PROMPT = (
+    "You are a professional fact-checking analyst working inside a consumer fact-checking "
+    "app. You will receive a piece of content, two AI classifier verdicts with confidence "
+    "scores, and web search evidence. Your job is to decide how truthful the content is. "
+    "You are the final judge: your verdict overrides the classifier verdicts.\n"
+    'Respond ONLY with a valid JSON object in this exact shape:\n'
+    '{"verdict": "REAL" or "FAKE" or "UNKNOWN", "confidence": <integer 0-100>, '
+    '"explanation": "<2-3 plain-language sentences>", '
+    '"key_points": ["short point", "short point", "short point"], '
+    '"sources": ["source title", "source title"]}\n'
+    "Rules:\n"
+    "- Use the web evidence as the strongest signal; the classifiers are secondary.\n"
+    "- Choose UNKNOWN only when there is genuinely not enough evidence to decide.\n"
+    "- Never mention classifiers, models, AI, or internal details in the explanation.\n"
+    "- Keep the explanation and key points in simple language a general audience understands."
+)
 
 
-def generate_ai_explanation(text, classification, confidence):
+def synthesize_verdict(content, model_results, evidence):
     if not GROQ_AVAILABLE:
         return None
-    prompt = (
-        f"Analyze this content that an AI model classified as '{classification}' "
-        f"with {confidence:.1f}% confidence.\n\n"
-        f"Content: {text[:400]}...\n\n"
-        f"Provide a concise professional explanation (2-4 sentences) of why this "
-        f"classification is or isn't correct. If the classification appears wrong, "
-        f"say so explicitly and explain."
-    )
+    payload = {
+        "content": content,
+        "classifier_verdicts": {k: v for k, v in model_results.items() if v},
+        "web_evidence": [
+            {"title": s["title"], "content": s["content"][:200]} for s in evidence
+        ],
+    }
+    prompt = json.dumps(payload, ensure_ascii=False)[:6000]
     for model in GROQ_MODELS:
         try:
             response = GROQ_CLIENT.chat.completions.create(
                 model=model,
-                messages=[{"role": "user", "content": prompt}],
-                temperature=0.3,
-                max_tokens=400,
+                messages=[
+                    {"role": "system", "content": SYSTEM_PROMPT},
+                    {"role": "user", "content": prompt},
+                ],
+                temperature=0.2,
+                max_tokens=700,
+                response_format={"type": "json_object"},
             )
-            content = response.choices[0].message.content
-            if content and content.strip():
-                return content.strip()
+            data = parse_llm_json(response.choices[0].message.content)
+            if data and data.get("verdict") in ("REAL", "FAKE", "UNKNOWN"):
+                return data
         except Exception:
             continue
     return None
 
 
-def analyze_with_models(text, classifier_b1, classifier_b2):
-    text_stripped = text.strip()
-    use_brain1 = is_us_political(text_stripped)
-    brain_name = "Brain 1 (Political)" if use_brain1 else "Brain 2 (General)"
-    classifier = classifier_b1 if use_brain1 else classifier_b2
-
-    try:
-        results = classifier(text_stripped, truncation=True)
-        raw_label = results[0]["label"]
-        if use_brain1:
-            label = normalize_brain1_label(raw_label)
-        else:
-            label = normalize_brain2_label(raw_label)
-        confidence = float(results[0]["score"]) * 100.0
-    except Exception:
-        label, confidence, summary = get_fallback_analysis(text_stripped)
-        return label, confidence, summary, brain_name, False
-
-    summary, corrected = build_summary(text_stripped, label, confidence)
-    return corrected, confidence, summary, brain_name, True
-
-
-def build_summary(text, label, confidence):
-    ai_explanation = generate_ai_explanation(text, label, confidence)
-    if ai_explanation:
-        lowered = ai_explanation.lower()
-        markers = ["incorrect", "wrong", f"not {label.lower()}", "misclassification"]
-        if any(marker in lowered for marker in markers):
-            corrected = "REAL" if label == "FAKE" else "FAKE"
-            return ai_explanation, corrected
-        return ai_explanation, label
-
-    if TAVILY_AVAILABLE:
-        tavily_info = tavily_search(text)
-        if tavily_info:
-            return (
-                f"Content classified as {label} by the model with {confidence:.1f}% "
-                f"confidence, cross-referenced against live web sources.", label,
-            )
-    return (
-        f"Content classified as {label} by the model with {confidence:.1f}% confidence.", label,
-    )
+def fallback_verdict(model_results):
+    preferred = None
+    for key in ("general", "political"):
+        if model_results.get(key):
+            preferred = model_results[key]
+            break
+    if preferred is None:
+        return "UNKNOWN", 0, (
+            "We couldn't analyze this content right now. Please try again in a moment."
+        )
+    label = preferred["verdict"]
+    confidence = int(round(preferred["score"] * 100))
+    if label == "FAKE":
+        explanation = (
+            "This content contains patterns commonly associated with misleading or "
+            "false information. We recommend double-checking it against trusted sources."
+        )
+    elif label == "REAL":
+        explanation = (
+            "This content looks consistent with factual reporting. It matches patterns "
+            "commonly found in verified, reliable information."
+        )
+    else:
+        explanation = "We couldn't determine the truthfulness of this content with confidence."
+    return label, confidence, explanation
 
