app.py

import streamlit as st
import pandas as pd
import numpy as np
import joblib
import random
import html

if "lang" not in st.session_state:
    st.session_state.lang = "en"


def t(en, es):
    return es if st.session_state.lang == "es" else en


# =========================
# Configuración general
# =========================

st.set_page_config(
    page_title="DPRK UPR Response Prediction Tool",
    page_icon="🇰🇵",
    layout="wide"
)

# =========================
# CSS
# =========================

st.markdown("""
<style>
.block-container {
    max-width: 1100px;
    padding-top: 2rem;
    padding-bottom: 4rem;
}
h1 { margin-bottom: 0.3rem; }
h2, h3 { margin-top: 1.4rem; }
div[data-testid="stDataFrame"] { border-radius: 12px; }
.stRadio > div { gap: 1rem; }
.result-card-accept {
    padding: 2rem;
    border-radius: 18px;
    background-color: #ecfdf3;
    border: 1px solid #bbf7d0;
    margin-bottom: 1rem;
}
.result-card-reject {
    padding: 2rem;
    border-radius: 18px;
    background-color: #fef2f2;
    border: 1px solid #fecaca;
    margin-bottom: 1rem;
}
.warning-card {
    padding: 1rem 1.2rem;
    border-radius: 14px;
    background-color: #fffbeb;
    border: 1px solid #fde68a;
    margin-bottom: 1rem;
}
.intro-highlight {
    display: inline-block;
    margin-top: 0.85rem;
    margin-bottom: 1.2rem;
    padding: 0.35rem 0.6rem;
    background-color: #fef3a6;
    border-radius: 6px;
    font-weight: 600;
}
.opening-copy {
    font-size: 1.06rem;
    margin-bottom: 0.9rem;
    line-height: 1.6;
}
.body-copy,
.stMarkdown p {
    text-align: justify;
}
[data-testid="stImage"] img {
    border-radius: 6px;
}
</style>
""", unsafe_allow_html=True)


def section_header(text):
    st.markdown(f'<h2 style="color:#FF4B4B;">{text}</h2>', unsafe_allow_html=True)


st.image("UN.jpg", use_container_width=True)

st.title("🇰🇵 DPRK UPR Response Prediction Tool")

with st.container(border=True):
    read_in_spanish = st.toggle(
        "App en Español",
        value=st.session_state.lang == "es",
    )
st.session_state.lang = "es" if read_in_spanish else "en"

opening_intro = t(
    "This tool simulates how the Democratic People's Republic of Korea (DPRK) might respond to a recommendation during the next Universal Periodic Review (UPR) process.",
    "La presente herramienta permite estimar posibles respuestas de la República Popular Democrática de Corea (RPDC) frente a recomendaciones formuladas durante el próximo ciclo del Examen Periódico Universal (EPU).",
)

opening_body = t(
    "Using historical UPR data and machine learning models, the tool estimates the <strong>probability that a recommendation will be accepted or rejected</strong> based on political, thematic, and discursive variables.<br><br><strong>❗ Predictions are probabilistic and should be interpreted as analytical scenarios rather than deterministic forecasts</strong>.",
    "Utilizando datos históricos del EPU y modelos de aprendizaje automático, la aplicación estima la <strong>probabilidad de aceptación o rechazo de una recomendación</strong> a partir de variables políticas, temáticas y discursivas.<br><br><strong>❗ Las predicciones son probabilísticas y deben interpretarse como escenarios analíticos, no como pronósticos deterministas</strong>.",
)

st.markdown(
    f'<div class="opening-copy body-copy">{opening_intro}</div>',
    unsafe_allow_html=True,
)

st.markdown(
    f'<div class="opening-copy body-copy">{opening_body}</div>',
    unsafe_allow_html=True,
)

st.markdown(
    f'<div class="intro-highlight">{t("Complete the fields below to generate a prediction.", "Complete los campos a continuación para generar una predicción.")}</div>',
    unsafe_allow_html=True,
)

