generate_dataset.py

#!/usr/bin/env python3
"""
African Judicial Access Indicators — Synthetic Dataset Generator
================================================================
Electric Sheep Africa | electricsheepafrica on HuggingFace

Generates synthetic judicial access indicator data for sub-Saharan African
countries at the subnational (region) level, parameterized from peer-reviewed
literature, WJP indices, Afrobarometer surveys, and national reports.

PHASE 2 — Causal & Correlation Structure (DAG)
================================================
ROOT NODES (sampled independently):
  ├── country           → Categorical (weighted by population)
  ├── year              → Uniform [2018–2025]
  └── region_type       → Categorical {urban, peri_urban, rural, remote_rural}

INTERMEDIATE NODES (sampled conditionally):
  ├── population        → Conditional on country, year
  ├── lawyer_density    → Conditional on country tier, region_type [Ref 16]
  ├── court_density     → Conditional on country tier, region_type
  ├── distance_to_court → Conditional on region_type, court_density [Ref 1]
  ├── travel_time_hours → Conditional on distance, region_type
  ├── legal_aid_coverage→ Conditional on country tier, region_type [Refs 6,7]
  ├── legal_aid_per_cap → Conditional on country tier [Ref 10]
  ├── paralegal_density → Conditional on country, region_type [Ref 4]
  ├── cost_pct_income   → Conditional on region_type, country tier [Ref 2]
  ├── bribery_rate      → Conditional on country tier [Ref 9]
  └── gender_access_ratio → Conditional on country tier, region_type [Ref 12,15]

LEAF NODES (derived / looked up):
  ├── court_contact_rate → Conditional on distance, cost, trust [Ref 9]
  ├── trust_in_courts   → Conditional on bribery, country tier [Ref 9]
  ├── customary_justice_use → Conditional on region_type, legal_aid [Ref 6]
  ├── wjp_score         → Looked up by country with noise [Ref 5]
  └── access_level      → Classification: high, moderate, low, very_low

CORRELATION TARGET MATRIX (lower triangle):
              lawyer_d  court_d  distance  legal_aid  cost   bribery  trust
lawyer_dens   1.00
court_dens    0.50      1.00
distance     -0.40     -0.60     1.00
legal_aid     0.45      0.35    -0.35      1.00
cost_pct     -0.30     -0.25     0.40     -0.35      1.00
bribery      -0.35     -0.30     0.25     -0.40      0.30   1.00
trust         0.25      0.30    -0.20      0.35     -0.25  -0.60     1.00

Sources: Afrobarometer R6 [Ref 9] for trust-bribery r≈-0.6; distance-contact
         inverse from Benyawa 2023 [Ref 1]; Soboka 2019 [Ref 2] for cost barriers.

References
----------
[1]  Benyawa (2023), "How Far Are Kenya's Courts?", J. of African Law
[2]  Soboka (2019), "What does justice cost in South Africa?", SA J. Human Rights
[3]  Bilchitz & Williams (2020), PER/PELJ
[4]  Cambridge UP (2018), Community Paralegals and the Pursuit of Justice
[5]  World Justice Project, Rule of Law Index 2024/2025
[6]  UNODC (2011), Access to Legal Aid in Africa
[7]  UNDP (2014), Legal Aid Service Provision in Africa
[8]  UNODC (2014), Handbook on Legal Aid in Africa
[9]  Afrobarometer (2017), Policy Paper No. 39
[10] Legal Aid South Africa, Annual Report 2022/23
[11] HiiL (2023), Justice Needs Survey: Nigeria
[12] UN Women (2021), Access to Justice for Women ESA
[13] World Bank (2019), Cost-Benefit Analysis of Legal Aid
[14] ACCORD (2012), Abunzi Mediation in Rwanda
[15] Brookings (2022), Women and Access to Justice in Africa
[16] World Population Review (2026), Lawyers per Capita
[17] NYU CIC (2023), Front-line Justice Services Policy Brief
"""

import argparse
import os
import sys

import numpy as np
import pandas as pd

# ============================================================================
# SECTION 1: LITERATURE-INFORMED PARAMETERS
# ============================================================================

