app.py

"""
Oil Intelligence Terminal — v4
Per-article Brent sentiment classification with structured rendering
"""

import streamlit as st
import subprocess
from collections import Counter
from datetime import datetime
import pandas as pd

st.set_page_config(page_title="Oil Intelligence", page_icon="🛢", layout="wide", initial_sidebar_state="collapsed")

# ── CSS ───────────────────────────────────────────────────────────────────────
st.markdown("""
<style>
@import url('https://fonts.googleapis.com/css2?family=IBM+Plex+Mono:wght@400;500;600&family=IBM+Plex+Sans:wght@300;400;500;600&display=swap');
html, body, [class*="css"] { font-family: 'IBM Plex Sans', sans-serif; background: #0a0a0a; color: #d8d0c0; }
.stApp { background: #0a0a0a; }
#MainMenu, footer, header { visibility: hidden; }
.block-container { padding: 1.5rem 2.5rem; max-width: 1200px; }

/* Header */
.hdr-title { font-family:'IBM Plex Mono',monospace; font-size:1rem; font-weight:500;
             color:#f0a500; letter-spacing:0.15em; text-transform:uppercase; margin:0; }
.hdr-sub   { font-family:'IBM Plex Mono',monospace; font-size:0.6rem; color:#333;
             letter-spacing:0.08em; margin-top:0.2rem; }

/* Indicators */
.ind-strip { display:flex; gap:0.6rem; margin-bottom:1.5rem; }
.ind-card  { flex:1; background:#0d0d0d; border:1px solid #181818; border-radius:2px;
             padding:0.6rem 0.8rem; min-width:0; }
.ind-lbl   { font-family:'IBM Plex Mono',monospace; font-size:0.5rem; color:#444;
             letter-spacing:0.1em; text-transform:uppercase; white-space:nowrap;
             overflow:hidden; text-overflow:ellipsis; }
.ind-val   { font-family:'IBM Plex Mono',monospace; font-size:1.1rem; font-weight:600;
             color:#f0a500; margin-top:0.15rem; }
.ind-chg   { font-family:'IBM Plex Mono',monospace; font-size:0.55rem; color:#555;
             margin-top:0.1rem; }
.ind-up    { color:#4caf50; }
.ind-down  { color:#ef5350; }

/* Tabs */
.stTabs [data-baseweb="tab-list"] { gap:0; border-bottom:1px solid #1a1a1a; background:transparent; }
.stTabs [data-baseweb="tab"] {
    font-family:'IBM Plex Mono',monospace; font-size:0.6rem; letter-spacing:0.1em;
    text-transform:uppercase; color:#444; background:transparent; border:none;
    border-bottom:2px solid transparent; padding:0.6rem 1.2rem; }
.stTabs [aria-selected="true"] { color:#f0a500!important; border-bottom:2px solid #f0a500!important; background:transparent!important; }

/* Output boxes */
.out  { background:#0d0d0d; border:1px solid #181818; border-left:3px solid #f0a500;
        border-radius:2px; padding:1.2rem; font-size:0.82rem; line-height:1.65;
        color:#bbb; margin-top:0.8rem; white-space:pre-wrap; }
.memo { background:#0d0d0d; border:1px solid #181818; border-left:3px solid #d8d0c0;
        border-radius:2px; padding:1.5rem; font-size:0.82rem; line-height:1.7;
        color:#bbb; margin-top:0.8rem; white-space:pre-wrap; }
.risk-out { background:#0d0d0d; border:1px solid #181818; border-left:3px solid #ef5350;
            border-radius:2px; padding:1.2rem; font-size:0.82rem; line-height:1.65;
            color:#bbb; margin-top:0.8rem; white-space:pre-wrap; }

/* Article cards (NEW) */
.article-card {
    background:#0d0d0d; border:1px solid #181818; border-radius:2px;
    padding:1rem 1.2rem; margin-top:0.6rem;
}
.article-headline {
    font-family:'IBM Plex Sans',sans-serif; font-size:0.9rem; font-weight:500;
    color:#e8e0d0; line-height:1.4; margin-bottom:0.4rem;
}
.article-event {
    font-family:'IBM Plex Sans',sans-serif; font-size:0.75rem; color:#888;
    line-height:1.5; margin-bottom:0.8rem;
}
.article-impact-row {
    display:flex; gap:0.4rem; flex-wrap:wrap; margin-bottom:0.6rem;
}
.impact-pill {
    font-family:'IBM Plex Mono',monospace; font-size:0.55rem; letter-spacing:0.08em;
    text-transform:uppercase; padding:0.2rem 0.55rem; border-radius:2px;
    border:1px solid; white-space:nowrap;
}
.impact-pill .lbl { color:#666; font-weight:400; margin-right:0.3rem; }
.imp-bullish { background:#001a08; border-color:#003a18; color:#4caf50; }
.imp-bearish { background:#1a0000; border-color:#3a0000; color:#ef5350; }
.imp-neutral { background:#0f0f0f; border-color:#222; color:#777; }
.imp-mixed   { background:#1a1000; border-color:#3a2400; color:#ffa726; }