# =========================
# Cargar modelo y archivos
# =========================

@st.cache_resource
def load_artifacts():
    model = joblib.load("modelo_xgb.pkl")

    alignment_lookup = pd.read_csv("alignment_lookup.csv")
    areas_lookup = pd.read_csv("areas_lookup.csv")
    historical_recommendations = pd.read_csv("historical_recommendations.csv")

    return (
        model,
        alignment_lookup,
        areas_lookup,
        historical_recommendations
    )


try:
    (
        model,
        alignment_lookup,
        areas_lookup,
        historical_recommendations
    ) = load_artifacts()

except FileNotFoundError:
    st.error(
        t(
            "Missing artifact. Make sure these files are in the same folder as app.py: "
            "modelo_xgb.pkl, alignment_lookup.csv, areas_lookup.csv, "
            "historical_recommendations.csv.",
            "Falta un archivo necesario. Asegúrese de que estos archivos estén en la misma carpeta que app.py: "
            "modelo_xgb.pkl, alignment_lookup.csv, areas_lookup.csv e "
            "historical_recommendations.csv.",
        )
    )
    st.stop()

except Exception as e:
    st.error(t(f"Error loading artifacts: {e}", f"Error al cargar archivos: {e}"))
    st.stop()

# Obtener directamente del modelo las columnas exactas de entrenamiento
model_columns = model.get_booster().feature_names

if not model_columns:
    st.error(
        t(
            "The model does not contain feature names. Retrain and save it using a pandas DataFrame with named columns.",
            "El modelo no contiene nombres de variables. Vuelva a entrenarlo y guardarlo usando un DataFrame de pandas con columnas nombradas.",
        )
    )
    st.stop()

model_columns = list(model_columns)

# =========================
# Normalizar lookups
# =========================

alignment_lookup.columns = alignment_lookup.columns.str.strip()
areas_lookup.columns = areas_lookup.columns.str.strip()

required_alignment_cols = {"Recommending_State", "Aligment_Score"}

if not required_alignment_cols.issubset(set(alignment_lookup.columns)):
    st.error(
        t(
            "alignment_lookup.csv must contain the columns: "
            "Recommending_State and Aligment_Score.",
            "alignment_lookup.csv debe contener las columnas: "
            "Recommending_State y Aligment_Score.",
        )
    )
    st.stop()

area_lookup_col = areas_lookup.columns[0]

alignment_lookup["Recommending_State"] = (
    alignment_lookup["Recommending_State"].astype(str).str.strip()
)

areas_lookup[area_lookup_col] = (
    areas_lookup[area_lookup_col].astype(str).str.strip()
)

available_areas = sorted(
    [x for x in areas_lookup[area_lookup_col].dropna().unique() if x != ""]
)

model_area_columns = [
    col for col in model_columns
    if col.startswith("Macro_Area_")
]
model_areas = sorted([
    col.replace("Macro_Area_", "", 1)
    for col in model_area_columns
])

areas_for_app = sorted(set(available_areas).intersection(set(model_areas)))

SENSITIVITY_BY_AREA = {
    "Civil Liberties & Political Rights": 2,
    "Disability Rights": 0,
    "Economic & Social Rights": 0,
    "Environment, Climate & Emergencies": 0,
    "Equality & Non-Discrimination": 1,
    "Gender Rights & Violence": 1,
    "Governance, Institutions & Rule of Law": 1,
    "Human Rights Education & Awareness": 0,
    "International Mechanisms & Cooperation": 1,
}

if len(areas_for_app) == 0:
    st.error(
        t(
            "No thematic areas in areas_lookup.csv match the Macro_Area dummy columns used by the model. "
            "Regenerate areas_lookup.csv from the same dataset used to train the model.",
            "Ninguna área temática de areas_lookup.csv coincide con las columnas dummy Macro_Area utilizadas por el modelo. "
            "Regenere areas_lookup.csv a partir del mismo dataset usado para entrenar el modelo.",
        )
    )
    st.stop()

missing_from_model = sorted(set(available_areas) - set(model_areas))

if len(missing_from_model) > 0:
    st.markdown(f"""
    <div class="warning-card">
    {t(
        "Some thematic areas in <code>areas_lookup.csv</code> are not present in the trained model columns "
        "and were excluded from the selector. This usually means the lookup file and the model were not "
        "generated from exactly the same training dataset.",
        "Algunas áreas temáticas de <code>areas_lookup.csv</code> no están presentes en las columnas con las que se entrenó el modelo "
        "y fueron excluidas del selector. Esto normalmente significa que el archivo lookup y el modelo no fueron "
        "generados a partir exactamente del mismo dataset de entrenamiento.",
    )}
    </div>
    """, unsafe_allow_html=True)