# --- 1A. Country metadata ---
# WJP scores [Ref 5], lawyer density [Ref 16], access tier
COUNTRIES = {
    "South Africa": {"pop_2024": 62.0, "wjp": 0.56, "tier": "high",     "growth": 0.008,
                     "lawyer_density": 37.04, "legal_aid_pc": 2.25},
    "Botswana":     {"pop_2024":  2.6, "wjp": 0.60, "tier": "high",     "growth": 0.015,
                     "lawyer_density": 25.0,  "legal_aid_pc": 1.50},
    "Mauritius":    {"pop_2024":  1.3, "wjp": 0.60, "tier": "high",     "growth": 0.001,
                     "lawyer_density": 45.21, "legal_aid_pc": 1.80},
    "Rwanda":       {"pop_2024": 14.0, "wjp": 0.63, "tier": "high",     "growth": 0.024,
                     "lawyer_density": 10.0,  "legal_aid_pc": 0.40},
    "Kenya":        {"pop_2024": 55.0, "wjp": 0.45, "tier": "moderate", "growth": 0.021,
                     "lawyer_density": 15.82, "legal_aid_pc": 0.02},
    "Ghana":        {"pop_2024": 34.0, "wjp": 0.54, "tier": "moderate", "growth": 0.021,
                     "lawyer_density":  7.26, "legal_aid_pc": 0.03},
    "Tanzania":     {"pop_2024": 67.0, "wjp": 0.46, "tier": "moderate", "growth": 0.030,
                     "lawyer_density":  4.0,  "legal_aid_pc": 0.05},
    "Senegal":      {"pop_2024": 18.0, "wjp": 0.56, "tier": "moderate", "growth": 0.027,
                     "lawyer_density":  6.5,  "legal_aid_pc": 0.04},
    "Nigeria":      {"pop_2024": 230.0,"wjp": 0.41, "tier": "low",      "growth": 0.025,
                     "lawyer_density": 11.0,  "legal_aid_pc": 0.01},
    "DRC":          {"pop_2024": 105.0,"wjp": 0.34, "tier": "low",      "growth": 0.032,
                     "lawyer_density": 13.93, "legal_aid_pc": 0.005},
    "Uganda":       {"pop_2024": 49.0, "wjp": 0.38, "tier": "low",      "growth": 0.032,
                     "lawyer_density":  5.0,  "legal_aid_pc": 0.02},
    "Malawi":       {"pop_2024": 21.0, "wjp": 0.52, "tier": "low",      "growth": 0.026,
                     "lawyer_density":  2.5,  "legal_aid_pc": 0.02},
    "Mozambique":   {"pop_2024": 33.0, "wjp": 0.37, "tier": "low",      "growth": 0.028,
                     "lawyer_density":  2.0,  "legal_aid_pc": 0.005},
}

# --- 1B. Region type distribution ---
# SSA: ~35-40% urban, 60-65% rural [World Bank]
REGION_TYPES = {
    "urban":        0.20,
    "peri_urban":   0.18,
    "rural":        0.42,
    "remote_rural": 0.20,
}

# --- 1C. Distance to court by region type (km) [Ref 1: Kenya avg 22km] ---
DISTANCE_KM = {
    "urban":        {"mean":  5, "sd": 3,  "min": 0.5, "max": 20},
    "peri_urban":   {"mean": 15, "sd": 8,  "min": 3,   "max": 40},
    "rural":        {"mean": 40, "sd": 20, "min": 10,  "max": 100},
    "remote_rural": {"mean": 90, "sd": 40, "min": 30,  "max": 250},
}

# Tier multiplier for distance [low-access countries have worse infrastructure]
DISTANCE_TIER_MULT = {"high": 0.7, "moderate": 1.0, "low": 1.4}

# --- 1D. Travel time model: time = distance / speed + overhead ---
# Speed varies by region type and infrastructure quality
TRAVEL_SPEED_KPH = {
    "urban":        {"mean": 20, "sd": 5},   # congested, public transport
    "peri_urban":   {"mean": 25, "sd": 8},
    "rural":        {"mean": 15, "sd": 5},   # poor roads, mixed transport
    "remote_rural": {"mean": 10, "sd": 4},   # walking, motorbike, boat
}
TRAVEL_OVERHEAD_HOURS = {
    "urban": 0.2, "peri_urban": 0.3, "rural": 0.5, "remote_rural": 1.0,
}

