mirasai/media-profiling

                         MEDIA PROFILER - COMPLETE PIPELINE
================================================================================

                                   INPUT
                                     │
                                     ▼
┌─────────────────────────────────────────────────────────────────────────────┐
│  1. SCRAPE NODE (scraper.py)                                                │
│     ┌─────────────────────────────────────────────────────────────────┐     │
│     │  MediaScraper                                                   │     │
│     │  ├── Fetch homepage + sitemap                                   │     │
│     │  ├── Collect up to 20 articles                                  │     │
│     │  ├── Separate News vs Opinion articles                          │     │
│     │  │   ├── URL patterns (/opinion/, /editorial/)                  │     │
│     │  │   ├── Schema.org metadata                                    │     │
│     │  │   └── Title patterns ("Opinion:", "Editorial:")              │     │
│     │  └── Extract site metadata                                      │     │
│     │       ├── About page (ownership, funding)                       │     │
│     │       ├── Author information                                    │     │
│     │       └── Location disclosure                                   │     │
│     └─────────────────────────────────────────────────────────────────┘     │
└─────────────────────────────────────────────────────────────────────────────┘
                                     │
                                     ▼
┌─────────────────────────────────────────────────────────────────────────────┐
│  2. ANALYZE NODE (refactored_analyzers.py + research.py)                    │
│                                                                             │
│     ┌───────────────────────────┐   ┌───────────────────────────┐           │
│     │    CONTENT ANALYZERS      │   │   METADATA ANALYZERS      │           │
│     ├───────────────────────────┤   ├───────────────────────────┤           │
│     │ EditorialBiasAnalyzer     │   │ TrafficLongevityAnalyzer  │           │
│     │ PseudoscienceAnalyzer     │   │ MediaTypeAnalyzer         │           │
│     │ SourcingAnalyzer          │   │ OpinionAnalyzer           │           │
│     │ FactCheckSearcher         │   │                           │           │
│     │ TransparencyAnalyzer      │   │                           │           │
│     │ OneSidednessAnalyzer      │   │                           │           │
│     └───────────────────────────┘   └───────────────────────────┘           │
│                                                                             │
│     ┌───────────────────────────────────────────────────────────────────┐   │
│     │  WEB RESEARCH (MediaResearcher) — pluggable search backend       │   │
│     │  ├── History research (founding, key events)                      │   │
│     │  ├── Ownership research (owner, funding, headquarters)            │   │
│     │  ├── External analysis (MBFC, NewsGuard, academic reviews)        │   │
│     │  └── Batch search support (combines queries for efficiency)       │   │
│     └───────────────────────────────────────────────────────────────────┘   │
│                                                                             │
│     ┌───────────────────────────────────────────────────────────────────┐   │
│     │  FREEDOM RATING LOOKUP                                           │   │
│     │  ├── RSF Press Freedom Index (2025.csv)                          │   │
│     │  └── Freedom House ratings → credibility penalty                 │   │
│     └───────────────────────────────────────────────────────────────────┘   │
└─────────────────────────────────────────────────────────────────────────────┘
                                     │
                                     ▼
┌─────────────────────────────────────────────────────────────────────────────┐
│  3. REPORT NODE (research.py + report_generator.py)                         │
│     ┌─────────────────────────────────────────────────────────────────┐     │
│     │  MediaProfiler                                                  │     │
│     │  ├── Calculate bias score (4 weighted subscores)                │     │
│     │  ├── Calculate factuality score (4 weighted components)         │     │
│     │  ├── Look up press freedom rating                               │     │
│     │  ├── Calculate credibility (point-based + exceptions)           │     │
│     │  └── Generate comprehensive MBFC-style report with evidence     │     │
│     └─────────────────────────────────────────────────────────────────┘     │
└─────────────────────────────────────────────────────────────────────────────┘
                                     │
                                     ▼
                                  OUTPUT