 
-def get_fallback_analysis(text):
-    fake_indicators = ['fake', 'hoax', 'conspiracy', 'false', 'lie', 'scam', 'fraud', 'misleading']
-    real_indicators = ['study', 'research', 'according', 'official', 'confirmed', 'verified', 'report']
-    text_lower = text.lower()
-    fake_score = sum(1 for word in fake_indicators if word in text_lower)
-    real_score = sum(1 for word in real_indicators if word in text_lower)
-    if fake_score > real_score:
-        return "FAKE", random.uniform(60.0, 80.0), "Fallback heuristic analysis: Likely FAKE content detected."
-    if real_score > fake_score:
-        return "REAL", random.uniform(60.0, 80.0), "Fallback heuristic analysis: Likely REAL content detected."
-    return "UNCERTAIN", 50.0, "Fallback heuristic analysis: Unable to classify definitively."
+def run_analysis(text, clf_political, clf_general):
+    content = text.strip()
+    if len(content) > MAX_ANALYSIS_CHARS:
+        content = content[:MAX_ANALYSIS_CHARS]
+    model_results = classify_with_models(content, clf_political, clf_general)
+    evidence = tavily_search(content)
+    llm_data = None
+    if GROQ_AVAILABLE:
+        llm_data = synthesize_verdict(content, model_results, evidence)
+    if llm_data:
+        verdict = llm_data.get("verdict", "UNKNOWN")
+        confidence = int(llm_data.get("confidence", 0))
+        explanation = llm_data.get("explanation", "")
+        key_points = llm_data.get("key_points", [])
+        sources = llm_data.get("sources", [])
+    else:
+        verdict, confidence, explanation = fallback_verdict(model_results)
+        key_points = []
+        sources = []
+    return {
+        "verdict": verdict,
+        "confidence": confidence,
+        "explanation": explanation,
+        "key_points": key_points,
+        "sources": sources,
+        "evidence": evidence,
+    }
 
 
 @st.cache_data(show_spinner=False, ttl=300)
 def fetch_web_content(url):
     try:
         headers = {
-            'User-Agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 '
-                          'Chrome/124.0.0.0 Safari/537.36'
+            "User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 "
+                          "Chrome/124.0.0.0 Safari/537.36"
         }
         response = requests.get(url, headers=headers, timeout=15)
         response.raise_for_status()
-        soup = BeautifulSoup(response.content, 'html.parser')
+        soup = BeautifulSoup(response.content, "html.parser")
 
-        for element in soup(['script', 'style', 'nav', 'footer', 'aside']):
+        for element in soup(["script", "style", "nav", "footer", "aside"]):
             element.decompose()
 
-        title_tag = soup.find('title')
+        title_tag = soup.find("title")
         title = title_tag.get_text(strip=True) if title_tag else "No title found"
 
-        paragraphs = soup.find_all('p')
+        paragraphs = soup.find_all("p")
         content = " ".join(
-            p.get_text(strip=True)
-            for p in paragraphs
-            if len(p.get_text(strip=True)) > 20
+            p.get_text(strip=True) for p in paragraphs if len(p.get_text(strip=True)) > 20
         )
-
         full_text = f"{title}\n\n{content}"
         return {
-            'success': True,
-            'title': title,
-            'content': content,
-            'full_text': full_text,
-            'word_count': len(full_text.split()),
-            'url': url,
+            "success": True,
+            "title": title,
+            "full_text": full_text,
+            "word_count": len(full_text.split()),
         }
     except Exception as e:
-        return {'success': False, 'error': str(e)}
+        return {"success": False, "error": str(e)}
 