# --- 1E. Lawyer density by region type [Ref 16: >80% concentrated in capitals] ---
# Multiplier applied to country's national lawyer_density
LAWYER_REGION_MULT = {
    "urban":        2.5,   # Capital/city concentration [Ref 16]
    "peri_urban":   1.0,
    "rural":        0.2,
    "remote_rural": 0.05,  # Near zero in remote areas
}

# --- 1F. Court density per 100K by tier [Refs from Phase 1] ---
# SA: ~1.17; Tanzania: ~1.2-1.5; Rwanda: ~3.0 (formal); Kenya: ~0.22; DRC: <0.10
COURT_DENSITY = {
    "high":     {"mean": 2.0, "sd": 0.8, "min": 0.5, "max": 5.0},
    "moderate": {"mean": 0.8, "sd": 0.4, "min": 0.1, "max": 2.0},
    "low":      {"mean": 0.3, "sd": 0.2, "min": 0.05, "max": 1.0},
}
COURT_REGION_MULT = {
    "urban": 2.0, "peri_urban": 1.2, "rural": 0.5, "remote_rural": 0.15,
}

# --- 1G. Legal aid coverage % [Refs 6, 7, 10] ---
LEGAL_AID_COVERAGE = {
    "high":     {"mean": 20, "sd": 8,  "min": 5,  "max": 40},
    "moderate": {"mean":  7, "sd": 3,  "min": 2,  "max": 15},
    "low":      {"mean":  3, "sd": 2,  "min": 0.5,"max": 8},
}
LEGAL_AID_REGION_MULT = {
    "urban": 2.0, "peri_urban": 1.0, "rural": 0.4, "remote_rural": 0.1,
}

# --- 1H. Paralegal / community justice density per 100K [Refs 4, 14, 17] ---
# Rwanda: 270/100K (Abunzi); SA: ~1-2; Uganda: ~1-2; Kenya: ~1-3
PARALEGAL_DENSITY = {
    "high":     {"mean": 30, "sd": 60, "min": 1, "max": 300},  # Rwanda pulls mean up
    "moderate": {"mean":  3, "sd": 2,  "min": 0.5, "max": 10},
    "low":      {"mean":  1, "sd": 1,  "min": 0.1, "max": 5},
}

# --- 1I. Cost of justice as % monthly income [Ref 2: SA 250% for poor] ---
COST_PCT_INCOME = {
    "urban":        {"mean": 60,  "sd": 40,  "min": 10, "max": 250},
    "peri_urban":   {"mean": 100, "sd": 50,  "min": 15, "max": 350},
    "rural":        {"mean": 150, "sd": 70,  "min": 25, "max": 450},
    "remote_rural": {"mean": 200, "sd": 90,  "min": 40, "max": 550},
}
COST_TIER_MULT = {"high": 0.6, "moderate": 1.0, "low": 1.4}

# --- 1J. Bribery rate % [Ref 9: Afrobarometer 30% SSA avg] ---
BRIBERY_RATE = {
    "high":     {"mean": 15, "sd": 5,  "min": 5,  "max": 30},
    "moderate": {"mean": 28, "sd": 8,  "min": 10, "max": 45},
    "low":      {"mean": 40, "sd": 10, "min": 15, "max": 60},
}

# --- 1K. Trust in courts % [Ref 9: SSA avg 53%; Tanzania 88%; Southern Africa 25%] ---
TRUST_COURTS = {
    "high":     {"mean": 70, "sd": 12, "min": 40, "max": 95},
    "moderate": {"mean": 50, "sd": 12, "min": 25, "max": 75},
    "low":      {"mean": 35, "sd": 10, "min": 15, "max": 55},
}

# --- 1L. Gender access ratio (women/men) [Refs 12, 15: 64% globally] ---
GENDER_ACCESS = {
    "high":     {"mean": 0.85, "sd": 0.06, "min": 0.65, "max": 0.98},
    "moderate": {"mean": 0.72, "sd": 0.08, "min": 0.50, "max": 0.90},
    "low":      {"mean": 0.58, "sd": 0.08, "min": 0.40, "max": 0.78},
}
GENDER_REGION_MULT = {
    "urban": 1.10, "peri_urban": 1.0, "rural": 0.90, "remote_rural": 0.80,
}