+-----------------------------------------------------------------------------+
|  BIAS SCORING (-10 to +10)             |  FACTUALITY SCORING (0-10)          |
|----------------------------------------|-------------------------------------|
|                                        |                                     |
|  EditorialBiasAnalyzer                 |  FactCheckSearcher          (40%)  |
|  4 weighted subscores:                 |  IFCN-approved fact-checker search |
|  - Economic Policy       (35%)         |                                     |
|  - Social Issues         (35%)         |  SourcingAnalyzer           (25%)  |
|  - News Reporting Style  (15%)         |  Link extraction + LLM quality     |
|  - Editorial Positions   (15%)         |                                     |
|                                        |  TransparencyAnalyzer       (25%)  |
|  -10 Left -> +10 Right                 |  About page, ownership, funding    |
|  7 labels: Extreme Left to             |                                     |
|            Extreme Right               |  OneSidednessAnalyzer       (10%)  |
|                                        |  Propaganda / one-sided reporting  |
+-----------------------------------------------------------------------------+
|                                                                             |
|                        CREDIBILITY (Point-Based)                            |
|  Factual (0-4 pts) + Bias (0-3 pts) + Traffic (0-2 pts)                    |
|  + Longevity bonus (1 pt if ≥10 years) + Freedom penalty (0 to -2 pts)     |
|  3 labels: High (≥7) / Medium (4-6) / Low (≤3)                             |
|  Exceptions: Mixed/Low/Very Low factuality → automatic Low                 |
|              Mostly Factual (score 3.6-4.5) → automatic Medium             |
|                                                                             |
+-----------------------------------------------------------------------------+

factuality_score = FailedFactChecks × 0.40 + Sourcing × 0.25 + Transparency × 0.25 + OneSidedness × 0.10

total_points = factual + bias + traffic + longevity + freedom_penalty

analyze(domain)
    │
    ├─► 1. WHOIS Lookup (always runs)
    │       └─► Extract creation_date → Calculate age_years
    │
    ├─► 2. Tranco Lookup (O(1) dict lookup)
    │       │   Source: https://tranco-list.eu/
    │       │   - Top 1M domains ranked by popularity
    │       │   - Auto-downloads if missing
    │       │
    │       ├─► Found?
    │       │       ├─► rank < 10,000    → HIGH traffic
    │       │       ├─► rank < 100,000   → MEDIUM traffic
    │       │       ├─► rank < 1,000,000 → LOW traffic
    │       │       └─► Return with confidence=1.0, source=TRANCO
    │       │
    │       └─► Not found? → Continue to step 3
    │
    └─► 3. LLM Fallback (via search backend)
            ├─► Search: "{domain} traffic stats similarweb hypestat semrush"
            ├─► Combine top 5 result snippets
            └─► Parse with structured LLM output → TrafficEstimate
                    ├─► traffic_tier: HIGH/MEDIUM/LOW/MINIMAL/UNKNOWN
                    ├─► monthly_visits_estimate (if found)
                    ├─► confidence: 0.0-1.0
                    └─► reasoning

OUTPUT: TrafficData
    ├── domain, creation_date, age_years
    ├── traffic_tier, traffic_confidence
    ├── traffic_source: TRANCO | LLM | FALLBACK
    ├── tranco_rank (if available)
    └── whois_success, whois_error