# =========================
# 1. País recomendante
# =========================

section_header(t("1. Recommending State", "1. Estado recomendante"))

st.markdown(t(
    """
Select the State issuing the recommendation.
""",
    """
Seleccione el Estado que emite la recomendación.
""",
), unsafe_allow_html=True)

recommending_country = st.selectbox(
    t(
        "Recommending State",
        "Estado recomendante",
    ),
    options=sorted(alignment_lookup["Recommending_State"].dropna().unique())
)

alignment_score = alignment_lookup.loc[
    alignment_lookup["Recommending_State"] == recommending_country,
    "Aligment_Score"
].iloc[0]

st.caption(
    t(
        f"**Alignment Score: {alignment_score:.3f}**. This score captures ideological proximity using UN General Assembly voting patterns, following the literature that treats voting behavior as a proxy for state preferences (Voeten, 2013). In this project, it is operationalized as the average absolute voting distance between each State and the DPRK, where 0 indicates no ideological proximity and 1 indicates complete alignment.",
        f"**Puntaje de alineamiento: {alignment_score:.3f}**. Este puntaje captura la proximidad ideológica a partir de patrones de votación en la Asamblea General de las Naciones Unidas, siguiendo la literatura que utiliza el comportamiento de voto como proxy de las preferencias estatales (Voeten, 2013). Aquí, el puntaje de alineamiento se ha operacionalizado como la distancia absoluta promedio entre el voto de cada Estado y el de la RPDC, donde 0 indica ausencia de proximidad ideológica y 1 indica alineamiento completo.",
    )
)

st.divider()

# =========================
# 2. Verb Intensity
# =========================

section_header(t("2. Verb Intensity", "2. Intensidad verbal"))

st.markdown(t(
    "Use the scale below to classify how demanding the main verb of the recommendation is.",
    "Utilice la escala a continuación para clasificar cuán imperativo es el verbo principal de la recomendación.",
), unsafe_allow_html=True)

verb_options = {
    1: t("Low-demand", "Baja exigencia"),
    2: t("Continuity", "Continuidad"),
    3: t("Consideration", "Consideración"),
    4: t("General action", "Acción general"),
    5: t("Specific action", "Acción específica")
}

verb_intensity = st.select_slider(
    t("Verb intensity", "Intensidad verbal"),
    options=list(verb_options.keys()),
    format_func=lambda x: verb_options[x]
)

with st.expander(t("See full verb intensity classification guide", "Ver guía completa de clasificación de intensidad verbal")):
    st.markdown(t(
        """
### Low demand
Call, request, seek, share

### Continuity
Continue, maintain, persevere, persist, pursue, sustain

### Consideration
Assess, consider, envisage, examine, reconsider, review, study

### General action
Act (general), address, bolster, comply, cooperate, develop, encourage, engage, enhance, ensure, facilitate, foster, improve, increase, intensify, make every effort, pay attention, prioritize, promote, reduce, restrict, step up, strengthen, support, take measures

### Specific action
Abolish, accept, access, adopt, agree, align, allow, amend, become a party, build, carry out, cease, clarify, close, create, criminalize, decriminalize, dismantle, disclose, eliminate, enact, end, enforce, establish, extend, fulfil, guarantee, give, grant, halt, hold accountable, implement, incorporate, introduce, investigate, invest, invite, issue, join, lift, permit, prevent (when concrete), prohibit, prosecute, provide, publish, put an end, ratify, recognize, reform, refrain, release, remove, repatriate, respond, respect, return, revise, secure, sign, stop, submit, train, undertake, withdraw
""",
        """
### Baja exigencia
Call, request, seek, share

### Continuidad
Continue, maintain, persevere, persist, pursue, sustain

### Consideración
Assess, consider, envisage, examine, reconsider, review, study

### Acción general
Act (general), address, bolster, comply, cooperate, develop, encourage, engage, enhance, ensure, facilitate, foster, improve, increase, intensify, make every effort, pay attention, prioritize, promote, reduce, restrict, step up, strengthen, support, take measures

### Acción específica
Abolish, accept, access, adopt, agree, align, allow, amend, become a party, build, carry out, cease, clarify, close, create, criminalize, decriminalize, dismantle, disclose, eliminate, enact, end, enforce, establish, extend, fulfil, guarantee, give, grant, halt, hold accountable, implement, incorporate, introduce, investigate, invest, invite, issue, join, lift, permit, prevent (when concrete), prohibit, prosecute, provide, publish, put an end, ratify, recognize, reform, refrain, release, remove, repatriate, respond, respect, return, revise, secure, sign, stop, submit, train, undertake, withdraw
""",
    ))

st.divider()

# =========================
# 3. Thematic Area
# =========================

section_header(t("3. Thematic Area", "3. Área temática"))

st.markdown(
    t(
        "Select the primary thematic area addressed by the recommendation.",
        "Seleccione el área temática principal abordada por la recomendación.",
    ),
    unsafe_allow_html=True
)

area = st.selectbox(
    t("Thematic area", "Área temática"),
    options=areas_for_app
)

with st.expander(
    t(
        "See thematic area classification guide",
        "Ver guía de clasificación de áreas temáticas"
    )
):
    st.markdown(
        t(
            """
The thematic areas below are simplified analytical groupings derived from recurring issue categories across United Nations human rights mechanisms and UPR documentation. They do not correspond to an official UN taxonomy and were created exclusively for analytical and modeling purposes.