# --- 1M. Customary justice use % [Ref 6: Malawi 80-90% rural] ---
CUSTOMARY_USE = {
    "urban":        {"mean": 15, "sd": 8,  "min": 2,  "max": 35},
    "peri_urban":   {"mean": 30, "sd": 12, "min": 10, "max": 55},
    "rural":        {"mean": 60, "sd": 15, "min": 25, "max": 85},
    "remote_rural": {"mean": 80, "sd": 10, "min": 50, "max": 98},
}

# --- 1N. Access level classification ---
# Composite score = weighted combination of indicators
# Thresholds calibrated so baseline produces ~15-25% each tier
ACCESS_LEVELS = {
    "high":      (0.55, 1.01),
    "moderate":  (0.38, 0.55),
    "low":       (0.22, 0.38),
    "very_low":  (0.00, 0.22),
}

# --- 1O. Scenario definitions ---
SCENARIOS = {
    "baseline": {
        "description": "Current SSA judicial access landscape (2018-2025)",
        "distance_mult": 1.0, "lawyer_mult": 1.0, "legal_aid_mult": 1.0,
        "cost_mult": 1.0, "paralegal_mult": 1.0,
    },
    "improved_access": {
        "description": "Countries investing in justice reform and legal aid expansion",
        "distance_mult": 0.8, "lawyer_mult": 1.3, "legal_aid_mult": 1.5,
        "cost_mult": 0.7, "paralegal_mult": 2.0,
    },
    "constrained": {
        "description": "Fiscal austerity / conflict reducing access to justice",
        "distance_mult": 1.2, "lawyer_mult": 0.8, "legal_aid_mult": 0.5,
        "cost_mult": 1.5, "paralegal_mult": 0.6,
    },
}

# WJP annual noise [Ref 5]
WJP_ANNUAL_NOISE_SD = 0.015


# ============================================================================
# SECTION 2: UTILITY FUNCTIONS
# ============================================================================

def trunc_normal(rng, mean, sd, low, high, size=1):
    """Sample from truncated normal distribution."""
    from scipy.stats import truncnorm
    if sd <= 0:
        return np.full(size, mean)
    a = (low - mean) / sd
    b = (high - mean) / sd
    return truncnorm.rvs(a, b, loc=mean, scale=sd, size=size, random_state=rng)


def sample_categorical(rng, categories, probabilities, size=1):
    """Weighted categorical sampling."""
    probs = np.array(probabilities, dtype=float)
    probs /= probs.sum()
    return rng.choice(categories, size=size, p=probs)


def classify_access(score):
    """Classify access level from composite score."""
    for label, (lo, hi) in ACCESS_LEVELS.items():
        if lo <= score < hi:
            return label
    return "very_low"


# ============================================================================
# SECTION 3: MAIN GENERATOR FUNCTION
# ============================================================================

def generate_judicial_access_data(scenario_name="baseline", n=10000, seed=42):
    """
    Generate synthetic African judicial access indicator records.

    SAMPLING ORDER (topological sort of DAG):
    ─────────────────────────────────────────
    1.  [Root]         → sample country
    2.  [Root]         → sample year
    3.  [Root]         → sample region_type
    4.  [Intermediate] → sample lawyer_density | country, region
    5.  [Intermediate] → sample court_density | tier, region
    6.  [Intermediate] → sample distance_to_court | region, tier
    7.  [Intermediate] → derive travel_time | distance, region
    8.  [Intermediate] → sample legal_aid_coverage | tier, region
    9.  [Intermediate] → lookup legal_aid_per_capita | country
    10. [Intermediate] → sample paralegal_density | tier, country
    11. [Intermediate] → sample cost_pct_income | region, tier
    12. [Intermediate] → sample bribery_rate | tier
    13. [Intermediate] → sample gender_access_ratio | tier, region
    14. [Intermediate] → sample customary_justice_use | region
    15. [Leaf]         → derive trust_in_courts | bribery, tier
    16. [Leaf]         → derive court_contact_rate | distance, cost, trust
    17. [Leaf]         → lookup wjp_score | country, year
    18. [Leaf]         → derive access_score & classify access_level
    ─────────────────────────────────────────
    """
    rng = np.random.default_rng(seed)
    scenario = SCENARIOS[scenario_name]