.article-reason {
    font-size:0.72rem; color:#888; line-height:1.55; margin-top:0.5rem;
    padding-top:0.5rem; border-top:1px solid #181818;
}
.article-meta {
    display:flex; justify-content:space-between; align-items:center;
    margin-top:0.6rem; font-family:'IBM Plex Mono',monospace; font-size:0.55rem;
    color:#444;
}
.confidence-bar {
    display:inline-block; width:80px; height:3px; background:#222;
    border-radius:1px; overflow:hidden; vertical-align:middle; margin-left:0.4rem;
}
.confidence-fill { height:100%; background:#f0a500; }

/* Aggregate summary card */
.agg-card {
    background:#0d0d0d; border:1px solid #181818; border-left:3px solid #f0a500;
    border-radius:2px; padding:1.2rem; margin-bottom:1rem;
}
.agg-title { font-family:'IBM Plex Mono',monospace; font-size:0.6rem; color:#666;
             letter-spacing:0.1em; text-transform:uppercase; margin-bottom:0.6rem; }
.agg-row { display:flex; gap:1.5rem; }
.agg-item { flex:1; }
.agg-count { font-family:'IBM Plex Mono',monospace; font-size:1.8rem; font-weight:600; }
.agg-label { font-family:'IBM Plex Mono',monospace; font-size:0.55rem;
             letter-spacing:0.1em; text-transform:uppercase; color:#555;
             margin-top:0.2rem; }
.agg-bullish { color:#4caf50; }
.agg-bearish { color:#ef5350; }
.agg-neutral { color:#777; }
.agg-mixed   { color:#ffa726; }

/* Source cards */
.src-card { background:#0b0b0b; border:1px solid #151515; border-radius:2px;
            padding:0.6rem 0.8rem; margin-top:0.4rem; }
.src-hdr  { display:flex; justify-content:space-between; font-family:'IBM Plex Mono',monospace;
            font-size:0.55rem; color:#555; }
.src-preview { color:#666; font-size:0.7rem; line-height:1.5; margin-top:0.3rem;
               max-height:60px; overflow:hidden; }
.src-url  { color:#4a9eff; font-size:0.55rem; word-break:break-all; text-decoration:none; }

/* Chips */
.chip { display:inline-block; font-family:'IBM Plex Mono',monospace; font-size:0.5rem;
        letter-spacing:0.08em; text-transform:uppercase; padding:0.15rem 0.4rem;
        border-radius:2px; margin-right:0.3rem; margin-top:0.6rem; }
.c-news       { background:#1a1200; color:#f0a500; border:1px solid #2a1a00; }
.c-prices     { background:#001a0a; color:#4caf50; border:1px solid #002a14; }
.c-eia        { background:#001020; color:#4a9eff; border:1px solid #001a30; }
.c-eia_report { background:#001020; color:#4a9eff; border:1px solid #001a30; }
.c-opec       { background:#1a000a; color:#f06292; border:1px solid #2a0014; }

/* Risk */
.sev-HIGH   { background:#1a0000; color:#ff5252; padding:0.1rem 0.4rem; border:1px solid #330000;
              border-radius:2px; font-size:0.5rem; font-family:'IBM Plex Mono',monospace; }
.sev-MEDIUM { background:#1a1000; color:#ffa726; padding:0.1rem 0.4rem; border:1px solid #332000;
              border-radius:2px; font-size:0.5rem; font-family:'IBM Plex Mono',monospace; }
.sev-LOW    { background:#111; color:#666; padding:0.1rem 0.4rem; border:1px solid #222;
              border-radius:2px; font-size:0.5rem; font-family:'IBM Plex Mono',monospace; }
.rsk-card { background:#0d0d0d; border:1px solid #181818; border-radius:2px;
            padding:0.6rem; text-align:center; }
.rsk-cnt  { font-family:'IBM Plex Mono',monospace; font-size:1.3rem; font-weight:600;
            margin-top:0.15rem; }
.rsk-active { color:#f0a500; }
.rsk-clear  { color:#222; }