 
-def process_analysis(user_input, input_method, classifier_b1, classifier_b2):
+def process_analysis(user_input, input_method, clf_political, clf_general):
     start_time = time.time()
-    with st.status("🧠 Analyzing with dual-AI system...", expanded=True) as status:
-        if input_method == "URL/Website" and user_input.startswith(('http://', 'https://')):
-            st.write("🌐 Fetching content from URL...")
+    with st.status("🔍 Analyzing...", expanded=True) as status:
+        if input_method == "URL/Website" and user_input.startswith(("http://", "https://")):
+            st.write("🌐 Reading the article...")
             web_data = fetch_web_content(user_input)
-            if web_data['success']:
-                text_to_analyze = web_data['full_text']
-                st.write(f"✅ Successfully extracted {web_data['word_count']} words")
+            if web_data["success"]:
+                text_to_analyze = web_data["full_text"]
+                st.write(f"✅ Got {web_data['word_count']} words from the page")
             else:
-                st.error(f"❌ Failed to fetch content: {web_data['error']}")
+                st.error(f"❌ Couldn't read that page: {web_data['error']}")
                 return
         else:
             text_to_analyze = user_input
 
-        if len(text_to_analyze) > 3000:
-            text_to_analyze = text_to_analyze[:3000]
-            st.write("✂️ Text truncated for optimal processing")
+        if len(text_to_analyze) > MAX_ANALYSIS_CHARS:
+            text_to_analyze = text_to_analyze[:MAX_ANALYSIS_CHARS]
+            st.write("✂️ Focused on the most relevant part of the content")
 
-        label, confidence, summary, brain_name, model_used = analyze_with_models(
-            text_to_analyze, classifier_b1, classifier_b2
-        )
+        st.write("🌐 Checking claims across the web...")
+        result = run_analysis(text_to_analyze, clf_political, clf_general)
 
         analysis_time = time.time() - start_time
-        status.update(label="✅ Analysis complete!", state="complete")
+        status.update(label="✅ Analysis ready!", state="complete")
 
     results = {
-        'verdict': label,
-        'confidence': confidence,
-        'summary': summary,
-        'analysis_time': analysis_time,
-        'brain': brain_name,
-        'model_used': model_used,
-        'input': user_input[:200] + "..." if len(user_input) > 200 else user_input,
-        'full_input': user_input,
+        **result,
+        "analysis_time": analysis_time,
+        "input": user_input[:200] + "..." if len(user_input) > 200 else user_input,
+        "full_input": user_input,
+        "timestamp": datetime.now().strftime("%b %d, %Y at %H:%M"),
     }
 
     st.session_state.current_results = results
     st.session_state.analysis_complete = True
 
-    if 'analysis_history' not in st.session_state:
+    if "analysis_history" not in st.session_state:
         st.session_state.analysis_history = []
     st.session_state.analysis_history.insert(0, results)
-    if len(st.session_state.analysis_history) > 10:
-        st.session_state.analysis_history = st.session_state.analysis_history[:10]
+    if len(st.session_state.analysis_history) > 20:
+        st.session_state.analysis_history = st.session_state.analysis_history[:20]
 
     st.rerun()
 