    # --- Root nodes ---
    country_names = list(COUNTRIES.keys())
    country_pops = np.array([COUNTRIES[c]["pop_2024"] for c in country_names])
    countries = sample_categorical(rng, country_names, country_pops / country_pops.sum(), size=n)
    years = rng.integers(2018, 2026, size=n)
    region_names = list(REGION_TYPES.keys())
    region_probs = list(REGION_TYPES.values())
    regions = sample_categorical(rng, region_names, region_probs, size=n)

    # Pre-allocate
    cols = {}
    lawyer_density = np.zeros(n)
    court_density = np.zeros(n)
    distance_km = np.zeros(n)
    travel_time = np.zeros(n)
    legal_aid_cov = np.zeros(n)
    legal_aid_pc = np.zeros(n)
    paralegal_dens = np.zeros(n)
    cost_pct = np.zeros(n)
    bribery = np.zeros(n)
    gender_ratio = np.zeros(n)
    customary_use = np.zeros(n)
    trust = np.zeros(n)
    contact_rate = np.zeros(n)
    wjp = np.zeros(n)
    access_score = np.zeros(n)
    access_level = np.empty(n, dtype=object)

    for i in range(n):
        c = countries[i]
        r = regions[i]
        meta = COUNTRIES[c]
        tier = meta["tier"]

        # Step 4: Lawyer density | country, region [Ref 16]
        base_ld = meta["lawyer_density"] * scenario["lawyer_mult"]
        ld = base_ld * LAWYER_REGION_MULT[r]
        ld += rng.normal(0, ld * 0.2)  # 20% noise
        lawyer_density[i] = round(np.clip(ld, 0.01, 150), 2)

        # Step 5: Court density | tier, region
        cd_params = COURT_DENSITY[tier]
        cd = trunc_normal(rng, cd_params["mean"], cd_params["sd"],
                          cd_params["min"], cd_params["max"])[0]
        cd *= COURT_REGION_MULT[r]
        court_density[i] = round(np.clip(cd, 0.01, 15), 3)

        # Step 6: Distance to court | region, tier [Ref 1]
        d_params = DISTANCE_KM[r]
        d = trunc_normal(rng, d_params["mean"] * DISTANCE_TIER_MULT[tier] * scenario["distance_mult"],
                         d_params["sd"], d_params["min"], d_params["max"])[0]
        # Inverse relationship with court density
        if court_density[i] > 0.5:
            d *= 0.8
        distance_km[i] = round(np.clip(d, 0.5, 300), 1)

        # Step 7: Travel time | distance, region
        spd_params = TRAVEL_SPEED_KPH[r]
        speed = max(rng.normal(spd_params["mean"], spd_params["sd"]), 2)
        overhead = TRAVEL_OVERHEAD_HOURS[r]
        tt = distance_km[i] / speed + overhead
        travel_time[i] = round(np.clip(tt, 0.1, 48), 2)

        # Step 8: Legal aid coverage | tier, region [Refs 6, 7]
        la_params = LEGAL_AID_COVERAGE[tier]
        la = trunc_normal(rng, la_params["mean"] * scenario["legal_aid_mult"],
                          la_params["sd"], la_params["min"],
                          la_params["max"] * scenario["legal_aid_mult"])[0]
        la *= LEGAL_AID_REGION_MULT[r]
        legal_aid_cov[i] = round(np.clip(la, 0.1, 60), 1)

        # Step 9: Legal aid per capita | country [Ref 10]
        la_pc = meta["legal_aid_pc"] * scenario["legal_aid_mult"]
        la_pc *= rng.lognormal(0, 0.3)
        legal_aid_pc[i] = round(np.clip(la_pc, 0.001, 10.0), 3)