/* Eval */
.ev-card  { background:#0d0d0d; border:1px solid #181818; border-radius:2px;
            padding:0.8rem; text-align:center; }
.ev-lbl   { font-family:'IBM Plex Mono',monospace; font-size:0.5rem; color:#444;
            letter-spacing:0.1em; text-transform:uppercase; }
.ev-val   { font-family:'IBM Plex Mono',monospace; font-size:1.6rem; font-weight:600;
            margin-top:0.3rem; }
.ev-pass  { color:#4caf50; } .ev-warn  { color:#ffa726; } .ev-fail  { color:#ef5350; }

/* Buttons */
.stButton > button {
    font-family:'IBM Plex Mono',monospace; font-size:0.6rem; letter-spacing:0.1em;
    text-transform:uppercase; background:transparent; border:1px solid #f0a500;
    color:#f0a500; padding:0.4rem 1.2rem; border-radius:2px; }
.stButton > button:hover { background:#f0a500; color:#0a0a0a; }

.stTextInput > div > div > input, .stTextArea > div > div > textarea {
    background:#0d0d0d; border:1px solid #181818; border-radius:2px;
    color:#d8d0c0; font-family:'IBM Plex Mono',monospace; font-size:0.75rem; }

.lbl { font-family:'IBM Plex Mono',monospace; font-size:0.55rem; letter-spacing:0.1em;
       text-transform:uppercase; color:#444; margin-bottom:0.3rem; }
hr { border-color:#141414; }
</style>
""", unsafe_allow_html=True)


# ── Loaders ───────────────────────────────────────────────────────────────────
@st.cache_resource(show_spinner="Loading Oil Intelligence...")
def load_rag():
    from RAG import OilIntelligence
    return OilIntelligence()

@st.cache_data(ttl=3600)
def load_indicators():
    from indicators import all_indicators
    return all_indicators()

@st.cache_data(ttl=3600)
def load_risk_scan():
    from risk_and_eval import RiskIntelligence
    return RiskIntelligence().scan_all()


def render_sources(sources: list[dict]):
    types = sorted(set(s.get("source", "") for s in sources))
    chips = "".join(f'<span class="chip c-{t}">{t.replace("_"," ")}</span>' for t in types)
    st.markdown(f'<div>{chips}</div>', unsafe_allow_html=True)
    st.markdown('<p class="lbl" style="margin-top:0.8rem;">Retrieved Sources</p>', unsafe_allow_html=True)
    for i, s in enumerate(sources):
        score = s.get("score", 0)
        date  = s.get("date", "")
        url   = s.get("url", "")
        src   = s.get("source", "").upper()
        prev  = s.get("text", "")[:300].replace("\n", " ")
        with st.expander(f"📎 {i+1}  ·  {src}  ·  {date}  ·  {score:.3f}", expanded=False):
            if url:
                st.markdown(f'<a class="src-url" href="{url}" target="_blank">{url}</a>', unsafe_allow_html=True)
            st.markdown(f'<div class="src-preview">{prev}...</div>', unsafe_allow_html=True)


def render_article_classification(item: dict):
    """Render one classified article as a structured card."""
    cls = item["classification"]
    chunk = item["chunk"]

    if cls.get("_parse_failed"):
        st.markdown(f'''
        <div class="article-card" style="border-left:3px solid #ef5350;">
            <div class="article-headline">⚠️ Classifier Parse Failed</div>
            <div class="article-event">{cls.get("_raw","")[:200]}</div>
        </div>''', unsafe_allow_html=True)
        return

    # Build impact pills row
    dimensions = [
        ("Supply",      cls.get("supply_effect", "neutral")),
        ("Demand",      cls.get("demand_effect", "neutral")),
        ("Geopolitical", cls.get("geopolitical_risk_effect", "neutral")),
        ("Short-Term",  cls.get("short_term_brent_impact", "neutral")),
        ("Long-Term",   cls.get("medium_long_term_brent_impact", "neutral")),
        ("Overall",     cls.get("overall_brent_impact", "neutral")),
    ]
    pills_html = ""
    for label, value in dimensions:
        v = value.lower() if isinstance(value, str) else "neutral"
        pills_html += f'<span class="impact-pill imp-{v}"><span class="lbl">{label}</span>{v.upper()}</span>'

    confidence = cls.get("confidence", 0)
    if isinstance(confidence, str):
        try: confidence = int(confidence)
        except: confidence = 0