analyze(url_or_domain, outlet_name?)
    │
    ├─► 1. Extract Domain & Outlet Name
    │       ├─► "nytimes.com" → "New York Times"
    │       └─► Uses known_names dict or generates from domain
    │
    ├─► 2. Search 5 Fact-Checker Sites (via search backend)
    │       │   Sites:
    │       │   ├── politifact.com
    │       │   ├── snopes.com
    │       │   ├── factcheck.org
    │       │   ├── fullfact.org
    │       │   ├── reuters.com/fact-check
    │       │   ├── apnews.com/ap-fact-check
    │       │   └── leadstories.com
    │       │
    │       │   Query format:
    │       │   site:{site} "{domain}" OR "{outlet_name}"
    │       │
    │       └─► Collect up to 3 results per site → Combine snippets
    │
    ├─► 3. LLM Parsing
    │       └─► Parse snippets → FactCheckLLMOutput
    │               ├─► findings: List[FactCheckFinding]
    │               │       ├── source_site (PolitiFact, Snopes, etc.)
    │               │       ├── claim_summary
    │               │       ├── verdict: TRUE/MOSTLY_TRUE/HALF_TRUE/MIXED/
    │               │       │            MOSTLY_FALSE/FALSE/PANTS_ON_FIRE/
    │               │       │            MISLEADING/UNPROVEN/NOT_RATED
    │               │       └── url (if available)
    │               ├─► failed_count (FALSE, MOSTLY_FALSE, PANTS_ON_FIRE, MISLEADING)
    │               ├─► total_count
    │               └─► confidence, reasoning
    │
    └─► 4. Score Calculation (MBFC 2025: linear 1:1 mapping)
            ├─► 0 failed checks    → 0.0
            ├─► 1 failed check     → 1.0
            ├─► 2 failed checks    → 2.0
            ├─► N failed checks    → N (capped at 10.0)
            └─► No data found      → 0.0 (with sufficient_coverage=False flag)

OUTPUT: FactCheckAnalysisResult
    ├── domain, outlet_name
    ├── failed_checks_count, total_checks_count
    ├── score: 0.0-10.0
    ├── source: SEARCH | FALLBACK
    ├── findings: List[FactCheckFinding]
    └── confidence, reasoning

analyze(articles: List[{text}])
    │
    ├─► 1. Extract Links from All Articles
    │       └─► Regex: https?://[^\s<>"')\]]+
    │
    ├─► 2. Extract Unique Domains
    │       │   Filter out social media:
    │       │   ├── twitter.com, x.com
    │       │   ├── facebook.com, instagram.com
    │       │   ├── youtube.com, tiktok.com
    │       │   ├── linkedin.com, reddit.com
    │       │   └── t.co (Twitter short links)
    │       │
    │       └─► No domains found? → Return score=5.0 (neutral)
    │
    └─► 3. LLM Quality Assessment
            └─► Assess each domain → SourcingLLMOutput
                    ├─► sources_assessed: List[SourceAssessment]
                    │       ├── domain
                    │       ├── quality: PRIMARY/WIRE_SERVICE/MAJOR_OUTLET/
                    │       │            CREDIBLE/UNKNOWN/QUESTIONABLE
                    │       └── reasoning
                    ├─► overall_quality_score: 0.0-10.0
                    ├─► has_primary_sources: bool
                    ├─► has_wire_services: bool
                    └─► overall_assessment

Quality Tiers:
    PRIMARY       → .gov, .edu, official sources, research papers
    WIRE_SERVICE  → Reuters, AP, AFP, UPI
    MAJOR_OUTLET  → NYT, BBC, WSJ, WaPo, Guardian, CNN
    CREDIBLE      → Regional papers, trade publications
    UNKNOWN       → Unfamiliar domains
    QUESTIONABLE  → Known unreliable sources

OUTPUT: SourcingAnalysisResult
    ├── score: 0.0-10.0 (0=excellent, 10=poor)
    ├── avg_sources_per_article
    ├── total_sources_found, unique_domains
    ├── has_hyperlinks, has_primary_sources, has_wire_services
    ├── source_assessments: List[SourceAssessment]
    └── confidence, reasoning

analyze(articles: List[{title, text}], url_or_domain?, outlet_name?)
    │
    ├─► 1. Format Articles for Analysis
    │       └─► Combine title + first 2000 chars of each article
    │
    └─► 2. LLM Analysis with MBFC Methodology
            │
            │   4 Weighted Subscores:
            │   ├── Economic Policy (35%): taxes, regulation, unions, trade
            │   ├── Social Issues (35%): abortion, LGBTQ+, guns, immigration
            │   ├── News Reporting Style (15%): loaded language, story selection
            │   └── Editorial Positions (15%): editorial page leaning
            │
            │   Additional Detection:
            │   ├── Loaded Language:
            │   │       ├── LEFT: "regime", "far-right", "fascist", "climate denier"
            │   │       └── RIGHT: "radical left", "woke", "cancel culture", "fake news"
            │   └── Story Selection Bias patterns
            │
            └─► Parse → EditorialBiasLLMOutput
                    ├─► overall_bias: EXTREME_LEFT/LEFT/LEFT_CENTER/CENTER/
                    │                 RIGHT_CENTER/RIGHT/EXTREME_RIGHT
                    ├─► bias_score: -10.0 to +10.0
                    ├─► economic_score, social_score, news_reporting_score, editorial_score
                    ├─► policy_positions: List[PolicyPosition]
                    ├─► uses_loaded_language: bool
                    ├─► loaded_language_examples: List[str]
                    └─► confidence, reasoning