 
-def render_analysis_interface(classifier_b1, classifier_b2):
-    st.markdown("### 🔍 Content Analysis")
-    input_method = st.selectbox(
-        "Select input method:",
-        ["Direct Text", "URL/Website", "File Upload"],
-        help="Choose how you want to provide content for fact-checking",
+def confidence_band(confidence):
+    if confidence >= 70:
+        return "High confidence"
+    if confidence >= 45:
+        return "Medium confidence"
+    return "Low confidence"
+
+
+def render_result(results, compact=False):
+    verdict = results["verdict"]
+    confidence = int(results.get("confidence", 0))
+    band = confidence_band(confidence)
+
+    chip_text = {"REAL": "Real", "FAKE": "Fake", "UNKNOWN": "Unclear"}.get(verdict, "Unclear")
+    chip_class = {
+        "REAL": "verdict-real",
+        "FAKE": "verdict-fake",
+        "UNKNOWN": "verdict-unknown",
+    }.get(verdict, "verdict-unknown")
+
+    st.markdown(
+        f"""
+        <div class="result-card">
+            <div class="verdict-head">
+                <div class="verdict-chip {chip_class}">{chip_text}</div>
+                <div class="conf-line">
+                    {band}
+                    <div class="conf-bar">
+                        <div class="conf-fill" style="width: {max(min(confidence, 100), 4)}%;"></div>
+                    </div>
+                </div>
+            </div>
+            <p class="explanation">{results.get("explanation", "")}</p>
+        """,
+        unsafe_allow_html=True,
     )
-    user_input = ""
-    if input_method == "Direct Text":
-        user_input = st.text_area(
-            "Enter text to analyze:",
-            height=150,
-            placeholder="Paste the content you want to fact-check here...",
-            help="Enter any text content for misinformation detection",
-        )
-    elif input_method == "URL/Website":
-        user_input = st.text_input(
-            "Enter website URL:",
-            placeholder="https://example.com/article",
-            help="Provide the URL of an article or webpage to analyze",
+
+    key_points = results.get("key_points") or []
+    if key_points:
+        points_html = "".join(
+            f'<div class="key-point"><span class="dot">•</span><span>{p}</span></div>'
+            for p in key_points
         )
-        if user_input and not user_input.startswith(('http://', 'https://')):
-            st.warning("⚠️ Please enter a complete URL starting with http:// or https://")
-    elif input_method == "File Upload":
-        uploaded_file = st.file_uploader(
-            "Upload text file:",
-            type=['txt', 'md'],
-            help="Upload a text file containing the content to analyze",
+        st.markdown(
+            f'<div class="sub-block"><div class="sub-title">Key takeaways</div>{points_html}</div>',
+            unsafe_allow_html=True,
         )
-        if uploaded_file:
-            try:
-                user_input = uploaded_file.getvalue().decode("utf-8", errors="replace")
-                st.success(f"✅ File loaded: {len(user_input)} characters")
-                if len(user_input) > 500:
-                    st.text_area(
-                        "Content preview:", user_input[:500] + "...", height=100, disabled=True
-                    )
-            except Exception as e:
-                st.error(f"❌ Error reading file: {str(e)}")
-                user_input = ""
 
-    st.markdown("---")
-    col1, col2, col3 = st.columns([3, 1, 1])
-    with col1:
-        analyze_btn = st.button(
-            "🧠 Analyze with Dual-AI",
-            type="primary",
-            disabled=not user_input.strip(),
-            help="Start the AI-powered fact-checking analysis",
+    sources = results.get("sources") or []
+    evidence = results.get("evidence") or []
+    source_links = []
+    if sources:
+        source_links.extend(sources)
+    if not source_links:
+        for e in evidence[:4]:
+            if e.get("url"):
+                source_links.append(f'{e["title"] or e["url"]} ({e["url"]})')
+    if source_links:
+        links_html = ""
+        for src in source_links[:5]:
+            url = None
+            label = src
+            if "http" in src:
+                end = src.rfind(" (http")
+                if end != -1:
+                    label = src[:end]
+                    url = src[end + 2:-1]
+            if url:
+                links_html += f'<a class="source-link" href="{url}" target="_blank">{label}</a>'
+            else:
+                links_html += f'<div class="source-link">{label}</div>'
+        st.markdown(
+            f'<div class="sub-block"><div class="sub-title">Sources</div>{links_html}</div>',
+            unsafe_allow_html=True,
         )
-    with col2:
-        if st.button("🔄 Clear", help="Clear current results and start over"):
-            st.session_state.analysis_complete = False
-            st.session_state.current_results = {}
-            st.rerun()
-    with col3:
-        export_enabled = st.session_state.get('analysis_complete', False)
-        if st.button("📄 Export", disabled=not export_enabled, help="Export analysis results"):
-            if export_enabled:
-                export_results()
 
-    if analyze_btn:
-        if not user_input.strip():
-            st.warning("⚠️ Please provide some content to analyze.")
-        elif len(user_input.strip()) < 10:
-            st.warning("⚠️ Please provide more content for meaningful analysis (minimum 10 characters).")
-        elif input_method == "URL/Website" and not user_input.startswith(('http://', 'https://')):
-            st.warning("⚠️ Please enter a valid URL starting with http:// or https://")
-        else:
-            process_analysis(user_input, input_method, classifier_b1, classifier_b2)
+    if results.get("timestamp"):
+        st.caption(f"Analyzed {results['timestamp']} · {results.get('analysis_time', 0):.1f}s")
+    st.markdown("</div>", unsafe_allow_html=True)
 