    headline = cls.get("headline", "—")
    event    = cls.get("core_event", "")
    reason   = cls.get("reason", "")
    url      = chunk.get("url", "")
    date     = chunk.get("date", "")

    st.markdown(f'''
    <div class="article-card">
        <div class="article-headline">{headline}</div>
        <div class="article-event">{event}</div>
        <div class="article-impact-row">{pills_html}</div>
        <div class="article-reason">{reason}</div>
        <div class="article-meta">
            <span>{date}{"  ·  " + url[:60] + "..." if url else ""}</span>
            <span>CONFIDENCE: {confidence}
                <span class="confidence-bar">
                    <span class="confidence-fill" style="width:{confidence}%;"></span>
                </span>
            </span>
        </div>
    </div>
    ''', unsafe_allow_html=True)


def render_aggregate(classifications: list[dict]):
    """Show counts of bullish/bearish/neutral/mixed across articles."""
    overalls = [c["classification"].get("overall_brent_impact", "neutral").lower()
                for c in classifications if not c["classification"].get("_parse_failed")]
    counts = Counter(overalls)

    items_html = ""
    for label in ["bullish", "bearish", "mixed", "neutral"]:
        count = counts.get(label, 0)
        items_html += f'''
        <div class="agg-item">
            <div class="agg-count agg-{label}">{count}</div>
            <div class="agg-label">{label}</div>
        </div>'''

    total = len(overalls)
    # Determine net sentiment
    if counts.get("bullish", 0) > counts.get("bearish", 0) * 1.5:
        net = "NET BULLISH"
        net_class = "agg-bullish"
    elif counts.get("bearish", 0) > counts.get("bullish", 0) * 1.5:
        net = "NET BEARISH"
        net_class = "agg-bearish"
    elif counts.get("mixed", 0) > total / 2:
        net = "MIXED"
        net_class = "agg-mixed"
    else:
        net = "BALANCED / NEUTRAL"
        net_class = "agg-neutral"

    st.markdown(f'''
    <div class="agg-card">
        <div class="agg-title">Brent Sentiment · {total} Articles Analyzed</div>
        <div class="agg-row">{items_html}</div>
        <div style="margin-top:0.8rem; font-family:'IBM Plex Mono',monospace; font-size:0.75rem;
                    letter-spacing:0.1em;" class="{net_class}">→ {net}</div>
    </div>''', unsafe_allow_html=True)


# ── Header with refresh ──────────────────────────────────────────────────────
col_hdr, col_refresh = st.columns([8, 2])
with col_hdr:
    st.markdown("""
    <div>
        <p class="hdr-title">🛢 Oil Intelligence Terminal</p>
        <p class="hdr-sub">RAG · Brent Sentiment Classifier · EIA · NewsAPI · Risk Monitoring</p>
    </div>""", unsafe_allow_html=True)
with col_refresh:
    st.markdown("<div style='height:0.5rem;'></div>", unsafe_allow_html=True)
    if st.button("🔄 Refresh Data", key="refresh"):
        with st.spinner("Pulling fresh data, re-scraping, rebuilding index..."):
            subprocess.run(["python", "day1_data_pipeline.py"], capture_output=True)
            subprocess.run(["python", "scraper.py"], capture_output=True)
            import RAG
            RAG.build_index(force=True)
        st.cache_data.clear()
        st.cache_resource.clear()
        st.rerun()

st.markdown('<hr style="margin:0.5rem 0 1rem 0;">', unsafe_allow_html=True)

# Indicators
indicators = load_indicators()
if indicators:
    ind_html = '<div class="ind-strip">'
    for ind in indicators[:6]:
        trend_cls = "ind-up" if ind.get("trend") == "up" else "ind-down" if ind.get("trend") == "down" else ""
        chg = f'<div class="ind-chg {trend_cls}">{ind.get("change","")}</div>' if ind.get("change") else ""
        ind_html += f'''
        <div class="ind-card">
            <div class="ind-lbl">{ind["label"]}</div>
            <div class="ind-val">{ind["value"]} <span style="font-size:0.55rem;color:#555;">{ind.get("unit","")}</span></div>
            {chg}
        </div>'''
    ind_html += '</div>'
    st.markdown(ind_html, unsafe_allow_html=True)

try:
    oil = load_rag()
except Exception as e:
    st.error(f"RAG init failed: {e}")
    st.stop()

tab1, tab2, tab3, tab4, tab5 = st.tabs(["Brief", "Q&A", "Memo", "Deal Risk", "Eval"])