### Civil Liberties & Political Rights
Freedom of expression, association and religion, political participation, civil liberties, access to information, restrictions on dissent, and fundamental freedoms.

### Disability Rights
Rights of persons with disabilities, accessibility, inclusion, non-discrimination, social protection, education, health care, and participation.

### Economic & Social Rights
Health, education, food, housing, poverty reduction, labor rights, social protection, welfare, and access to basic services.

### Environment, Climate & Emergencies
Environmental protection, climate change, disaster risk reduction, emergency response, resilience, and humanitarian situations linked to environmental or climate-related crises.

### Equality & Non-Discrimination
General equality protections, anti-discrimination measures, minority rights, inclusion policies, and protection of vulnerable or marginalized groups.

### Gender Rights & Violence
Gender equality, discrimination against women and girls, gender-based violence, reproductive rights, participation, and protection from harmful practices.

### Governance, Institutions & Rule of Law
Institutional reform, legal frameworks, judicial guarantees, due process, public administration, accountability, and rule of law.

### Human Rights Education & Awareness
Human rights education, awareness campaigns, training programs, dissemination of human rights norms, and capacity-building initiatives.

### International Mechanisms & Cooperation
Treaty ratification, reporting obligations, cooperation with UN mechanisms, special procedures, international commitments, and engagement with human rights bodies.
""",
            """
Las áreas temáticas que se presentan a continuación son agrupaciones analíticas simplificadas derivadas de categorías recurrentes utilizadas en los mecanismos internacionales de derechos humanos y en la documentación del Examen Periódico Universal (EPU). No corresponden a una taxonomía oficial de las Naciones Unidas y fueron creadas exclusivamente con fines analíticos y de modelización.

### Libertades civiles y derechos políticos
Libertad de expresión, asociación y religión, participación política, libertades civiles, acceso a la información, restricciones a la disidencia y libertades fundamentales.

### Derechos de las personas con discapacidad
Derechos de las personas con discapacidad, accesibilidad, inclusión, no discriminación, protección social, educación, salud y participación.

### Derechos económicos y sociales
Salud, educación, alimentación, vivienda, reducción de la pobreza, derechos laborales, protección social, bienestar y acceso a servicios básicos.

### Medio ambiente, clima y emergencias
Protección ambiental, cambio climático, reducción del riesgo de desastres, respuesta ante emergencias, resiliencia y situaciones humanitarias vinculadas a crisis ambientales o climáticas.

### Igualdad y no discriminación
Protecciones generales de igualdad, medidas antidiscriminatorias, derechos de minorías, políticas de inclusión y protección de grupos vulnerables o marginados.

### Derechos de género y violencia
Igualdad de género, discriminación contra mujeres y niñas, violencia basada en género, derechos reproductivos, participación y protección frente a prácticas nocivas.

### Gobernanza, instituciones y Estado de derecho
Reformas institucionales, marcos legales, garantías judiciales, debido proceso, administración pública, rendición de cuentas y Estado de derecho.

### Educación y sensibilización en derechos humanos
Educación en derechos humanos, campañas de sensibilización, programas de formación, difusión de normas de derechos humanos e iniciativas de fortalecimiento de capacidades.

### Mecanismos internacionales y cooperación
Ratificación de tratados, obligaciones de reporte, cooperación con mecanismos de Naciones Unidas, procedimientos especiales, compromisos internacionales y vinculación con organismos de derechos humanos.
"""
        ),
        unsafe_allow_html=True
    )

st.divider()

# =========================
# 4. Political Sensitivity
# =========================

section_header(t("4. Political Sensitivity", "4. Sensibilidad política"))

st.markdown(t(
    """
Classify how politically sensitive the recommendation is for the DPRK regime. A suggested value is provided based on the thematic area selected above, but you may adjust it if needed.
""",
    """
Clasifique cuán políticamente sensible es la recomendación para el régimen norcoreano. Se ofrece un valor sugerido en función del área temática seleccionada arriba, pero puede ajustarlo si lo considera necesario.
""",
), unsafe_allow_html=True)

suggested_sensitivity = SENSITIVITY_BY_AREA.get(area, 1)

sensitivity_options = {
    0: t("Low sensitivity", "Baja sensibilidad"),
    1: t("Medium sensitivity", "Sensibilidad media"),
    2: t("High sensitivity", "Alta sensibilidad")
}

regime_sensitivity = st.select_slider(
    t("Political sensitivity", "Sensibilidad política"),
    options=list(sensitivity_options.keys()),
    value=suggested_sensitivity,
    format_func=lambda x: sensitivity_options[x],
)

with st.expander(t("See full political sensitivity classification guide", "Ver guía completa de clasificación de sensibilidad política")):
    st.markdown(t(
        """
### Low sensitivity
Recommendations focused on thematic areas generally perceived as less politically threatening by the DPRK regime.