 
 def export_results():
-    if not st.session_state.get('current_results'):
-        st.warning("⚠️ No results to export!")
-        return
     results = st.session_state.current_results
     export_data = {
-        'analysis_timestamp': datetime.now().isoformat(),
-        'input_text': results.get('full_input', results.get('input', '')),
-        'verdict': results.get('verdict', ''),
-        'confidence_score': float(results.get('confidence', 0)),
-        'ai_summary': results.get('summary', ''),
-        'analysis_model': results.get('brain', ''),
-        'analysis_time': results.get('analysis_time', 0),
+        "timestamp": datetime.now().isoformat(),
+        "input_text": results.get("full_input", results.get("input", "")),
+        "verdict": results.get("verdict", ""),
+        "confidence": int(results.get("confidence", 0)),
+        "explanation": results.get("explanation", ""),
+        "key_points": results.get("key_points", []),
+        "sources": results.get("sources", []),
     }
     json_string = json.dumps(export_data, indent=2, default=str, ensure_ascii=False)
     st.download_button(
-        label="📥 Download Analysis Report",
+        label="📥 Download report",
         data=json_string,
-        file_name=f"credo_ai_analysis_{datetime.now().strftime('%Y%m%d_%H%M%S')}.json",
+        file_name=f"credo_report_{datetime.now().strftime('%Y%m%d_%H%M%S')}.json",
         mime="application/json",
     )
-    st.success("📄 Analysis report ready for download!")
 
 
-def render_analysis_results(results):
-    st.markdown("### ✨ AI-Powered Analysis Summary")
-    st.markdown(
-        f"""<div class="summary-box">{results['summary']}</div>""",
-        unsafe_allow_html=True,
+def render_input_interface(clf_political, clf_general):
+    input_method = st.selectbox(
+        "How would you like to provide the content?",
+        ["Paste text", "Article link", "Upload a file"],
     )
-    col1, col2 = st.columns(2, gap="large")
+    user_input = ""
+    if input_method == "Paste text":
+        user_input = st.text_area(
+            "Paste a claim, headline, or article excerpt",
+            height=160,
+            placeholder="Paste the text you want to fact-check...",
+        )
+    elif input_method == "Article link":
+        user_input = st.text_input(
+            "Article URL",
+            placeholder="https://example.com/article",
+        )
+        if user_input and not user_input.startswith(("http://", "https://")):
+            st.warning("Please enter a complete URL starting with http:// or https://")
+    else:
+        uploaded = st.file_uploader("Upload a text file", type=["txt", "md"])
+        if uploaded:
+            user_input = uploaded.getvalue().decode("utf-8", errors="replace")
+            st.success(f"Loaded {len(user_input)} characters")
+
+    col1, col2 = st.columns([3, 1])
     with col1:
-        st.markdown("### 🎯 Primary Verdict")
-        verdict = results['verdict']
-        confidence = results['confidence']
-        verdict_class = 'verdict-real'
-        if verdict == 'FAKE':
-            verdict_class = 'verdict-fake'
-        elif verdict == 'UNCERTAIN':
-            verdict_class = 'verdict-uncertain'
-        st.markdown(
-            f"""
-            <div class="verdict-container {verdict_class}">
-                <div class="verdict-text">{verdict}</div>
-            </div>
-            <div style="text-align: center; margin-top: 1rem; font-size: 1.4rem; font-weight: 700; color: #334155;">
-                {confidence:.1f}% Confidence
-            </div>
-            """,
-            unsafe_allow_html=True,
+        analyze_btn = st.button(
+            "🧠 Fact-check now",
+            type="primary",
+            disabled=not user_input.strip(),
+            use_container_width=True,
         )
-        st.progress(min(float(confidence) / 100.0, 1.0), text="Model confidence")
     with col2:
-        st.markdown("### 📊 Analysis Details")
-        st.metric("Processing Time", f"{results.get('analysis_time', 0):.2f}s")
-        st.metric("Content Length", f"{len(results.get('input', '').split())} words")
-        st.metric("Analysis Model", results.get('brain', 'Brain 2 (General)'))
-        if results.get('model_used'):
-            st.metric("Summary Layer", "Groq AI")
+        if st.button("🗑️ Clear", disabled=not (user_input or st.session_state.analysis_complete)):
+            st.session_state.analysis_complete = False
+            st.session_state.current_results = {}
+            st.rerun()
+
+    if analyze_btn:
+        if not user_input.strip():
+            st.warning("Please add some content to fact-check.")
+        elif len(user_input.strip()) < 10:
+            st.warning("Please add a bit more text — at least 10 characters.")
+        elif input_method == "Article link" and not user_input.startswith(("http://", "https://")):
+            st.warning("Please enter a valid URL.")
         else:
-            st.metric("Summary Layer", "Built-in fallback")
+            process_analysis(user_input, input_method, clf_political, clf_general)
 