# ── TAB 1: Brent Sentiment Brief ─────────────────────────────────────────────
with tab1:
    st.markdown('<p class="lbl">Brent Crude Sentiment Analysis</p>', unsafe_allow_html=True)
    st.markdown('<p style="color:#555; font-size:0.7rem; margin-bottom:1rem;">'
                'Per-article structured classification for front-month Brent price direction.</p>',
                unsafe_allow_html=True)

    if st.button("Run Brent Sentiment Analysis", key="news_btn"):
        with st.spinner("Classifying articles across 6 dimensions each..."):
            classifications, sources = oil.news_brief_classified()

        if not classifications:
            st.warning("No news articles found in corpus. Hit Refresh Data to pull fresh news.")
        else:
            # Aggregate summary
            render_aggregate(classifications)

            # Per-article cards
            st.markdown('<p class="lbl" style="margin-top:1.2rem;">Article-Level Classifications</p>',
                        unsafe_allow_html=True)
            for item in classifications:
                render_article_classification(item)


# ── TAB 2: Q&A ───────────────────────────────────────────────────────────────
with tab2:
    st.markdown('<p class="lbl">Ask the Data</p>', unsafe_allow_html=True)
    presets = [
        "Current US crude inventory?",
        "WTI price trend last month?",
        "Key supply risks now?",
        "WTI-Brent spread?",
    ]
    pcols = st.columns(len(presets))
    selected = None
    for i, (c, p) in enumerate(zip(pcols, presets)):
        with c:
            if st.button(p, key=f"p_{i}"):
                selected = p

    question = st.text_input("q", value=selected or "", placeholder="Ask anything about oil markets...", label_visibility="collapsed")
    if st.button("Ask", key="ask_btn") and question:
        with st.spinner("Retrieving..."):
            answer, sources = oil.ask_with_sources(question)
        st.markdown(f'<div class="out">{answer}</div>', unsafe_allow_html=True)
        render_sources(sources)


# ── TAB 3: Memo ──────────────────────────────────────────────────────────────
with tab3:
    st.markdown('<p class="lbl">Investment Memo Generator</p>', unsafe_allow_html=True)
    memo_topic = st.text_input("mt", placeholder="e.g. WTI Q3 2026 outlook / OPEC production cuts / shale thesis", label_visibility="collapsed")
    if st.button("Generate Memo", key="memo_btn") and memo_topic:
        with st.spinner("Generating..."):
            answer, sources = oil.memo_with_sources(memo_topic)
        st.markdown(f'''
        <div style="font-family:IBM Plex Mono,monospace; font-size:0.55rem; color:#333;
                    letter-spacing:0.08em; margin-top:1rem;">
            INVESTMENT MEMORANDUM · {datetime.now().strftime("%B %d, %Y").upper()}
        </div>''', unsafe_allow_html=True)
        st.markdown(f'<div class="memo">{answer}</div>', unsafe_allow_html=True)
        st.download_button(
            "Download .txt",
            data=f"INVESTMENT MEMORANDUM\n{datetime.now().strftime('%B %d, %Y')}\nTopic: {memo_topic}\n\n{answer}",
            file_name=f"memo_{datetime.now().strftime('%Y%m%d')}.txt", mime="text/plain",
        )
        render_sources(sources)


# ── TAB 4: Deal Risk ─────────────────────────────────────────────────────────
with tab4:
    st.markdown('<p class="lbl">Risk Monitor</p>', unsafe_allow_html=True)
    risk_report = load_risk_scan()

    cat_html = '<div style="display:flex; gap:0.5rem; margin-bottom:1.2rem;">'
    for cat in risk_report["categories"]:
        cnt_cls = "rsk-active" if cat["count"] > 0 else "rsk-clear"
        sts_color = "#ef5350" if cat["count"] > 0 else "#333"
        cat_html += f'''
        <div class="rsk-card" style="flex:1;">
            <div style="font-size:1.2rem;">{cat["icon"]}</div>
            <div class="ind-lbl">{cat["label"]}</div>
            <div class="rsk-cnt {cnt_cls}">{cat["count"]}</div>
            <div style="color:{sts_color}; font-size:0.5rem; font-family:'IBM Plex Mono',monospace; letter-spacing:0.1em;">{cat["status"]}</div>
        </div>'''
    cat_html += '</div>'
    st.markdown(cat_html, unsafe_allow_html=True)