        # Step 10: Paralegal density | tier [Refs 4, 14, 17]
        pl_params = PARALEGAL_DENSITY[tier]
        # Rwanda special case: Abunzi system [Ref 14]
        if c == "Rwanda":
            pl = trunc_normal(rng, 270 * scenario["paralegal_mult"], 50, 100, 400)[0]
        else:
            pl = trunc_normal(rng, pl_params["mean"] * scenario["paralegal_mult"],
                              pl_params["sd"], pl_params["min"], pl_params["max"])[0]
        paralegal_dens[i] = round(np.clip(pl, 0.05, 400), 1)

        # Step 11: Cost as % monthly income | region, tier [Ref 2]
        cost_params = COST_PCT_INCOME[r]
        co = trunc_normal(rng, cost_params["mean"] * COST_TIER_MULT[tier] * scenario["cost_mult"],
                          cost_params["sd"], cost_params["min"], cost_params["max"])[0]
        cost_pct[i] = round(np.clip(co, 5, 700), 0)

        # Step 12: Bribery rate | tier [Ref 9]
        br_params = BRIBERY_RATE[tier]
        br = trunc_normal(rng, br_params["mean"], br_params["sd"],
                          br_params["min"], br_params["max"])[0]
        bribery[i] = round(np.clip(br, 1, 70), 1)

        # Step 13: Gender access ratio | tier, region [Refs 12, 15]
        ga_params = GENDER_ACCESS[tier]
        ga = trunc_normal(rng, ga_params["mean"], ga_params["sd"],
                          ga_params["min"], ga_params["max"])[0]
        ga *= GENDER_REGION_MULT[r]
        gender_ratio[i] = round(np.clip(ga, 0.30, 1.0), 3)

        # Step 14: Customary justice use | region [Ref 6]
        cj_params = CUSTOMARY_USE[r]
        cj = trunc_normal(rng, cj_params["mean"], cj_params["sd"],
                          cj_params["min"], cj_params["max"])[0]
        customary_use[i] = round(np.clip(cj, 1, 99), 1)

        # Step 15: Trust in courts | bribery, tier [Ref 9]
        # Trust inversely related to bribery: r ≈ -0.6
        tr_params = TRUST_COURTS[tier]
        tr_base = trunc_normal(rng, tr_params["mean"], tr_params["sd"],
                               tr_params["min"], tr_params["max"])[0]
        # Inject bribery effect: higher bribery → lower trust
        bribery_effect = -0.6 * (bribery[i] - 30)  # centered on SSA avg 30%
        tr = tr_base + bribery_effect
        trust[i] = round(np.clip(tr, 10, 98), 1)

        # Step 16: Court contact rate | distance, cost, trust [Ref 9]
        # SSA avg: 13%; range 4-28%. Weaker effects + more noise to avoid
        # overly mechanical correlations.
        base_contact = 13.0
        # Distance penalty: each 10km above 20 reduces by 0.3pp
        dist_effect = -0.3 * max(0, (distance_km[i] - 20)) / 10
        # Cost penalty: each 50% above 100% reduces by 0.5pp
        cost_effect = -0.5 * max(0, (cost_pct[i] - 100)) / 50
        # Trust bonus: each 10pp above 50% adds 0.6pp
        trust_effect = 0.6 * (trust[i] - 50) / 10
        cr = base_contact + dist_effect + cost_effect + trust_effect
        cr += rng.normal(0, 5)  # larger noise to decouple
        contact_rate[i] = round(np.clip(cr, 1, 40), 1)

        # Step 17: WJP score | country, year [Ref 5]
        base_wjp = meta["wjp"]
        years_to_2025 = 2025 - years[i]
        wjp_noise = rng.normal(0, WJP_ANNUAL_NOISE_SD * np.sqrt(years_to_2025 + 1))
        wjp[i] = round(np.clip(base_wjp + wjp_noise, 0.15, 0.90), 3)