Typical thematic areas:
- Economic & Social Rights
- Disability Rights
- Environment, Climate & Emergencies
- Human Rights Education & Awareness

<u>Examples</u>: food, health, education, housing, poverty, disability rights, humanitarian cooperation, disaster response, basic services, awareness-raising, training.

### Medium sensitivity
Recommendations focused on thematic areas that may imply institutional, legal, or normative change, but do not necessarily challenge regime control directly.

Typical thematic areas:
- Gender Rights & Violence
- Equality & Non-Discrimination
- Governance, Institutions & Rule of Law
- International Mechanisms & Cooperation

<u>Examples</u>: gender equality, anti-discrimination measures, legal reform, institutional strengthening, treaty ratification, reporting obligations, cooperation with UN mechanisms, capacity-building.

### High sensitivity
Recommendations focused on thematic areas directly linked to political control, repression, accountability, or fundamental civil and political freedoms.

Typical thematic areas:
- Civil Liberties & Political Rights

<u>Examples</u>: torture, prison camps, executions, death penalty, enforced disappearances, freedom of expression, freedom of association, freedom of religion, political participation, access to information, dissent, ICC, COI.
""",
        """
### Baja sensibilidad
Recomendaciones centradas en áreas temáticas generalmente percibidas como menos amenazantes políticamente por el régimen norcoreano.

Áreas temáticas típicas:
- Derechos económicos y sociales
- Derechos de las personas con discapacidad
- Medio ambiente, clima y emergencias
- Educación y sensibilización en derechos humanos

<u>Ejemplos</u>: alimentación, salud, educación, vivienda, pobreza, derechos de las personas con discapacidad, cooperación humanitaria, respuesta ante desastres, servicios básicos, sensibilización, formación.

### Sensibilidad media
Recomendaciones centradas en áreas temáticas que pueden implicar cambios institucionales, legales o normativos, pero que no necesariamente cuestionan directamente el control del régimen.

Áreas temáticas típicas:
- Derechos de género y violencia
- Igualdad y no discriminación
- Gobernanza, instituciones y Estado de derecho
- Mecanismos internacionales y cooperación

<u>Ejemplos</u>: igualdad de género, medidas antidiscriminatorias, reforma legal, fortalecimiento institucional, ratificación de tratados, obligaciones de reporte, cooperación con mecanismos de Naciones Unidas, fortalecimiento de capacidades.

### Alta sensibilidad
Recomendaciones centradas en áreas temáticas directamente vinculadas al control político, la represión, la rendición de cuentas o las libertades civiles y políticas fundamentales.

Área temática típica:
- Libertades civiles y derechos políticos

<u>Ejemplos</u>: tortura, campos de prisioneros, ejecuciones, pena de muerte, desapariciones forzadas, libertad de expresión, libertad de asociación, libertad religiosa, participación política, acceso a la información, disidencia, CPI, COI.
"""
    ), unsafe_allow_html=True)

st.divider()

# =========================
# 5. External Scrutiny
# =========================

section_header(t("5. External Scrutiny", "5. Escrutinio externo"))

st.markdown(t(
    """
Indicate whether the recommendation includes a concrete request for external access,
monitoring, reporting, visits, accountability mechanisms, or international oversight.
""",
    """
Indique si la recomendación incluye una solicitud concreta de acceso externo,
monitoreo, informes, visitas, mecanismos de rendición de cuentas o supervisión internacional.
""",
), unsafe_allow_html=True)

external_scrutiny = st.toggle(
    t("External scrutiny ON", "Escrutinio externo ACTIVADO"),
    value=False
)

external_scrutiny = int(external_scrutiny)

with st.expander(t("See full external scrutiny classification guide", "Ver guía completa de clasificación de escrutinio externo")):
    st.markdown(t(
        """
### Off · No external scrutiny
No explicit request for monitoring, reporting, visits, accountability mechanisms, or external intervention.