 
-if 'analysis_complete' not in st.session_state:
+if "analysis_complete" not in st.session_state:
     st.session_state.analysis_complete = False
-if 'current_results' not in st.session_state:
+if "current_results" not in st.session_state:
     st.session_state.current_results = {}
-if 'analysis_history' not in st.session_state:
+if "analysis_history" not in st.session_state:
     st.session_state.analysis_history = []
 
 with st.sidebar:
     st.markdown(
         """
-        <div style="text-align: center; padding: 1rem 0; margin-bottom: 2rem;">
-            <div style="font-size: 2.4rem; margin-bottom: 0.4rem;">🧠</div>
-            <h2 style="color: #4f46e5; margin: 0;">Credo AI</h2>
-            <p style="color: #94a3b8; margin: 0.5rem 0 0 0; font-size: 0.9rem;">Truth Detection Platform</p>
+        <div class="brand">
+            <div class="brand-logo">🧠</div>
+            <div class="brand-name">Credo</div>
+            <p class="brand-tag">Truth, verified.</p>
         </div>
         """,
         unsafe_allow_html=True,
     )
-
-    page = st.radio(
-        "Navigate:",
-        ["🚀 Live Analysis", "📜 History", "ℹ️ About"],
-        key="navigation",
-    )
-
-    if st.session_state.analysis_history:
-        st.markdown("---")
-        st.markdown("### 📈 Quick Stats")
-        total = len(st.session_state.analysis_history)
-        fake_count = sum(
-            1 for h in st.session_state.analysis_history if h.get('verdict') == 'FAKE'
-        )
-        st.metric("Total Analyses", total)
-        if total > 0:
-            st.metric("Fake Rate", f"{(fake_count / total * 100):.0f}%")
-
+    page = st.radio("Menu", ["✅ Fact Check", "🕘 History", "ℹ️ How it works"])
     st.markdown("---")
-    st.markdown("### 🔧 Status")
-    if GROQ_AVAILABLE:
-        st.success("🟢 Groq Enhanced")
-    else:
-        st.warning("🟡 Basic Mode")
-
-    st.markdown("---")
-    if st.button("🗑️ Clear History", help="Clear all analysis history"):
+    if st.button("🗑️ Clear history"):
         st.session_state.analysis_history = []
         st.session_state.analysis_complete = False
         st.session_state.current_results = {}
-        st.success("History cleared!")
+        st.success("History cleared")
         time.sleep(1)
         st.rerun()
-
-if page == "🚀 Live Analysis":
     st.markdown(
         """
-        <div class="hero-container">
-            <h1 class="main-title">🧠 Credo AI Platform</h1>
-            <p class="hero-subtitle">
-                Next-generation misinformation detection powered by
-                <strong>dual-AI architecture</strong>. Analyze text, articles, and claims
-                with speed and insight.
-            </p>
-            <div class="metrics-container">
-                <div class="metric-card">
-                    <span class="metric-value">2</span>
-                    <span class="metric-label">AI Brains</span>
-                </div>
-                <div class="metric-card">
-                    <span class="metric-value">FAKE/REAL</span>
-                    <span class="metric-label">Verdict</span>
-                </div>
-                <div class="metric-card">
-                    <span class="metric-value">&lt;3s</span>
-                    <span class="metric-label">Analysis Time</span>
-                </div>
-            </div>
+        <div class="trust-row" style="flex-direction: column; gap: 0.5rem; margin-top: 1rem;">
+            <div class="trust-chip">Free to use</div>
+            <div class="trust-chip">Private — nothing is stored</div>
+            <div class="trust-chip">Results in seconds</div>
         </div>
         """,
         unsafe_allow_html=True,
     )
 
-    if not GROQ_AVAILABLE:
-        st.info(
-            "🔑 **Optional Setup:** Add GROQ_API_KEY in Space Settings → Variables and "
-            "Secrets for AI-powered summaries with Groq. The platform works without it "
-            "using intelligent fallback analysis."
-        )
+if page == "✅ Fact Check":
+    st.markdown(
+        """
+        <div class="hero">
+            <h1 class="hero-title">Check any claim, <span>before you share it.</span></h1>
+            <p class="hero-sub">Paste a headline, an article, or a link — get a clear
+            verdict with sources in seconds.</p>
+        </div>
+        """,
+        unsafe_allow_html=True,
+    )
 