    high_docs = [d for d in risk_report["documents"] if d["severity"] == "HIGH"]
    if high_docs:
        st.markdown('<p class="lbl">High-Severity Signals</p>', unsafe_allow_html=True)
        for doc in high_docs[:3]:
            risks_html = " ".join(f'<span class="sev-{r["severity"]}">{r["icon"]} {r["label"]}</span>' for r in doc["risks"][:3])
            st.markdown(f'''
            <div class="src-card" style="border-left:3px solid #ef5350;">
                <div class="src-hdr"><span>{doc["source"].upper()} · {doc["date"]}</span><span class="sev-HIGH">HIGH</span></div>
                <div style="margin-top:0.3rem;">{risks_html}</div>
                <div class="src-preview">{doc["preview"]}...</div>
            </div>''', unsafe_allow_html=True)
        st.markdown("<br>", unsafe_allow_html=True)

    st.markdown('<p class="lbl">Deal Scenario Assessment</p>', unsafe_allow_html=True)
    scenario = st.text_area(
        "sc", placeholder="e.g. Long 500k bbl WTI Q1 2027, Permian sourced, Houston → Rotterdam",
        height=80, label_visibility="collapsed",
    )
    if st.button("Assess Risk", key="risk_btn") and scenario:
        with st.spinner("Scanning..."):
            answer, sources = oil.deal_risk_with_sources(scenario)
        st.markdown(f'<div class="risk-out">{answer}</div>', unsafe_allow_html=True)
        render_sources(sources)


# ── TAB 5: Eval ──────────────────────────────────────────────────────────────
with tab5:
    st.markdown('<p class="lbl">System Evaluation</p>', unsafe_allow_html=True)

    if st.button("Run Evaluation Suite", key="eval_btn"):
        from risk_and_eval import RAGEvaluator
        with st.spinner("Running 6-question eval suite..."):
            ev = RAGEvaluator(oil).evaluate()

        metrics = [
            ("Retrieval Precision",      ev["retrieval_precision"]),
            ("Keyword Coverage",         ev["keyword_coverage"]),
            ("Numeric Accuracy",         ev["numeric_accuracy"]),
            ("Hallucination Resistance", ev["hallucination_resistance"]),
        ]
        mcols = st.columns(4)
        for col, (label, score) in zip(mcols, metrics):
            with col:
                cls = "ev-pass" if score >= 0.8 else "ev-warn" if score >= 0.5 else "ev-fail"
                st.markdown(f'''
                <div class="ev-card">
                    <div class="ev-lbl">{label}</div>
                    <div class="ev-val {cls}">{score:.0%}</div>
                </div>''', unsafe_allow_html=True)

        st.markdown("<br>", unsafe_allow_html=True)
        st.markdown('<p class="lbl">Question-Level Results</p>', unsafe_allow_html=True)
        for r in ev["results"]:
            icon = "✅" if r["overall_pass"] else "❌"
            with st.expander(f'{icon}  [{r["category"]}]  {r["question"]}', expanded=False):
                st.markdown(f'<div class="src-preview" style="max-height:none;">{r["answer"][:600]}</div>', unsafe_allow_html=True)
                ccols = st.columns(4)
                for col, (lbl, ok) in zip(ccols, [("Retrieval", r["retrieval_pass"]),
                    ("Keywords", r["keyword_pass"]), ("Numeric", r["numeric_pass"]),
                    ("Refusal", r["refusal_pass"])]):
                    c = "#4caf50" if ok else "#ef5350"
                    i = "✓" if ok else "✗"
                    with col:
                        st.markdown(f'<div style="text-align:center;color:{c};font-family:IBM Plex Mono,monospace;font-size:0.65rem;">{i} {lbl}</div>', unsafe_allow_html=True)
    else:
        st.markdown('''
        <div class="out" style="border-left-color:#333;">
        <strong>Methodology</strong> — 6 curated questions across 4 dimensions:

        <strong>Retrieval Precision</strong> — correct source types retrieved
        <strong>Keyword Coverage</strong> — expected domain terms in answer
        <strong>Numeric Accuracy</strong> — numbers present when expected
        <strong>Hallucination Resistance</strong> — refuses impossible queries
        </div>''', unsafe_allow_html=True)


st.markdown('''
<hr style="margin-top:2rem;">
<p style="font-family:IBM Plex Mono,monospace; font-size:0.5rem; color:#1a1a1a; text-align:center; letter-spacing:0.1em;">
LLAMA 3.2 · FAISS · EIA · NEWSAPI · RESEARCH ONLY
</p>''', unsafe_allow_html=True)