### On · External scrutiny
Explicit mention of external monitoring or accountability mechanisms, including references to: Special Rapporteur, OHCHR, ICC, COI, treaty bodies, humanitarian access, international observers, external investigations, reporting obligations.
""",
        """
### Desactivado · Sin escrutinio externo
No hay solicitud explícita de monitoreo, informes, visitas, mecanismos de rendición de cuentas o intervención externa.

### Activado · Con escrutinio externo
Existe una mención explícita a monitoreo externo o mecanismos de rendición de cuentas, incluidas referencias a: Relator/a Especial, OHCHR, CPI, COI, órganos de tratados, acceso humanitario, observadores internacionales, investigaciones externas u obligaciones de informe.
""",
    ))

st.divider()

# =========================
# 6. DPRK behavioral scenario
# =========================

section_header(t("6. DPRK behavioral scenario", "6. Escenario de comportamiento de la RPDC"))

st.markdown(t(
    """
The DPRK’s first three UPR cycles suggest a more balanced and less differentiated response pattern, approximating a **50/50 coin flip dynamic** in which acceptance and rejection remain relatively even across recommendation areas.

By contrast, the fourth UPR cycle reflects a more selective and structured response strategy, increasingly resembling a **Pareto style 80/20 logic**. The DPRK accepts a larger share of recommendations overall (~80%), especially in less politically sensitive areas, while continuing to reject recommendations perceived as intrusive, accountability oriented, or directly challenging regime control (~20%).

Select the behavioral scenario that should guide the prediction.
""",
    """
Los tres primeros ciclos del EPU sugieren un patrón de respuesta más equilibrado y menos diferenciado por parte del régimen norcoreano, aproximándose a una **dinámica de moneda 50/50** en la que la aceptación y el rechazo permanecen relativamente parejos entre las distintas áreas de recomendación.

El cuarto ciclo del EPU, por el contrario, refleja una estrategia de respuesta más selectiva y estructurada, cada vez más cercana a una **lógica de Pareto 80/20**. La RPDC acepta una mayor proporción de recomendaciones en términos generales (~80%), especialmente en áreas menos sensibles políticamente, mientras sigue rechazando aquellas percibidas como intrusivas, orientadas a la rendición de cuentas o directamente desafiantes para el control del régimen (~20%).

Seleccione el patrón de comportamiento que servirá de base para la predicción.
""",
))

scenario = st.radio(
    "",
    options=["coin_flip", "pareto_80_20"],
    horizontal=True,
    label_visibility="collapsed",
    format_func=lambda x: {
        "coin_flip": t("50/50 coin flip dynamic", "Dinámica de moneda 50/50"),
        "pareto_80_20": t("80/20 Pareto logic", "Lógica de Pareto 80/20")
    }[x]
)

if scenario == "coin_flip":
    cycle = random.choice([1, 2, 3])
elif scenario == "pareto_80_20":
    cycle = 4

st.caption(
    t(f"Internal model value: Cycle {cycle}.", f"Valor interno del modelo: Ciclo {cycle}.")
)

st.divider()

# =========================
# Construcción del input
# =========================

input_data = pd.DataFrame([{
    "Cycle": cycle,
    "Aligment_Score": alignment_score,
    "Verb_Intensity": verb_intensity,
    "Regime_Sensitivity": regime_sensitivity,
    "External_Scrutiny": external_scrutiny,
    "Macro_Area": area
}])

input_data["SimplVerb_Intensity"] = input_data["Verb_Intensity"].replace({
    1: 1,
    2: 1,
    3: 2,
    4: 2,
    5: 3
})

input_data["Pressure_Index"] = (
    input_data["Verb_Intensity"]
    * input_data["External_Scrutiny"]
    * input_data["Regime_Sensitivity"]
)

input_data["Aligned_Pressure"] = (
    input_data["Aligment_Score"]
    * input_data["Verb_Intensity"]
)

input_data["Adjusted_Intensity"] = (
    input_data["Verb_Intensity"]
    * input_data["Regime_Sensitivity"]
)