-    classifier_b1, classifier_b2 = load_ai_models()
-    if classifier_b1 is None or classifier_b2 is None:
-        st.error("Failed to load AI models! Please restart the app or check logs.")
+    clf_political, clf_general = load_engines()
+    if clf_political is None and clf_general is None:
+        st.error("The analysis engines failed to start. Please reload the page to retry.")
     else:
-        render_analysis_interface(classifier_b1, classifier_b2)
-
-    if st.session_state.analysis_complete and st.session_state.current_results:
-        st.markdown("---")
-        st.markdown("## 📊 Analysis Results")
-        render_analysis_results(st.session_state.current_results)
-
-elif page == "📜 History":
-    st.markdown("# 📜 Analysis History")
+        render_input_interface(clf_political, clf_general)
+
+        if st.session_state.analysis_complete and st.session_state.current_results:
+            st.markdown("### Your result")
+            render_result(st.session_state.current_results)
+            st.markdown("---")
+            col_a, col_b = st.columns([1, 3])
+            with col_a:
+                export_results()
+            with col_b:
+                st.markdown(
+                    "⚠️ AI-generated analysis — always verify important information with "
+                    "trusted sources before sharing."
+                )
+
+elif page == "🕘 History":
+    st.markdown("# 🕘 Your history")
     if st.session_state.analysis_history:
+        st.markdown("### At a glance")
+        stats = st.columns(3)
         total = len(st.session_state.analysis_history)
-        fake_count = sum(
-            1 for h in st.session_state.analysis_history if h.get('verdict') == 'FAKE'
-        )
-        real_count = sum(
-            1 for h in st.session_state.analysis_history if h.get('verdict') == 'REAL'
-        )
-        st.markdown("### 📈 Summary Statistics")
-        stat_cols = st.columns(3)
-        with stat_cols[0]:
-            st.metric("Total Analyses", total)
-        with stat_cols[1]:
-            st.metric("Fake Content", fake_count)
-        with stat_cols[2]:
-            st.metric("Real Content", real_count)
+        fake_count = sum(1 for h in st.session_state.analysis_history if h.get("verdict") == "FAKE")
+        real_count = sum(1 for h in st.session_state.analysis_history if h.get("verdict") == "REAL")
+        stats[0].metric("Checks", total)
+        stats[1].metric("Marked fake", fake_count)
+        stats[2].metric("Marked real", real_count)
         st.markdown("---")
         for i, result in enumerate(st.session_state.analysis_history):
-            with st.expander(
-                f"#{i + 1} - {result.get('verdict', 'Unknown')} | {result.get('input', 'No input')}",
-                expanded=(i == 0),
-            ):
-                render_analysis_results(result)
+            chip = {
+                "REAL": "✅ Real",
+                "FAKE": "🚩 Fake",
+                "UNKNOWN": "⚠️ Unclear",
+            }.get(result.get("verdict"), "⚠️ Unclear")
+            with st.expander(f"{chip} · {result.get('input', '')}", expanded=(i == 0)):
+                render_result(result)
     else:
-        st.info(
-            "📚 **No Analysis History** - Your analysis history will appear here after you "
-            "perform some fact-checking analyses. Start by going to the Live Analysis page "
-            "and analyzing some content!"
-        )
+        st.info("Nothing here yet. Run your first fact check and it will show up here.")
 