        # Step 18: Access score & classification
        # Composite: normalize each indicator to [0,1] and weight
        ld_norm = np.clip(lawyer_density[i] / 50, 0, 1)          # 50/100K = best
        cd_norm = np.clip(court_density[i] / 3.0, 0, 1)          # 3/100K = best
        dist_norm = 1 - np.clip(distance_km[i] / 150, 0, 1)      # 0km = best
        la_norm = np.clip(legal_aid_cov[i] / 30, 0, 1)            # 30% = best
        cost_norm = 1 - np.clip(cost_pct[i] / 500, 0, 1)          # 0% = best
        trust_norm = np.clip(trust[i] / 100, 0, 1)                # 100% = best
        bribe_norm = 1 - np.clip(bribery[i] / 60, 0, 1)           # 0% = best
        gender_norm = np.clip(gender_ratio[i], 0, 1)

        score = (0.15 * ld_norm + 0.15 * cd_norm + 0.15 * dist_norm +
                 0.15 * la_norm + 0.10 * cost_norm + 0.10 * trust_norm +
                 0.10 * bribe_norm + 0.10 * gender_norm)
        access_score[i] = round(score, 3)
        access_level[i] = classify_access(score)

    # --- Assemble DataFrame ---
    df = pd.DataFrame({
        "record_id": np.arange(1, n + 1),
        "country": countries,
        "year": years,
        "region_type": regions,
        "lawyer_density_per_100k": lawyer_density,
        "court_density_per_100k": court_density,
        "distance_to_court_km": distance_km,
        "travel_time_hours": travel_time,
        "legal_aid_coverage_pct": legal_aid_cov,
        "legal_aid_per_capita_usd": legal_aid_pc,
        "paralegal_density_per_100k": paralegal_dens,
        "cost_pct_monthly_income": cost_pct,
        "bribery_rate_pct": bribery,
        "trust_in_courts_pct": trust,
        "gender_access_ratio": gender_ratio,
        "customary_justice_use_pct": customary_use,
        "court_contact_rate_pct": contact_rate,
        "wjp_rule_of_law_score": wjp,
        "access_score": access_score,
        "access_level": access_level,
    })

    # --- Print summary ---
    print(f"\n{'='*70}")
    print(f"AFRICAN JUDICIAL ACCESS INDICATORS — GENERATION SUMMARY")
    print(f"Scenario: {scenario_name} | N: {n:,} | Seed: {seed}")
    print(f"{'='*70}")
    print(f"\nCountry distribution:")
    print(df["country"].value_counts().to_string())
    print(f"\nRegion type distribution:")
    print(df["region_type"].value_counts(normalize=True).mul(100).round(1).to_string())
    print(f"\nAccess level distribution:")
    print(df["access_level"].value_counts(normalize=True).mul(100).round(1).to_string())
    print(f"\nKey statistics:")
    for col in ["lawyer_density_per_100k", "distance_to_court_km", "travel_time_hours",
                 "legal_aid_coverage_pct", "cost_pct_monthly_income", "bribery_rate_pct",
                 "trust_in_courts_pct", "gender_access_ratio", "court_contact_rate_pct",
                 "access_score"]:
        print(f"  {col}: mean={df[col].mean():.2f}, sd={df[col].std():.2f}, "
              f"min={df[col].min():.2f}, max={df[col].max():.2f}")
    print(f"{'='*70}\n")

    return df


# ============================================================================
# SECTION 4: CLI ENTRY POINT
# ============================================================================

def main():
    parser = argparse.ArgumentParser(
        description="Generate synthetic African judicial access indicators dataset"
    )
    parser.add_argument("--scenario", choices=list(SCENARIOS.keys()), default="baseline")
    parser.add_argument("--n", type=int, default=10000)
    parser.add_argument("--seed", type=int, default=42)
    parser.add_argument("--output", type=str, default=None)
    args = parser.parse_args()

    df = generate_judicial_access_data(args.scenario, args.n, args.seed)

    if args.output is None:
        script_dir = os.path.dirname(os.path.abspath(__file__))
        output_dir = os.path.join(script_dir, "data")
        os.makedirs(output_dir, exist_ok=True)
        output_path = os.path.join(output_dir, f"{args.scenario}.csv")
    else:
        output_path = args.output
        os.makedirs(os.path.dirname(output_path) or ".", exist_ok=True)

    df.to_csv(output_path, index=False)
    print(f"Dataset saved to: {output_path}")
    print(f"Shape: {df.shape}")


if __name__ == "__main__":
    main()