# =========================
# Encoding con las columnas reales del modelo
# =========================

def build_model_input(raw_input, model_columns):
    """
    Construye una fila con exactamente las mismas columnas,
    nombres y orden que utilizó el modelo durante el entrenamiento.
    """
    encoded = pd.DataFrame(
        0.0,
        index=[0],
        columns=model_columns
    )

    numeric_and_derived_features = [
        "Cycle",
        "Aligment_Score",
        "Verb_Intensity",
        "Regime_Sensitivity",
        "External_Scrutiny",
        "SimplVerb_Intensity",
        "Pressure_Index",
        "Aligned_Pressure",
        "Adjusted_Intensity"
    ]

    for col in numeric_and_derived_features:
        if col in raw_input.columns and col in encoded.columns:
            encoded.loc[0, col] = float(raw_input.loc[0, col])

    selected_area = str(raw_input.loc[0, "Macro_Area"]).strip()
    area_dummy_col = f"Macro_Area_{selected_area}"

    if area_dummy_col not in encoded.columns:
        raise ValueError(
            f"The selected area is not present in the model columns: {area_dummy_col}"
        )

    encoded.loc[0, area_dummy_col] = 1.0

    encoded = encoded.reindex(
        columns=model_columns,
        fill_value=0.0
    )

    return encoded, area_dummy_col


try:
    input_encoded, area_dummy_col = build_model_input(
        raw_input=input_data,
        model_columns=model_columns
    )
except Exception as e:
    st.error(t(
        f"Error preparing model input: {e}",
        f"Error al preparar el input del modelo: {e}"
    ))
    st.stop()

# =========================
# Predicción
# =========================

def find_similar_recommendations(input_data, historical_df, top_n=1):
    df_sim = historical_df.copy()

    df_sim = df_sim[
        df_sim["Macro_Area"] == input_data.loc[0, "Macro_Area"]
    ].copy()

    if len(df_sim) < top_n:
        df_sim = historical_df.copy()

    numeric_cols = [
        "Aligment_Score",
        "Verb_Intensity",
        "Regime_Sensitivity",
        "External_Scrutiny",
        "Cycle",
    ]

    for col in numeric_cols + ["Position_Binary"]:
        df_sim[col] = pd.to_numeric(df_sim[col], errors="coerce")

    df_sim = df_sim.dropna(
        subset=numeric_cols + [
            "Position_Binary",
            "Recommendation",
            "Macro_Area"
        ]
    )

    if df_sim.empty:
        return df_sim

    feature_ranges = {}
    for col in numeric_cols:
        col_min = float(df_sim[col].min())
        col_max = float(df_sim[col].max())
        feature_ranges[col] = max(col_max - col_min, 1e-9)

    normalized_diffs = []
    for col in numeric_cols:
        diff = (
            abs(df_sim[col] - float(input_data.loc[0, col]))
            / feature_ranges[col]
        )
        normalized_diffs.append(diff)

    df_sim["distance"] = sum(normalized_diffs) / len(normalized_diffs)
    df_sim["similarity"] = (
        (1 - df_sim["distance"]).clip(lower=0, upper=1) * 100
    )

    df_sim = df_sim.sort_values(
        ["distance", "similarity"],
        ascending=[True, False]
    ).head(top_n)

    return df_sim


st.header(t("Prediction", "Predicción"))

if st.button(t("Run prediction", "Ejecutar predicción"), type="primary"):
    if list(input_encoded.columns) != model_columns:
        st.error(
            t(
                "The encoded input columns do not match the model columns.",
                "Las columnas del input codificado no coinciden con las columnas del modelo."
            )
        )
        st.write(t("Model columns:", "Columnas del modelo:"), model_columns)
        st.write(
            t("Input columns:", "Columnas del input:"),
            input_encoded.columns.tolist()
        )
        st.stop()

    pred = model.predict(input_encoded)[0]
    proba = model.predict_proba(input_encoded)[0]

    prob_reject = float(proba[0])
    prob_accept = float(proba[1])

    st.progress(prob_accept)

    if pred == 1:
        st.markdown(f"""
<div class="result-card-accept">
    <h2 style="margin-bottom:0.5rem;">{t("Likely ACCEPTED", "Probablemente ACEPTADA")}</h2>
    <p style="font-size:1.1rem;">
        {t("Estimated acceptance probability", "Probabilidad estimada de aceptación")}: <strong>{prob_accept:.1%}</strong>
    </p>
</div>
""", unsafe_allow_html=True)
    else:
        st.markdown(f"""
<div class="result-card-reject">
    <h2 style="margin-bottom:0.5rem;">{t("Likely REJECTED", "Probablemente RECHAZADA")}</h2>
    <p style="font-size:1.1rem;">
        {t("Estimated rejection probability", "Probabilidad estimada de rechazo")}: <strong>{prob_reject:.1%}</strong>
    </p>
</div>
""", unsafe_allow_html=True)

    col1, col2 = st.columns(2)

    with col1:
        st.metric(
            t("Acceptance probability", "Probabilidad de aceptación"),
            f"{prob_accept:.1%}"
        )