-elif page == "ℹ️ About":
-    st.markdown("# 🔬 About Credo AI")
+elif page == "ℹ️ How it works":
+    st.markdown("# ℹ️ How it works")
     st.markdown(
         """
-        <div class="glass-card">
-            <h2 style="color: #4f46e5; margin-bottom: 1rem;">🚀 Revolutionary Detection Technology</h2>
-            <p style="font-size: 1.2rem; color: #475569; line-height: 1.7;">
-                Credo AI represents a breakthrough in automated fact-checking, combining
-                <strong>two specialized neural networks</strong> with advanced language
-                understanding to deliver rapid, transparent misinformation detection.
-            </p>
+        <div class="steps-grid">
+            <div class="step-card">
+                <div class="step-num">1</div>
+                <p class="step-title">You share the content</p>
+                <p class="step-text">Paste a claim, headline, article, or a link to a page.</p>
+            </div>
+            <div class="step-card">
+                <div class="step-num">2</div>
+                <p class="step-title">We cross-check it</p>
+                <p class="step-text">Your content is checked against live web sources and
+                analyzed by AI trained to spot misleading patterns.</p>
+            </div>
+            <div class="step-card">
+                <div class="step-num">3</div>
+                <p class="step-title">You get a clear verdict</p>
+                <p class="step-text">A simple verdict — Real, Fake, or Unclear — with
+                confidence, key takeaways, and sources.</p>
+            </div>
+        </div>
+        """,
+        unsafe_allow_html=True,
+    )
+    st.markdown(
+        """
+        <div class="result-card">
+            <div class="sub-title">The honest part</div>
+            <p class="step-text">Credo is a helpful first filter, not a perfect truth machine.
+            AI pattern detection works best on news-style text and can miss false claims that
+            look completely normal. Always check important information with trusted sources
+            before you share it.</p>
         </div>
         """,
         unsafe_allow_html=True,
     )
-    tab1, tab2, tab3 = st.tabs(["🧠 AI Architecture", "📊 Performance", "🔬 Technology"])
-
-    with tab1:
-        st.markdown(
-            """
-            ### ⚡ Brain 2: The Specialist
-            - **Model:** `Arko007/fact-check1-v3-final` (DeBERTa-v3-large)
-            - **Function:** Rapid FAKE/REAL binary classification
-            - **Training:** 50,000+ verified news articles + calibration passes
-            - **Speed:** Sub-second inference time
-
-            ### 🧠 Brain 1: The Political Expert
-            - **Model:** `Arko007/fake-news-liar-political` (RoBERTa-base)
-            - **Function:** Binary political fact-checking (US-centric)
-            - **Training:** LIAR dataset converted to binary
-            - **Performance:** ~71% accuracy
-            - **Specialization:** Short political statement classification
-
-            ### ⚡ Groq Integration
-            - **Role:** Intelligent synthesis & explanation layer
-            - **Model:** `qwen/qwen3.6-27b` (fallback: `openai/gpt-oss-120b`)
-            - **Function:** Validates classifications and explains verdicts in plain language
-            - **Speed:** Blazing-fast inference on Groq LPU hardware
-            """
-        )
-
-    with tab2:
-        st.markdown("### 📈 Performance Metrics")
-        import pandas as pd
-
-        metrics_data = {
-            'Metric': ['Accuracy', 'Precision', 'Recall', 'F1-Score', 'Speed'],
-            'Brain 1': ['71.4%', 'N/A', 'N/A', 'N/A', 'N/A'],
-            'Brain 2': ['99.9%', '99.8%', '99.7%', '99.7%', '0.8s'],
-            'Combined': ['~95%', 'N/A', 'N/A', 'N/A', '<3s'],
-        }
-        st.dataframe(
-            pd.DataFrame(metrics_data),
-            width="stretch",
-            hide_index=True,
-        )
-        st.success("🏆 Credo AI blends specialized models to maximize coverage and accuracy.")
-
-    with tab3:
-        st.markdown(
-            """
-            ### 🛠️ Technology Stack
-
-            **🤖 Core AI/ML:**
-            - PyTorch deep learning framework
-            - Hugging Face Transformers for model handling
-            - RoBERTa & DeBERTa-v3 fine-tuned classifiers
-
-            **🌐 Web & Integration:**
-            - Streamlit for responsive UI
-            - Beautiful Soup for web scraping
-            - Groq Cloud API (`qwen/qwen3.6-27b`)
-            - Tavily real-time information search
-            - Custom CSS for enhanced UX
-
-            **⚡ Performance:**
-            - Intelligent caching system
-            - Memory-efficient processing
-            - Mobile-responsive design
-            - Privacy-first architecture
-            """
-        )
 
 st.markdown(
-    """
-    <div class="footer-enhanced">
-        <div class="footer-features">
-            <div class="footer-feature">
-                <div class="footer-feature-icon">🏆</div>
-                <div class="footer-feature-text">Award Winning</div>
-            </div>
-            <div class="footer-feature">
-                <div class="footer-feature-icon">⚡</div>
-                <div class="footer-feature-text">Lightning Fast</div>
-            </div>
-            <div class="footer-feature">
-                <div class="footer-feature-icon">🔒</div>
-                <div class="footer-feature-text">Privacy First</div>
-            </div>
-            <div class="footer-feature">
-                <div class="footer-feature-icon">🌍</div>
-                <div class="footer-feature-text">Global Impact</div>
-            </div>
-        </div>
-        <div style="font-size: 0.9rem; opacity: 0.85;">
-            Built with ❤️ for Hack2Skill Hackathon 2025 | 🐉 Data Dragons Team
-        </div>
-        <div style="font-size: 0.8rem; opacity: 0.6; margin-top: 0.5rem;">
-            Powered by Advanced AI • Making Truth Accessible to Everyone
-        </div>
-    </div>
-    """,
+    '<div class="footer">Made with ❤️ for the Hack2Skill Hackathon 2025 · '
+    "Credo — truth, verified.</div>",
     unsafe_allow_html=True,
 )