    with col2:
        st.metric(
            t("Rejection probability", "Probabilidad de rechazo"),
            f"{prob_reject:.1%}"
        )

    st.divider()

    st.subheader(
        t(
            "Most similar historical recommendation",
            "Recomendación histórica más similar"
        )
    )

    similar_recs = find_similar_recommendations(
        input_data=input_data,
        historical_df=historical_recommendations,
        top_n=1
    )

    if similar_recs.empty or int(round(float(similar_recs["similarity"].max()))) <= 0:
        st.info(
            t(
                "No similar historical recommendations were found.",
                "No se encontraron recomendaciones históricas similares."
            )
        )
    else:
        for _, row in similar_recs.iterrows():
            outcome = (
                t("Accepted", "Aceptada")
                if int(row["Position_Binary"]) == 1
                else t("Rejected", "Rechazada")
            )

            recommending_state_html = html.escape(str(row["Recommending_State"]))
            recommendation_html = html.escape(str(row["Recommendation"]))
            macro_area_html = html.escape(str(row["Macro_Area"]))
            outcome_html = html.escape(outcome)

            card_html = "\n".join([
                '<div class="similar-card">',
                f'<h4 style="margin-bottom:0.4rem;">{recommending_state_html} · {t("Cycle", "Ciclo")} {int(row["Cycle"])}</h4>',
                f'<p style="margin-bottom:0.8rem;">“{recommendation_html}”</p>',
                '<p style="margin-bottom:0;">',
                f'<strong>{t("Outcome", "Resultado")}:</strong> {outcome_html}<br>',
                f'<strong>{t("Thematic area", "Área temática")}:</strong> {macro_area_html}<br>',
                f'<strong>{t("Alignment score", "Puntaje de alineamiento")}:</strong> {row["Aligment_Score"]:.3f}<br>',
                f'<strong>{t("Similarity", "Similitud")}:</strong> {row["similarity"]:.0f}%',
                '</p>',
                '</div>',
            ])

            st.markdown(card_html, unsafe_allow_html=True)

    st.divider()

    st.subheader(t("Scenario summary", "Resumen del escenario"))

    summary_df = pd.DataFrame({
        "Parameter": [
            t("Recommending State", "Estado recomendante"),
            t("Alignment Score", "Puntaje de alineamiento"),
            t("Verb Intensity", "Intensidad verbal"),
            t("Regime Sensitivity", "Sensibilidad del régimen"),
            t("External Scrutiny", "Escrutinio externo"),
            t("Thematic Area", "Área temática"),
            t("Behavioral Scenario", "Escenario de comportamiento"),
            t("Internal Model Cycle", "Ciclo interno del modelo")
        ],
        "Value": [
            recommending_country,
            round(float(alignment_score), 3),
            verb_intensity,
            regime_sensitivity,
            external_scrutiny,
            area,
            {
                "coin_flip": t(
                    "50/50 coin flip dynamic",
                    "Dinámica de moneda 50/50"
                ),
                "pareto_80_20": t(
                    "80/20 Pareto logic",
                    "Lógica de Pareto 80/20"
                )
            }[scenario],
            cycle
        ]
    })

    st.dataframe(
        summary_df,
        hide_index=True,
        use_container_width=True
    )

    with st.expander(
        t(
            "Technical check: encoded input sent to model",
            "Chequeo técnico: input codificado enviado al modelo"
        )
    ):
        active_area_cols = [
            col for col in input_encoded.columns
            if col.startswith("Macro_Area_")
            and input_encoded.loc[0, col] == 1
        ]

        st.write(
            t(
                "Active thematic area column:",
                "Columna activa de área temática:"
            ),
            active_area_cols
        )
        st.write(
            t("Input shape:", "Dimensión del input:"),
            input_encoded.shape
        )
        st.dataframe(input_encoded, use_container_width=True)

st.divider()

st.caption(t(
    """
⚠️ This application combines historical UPR data, thematic coding, and machine learning models to simulate possible DPRK response patterns across different political scenarios. **Outputs should be understood as probabilistic behavioral scenarios rather than definitive predictions**.

👩🏻‍💻 Developed by [María de los Ángeles Lasa](http://marialasa.ar), PhD.
""",
    """
⚠️ Esta aplicación combina datos históricos del EPU, codificación temática y modelos de aprendizaje automático para simular posibles patrones de respuesta de la RPDC en distintos escenarios políticos. **Los resultados deben entenderse como escenarios de comportamiento probabilísticos y no como predicciones definitivas**.

👩🏻‍💻 Desarrollado por [María de los Ángeles Lasa](http://marialasa.ar), PhD.
""",
))