OUTPUT: EditorialBiasResult
    ├── domain, outlet_name
    ├── overall_bias: BiasDirection
    ├── bias_score: -10.0 to +10.0
    ├── economic_score, social_score, news_reporting_score, editorial_score
    ├── mbfc_label: "Extreme Left"/"Left"/"Left-Center"/"Least Biased"/"Right-Center"/"Right"/"Extreme Right"
    ├── policy_positions, loaded_language_examples
    ├── articles_analyzed
    └── confidence, reasoning

analyze(articles, url_or_domain?, outlet_name?, site_metadata?)
    │
    └─► LLM Analysis → TransparencyLLMOutput
            ├─► has_about_page: bool
            ├─► discloses_ownership: bool
            ├─► discloses_funding: bool
            ├─► discloses_location: bool
            ├─► identifies_authors: bool
            ├─► score: 0.0-10.0 (0=fully transparent, 10=no transparency)
            └─► confidence, reasoning

OUTPUT: TransparencyResult
    ├── score: 0.0-10.0
    ├── transparency indicators (about page, ownership, funding, location, authors)
    └── confidence, reasoning

analyze(articles, url_or_domain?, outlet_name?)
    │
    └─► LLM Analysis → OneSidednessLLMOutput
            ├─► one_sidedness_score: 0.0-10.0 (0=balanced, 10=extreme propaganda)
            ├─► propaganda_techniques_found: List[str]
            └─► confidence, reasoning

OUTPUT: OneSidednessResult
    ├── score: 0.0-10.0
    ├── propaganda_techniques_found
    └── confidence, reasoning

analyze(articles: List[{title, text}], url_or_domain?, outlet_name?)
    │
    ├─► 1. Format Articles for Analysis
    │       └─► Combine title + first 2000 chars of each article
    │
    └─► 2. LLM Analysis with Scientific Consensus
            │
            │   Categories to Detect:
            │   HEALTH: Anti-Vaccination, Alternative Medicine, COVID-19 Misinfo
            │   CLIMATE: Climate Denialism, 5G Conspiracy, Chemtrails, GMO Claims
            │   PARANORMAL: Astrology, Psychic Claims, Faith Healing
            │   CONSPIRACY: Flat Earth, Moon Landing Hoax, QAnon
            │
            │   Severity: PROMOTES / PRESENTS_UNCRITICALLY / MIXED / NONE_DETECTED
            │
            └─► Parse → PseudoscienceLLMOutput
                    ├─► indicators: List[PseudoscienceIndicator]
                    ├─► promotes_pseudoscience: bool
                    ├─► overall_severity: PseudoscienceSeverity
                    ├─► respects_scientific_consensus: bool
                    └─► confidence, reasoning

OUTPUT: PseudoscienceAnalysisResult
    ├── score: 0.0-10.0 (0=pro-science, 10=promotes pseudoscience)
    ├── promotes_pseudoscience: bool
    ├── categories_found, indicators
    ├── respects_scientific_consensus: bool
    └── confidence, reasoning

NOTE: If pseudoscience is detected, the FactCheck score is pushed to a minimum of 5.0.

resolve_outlet_name(url, domain)
    │
    ├─► 1. URL Heuristic
    │       ├─► Extract domain base: "bbc.com" → "bbc"
    │       ├─► Short names (≤4 chars) → UPPERCASE: "bbc" → "BBC"
    │       └─► Longer names → Title Case: "foxnews" → "Foxnews"
    │
    ├─► 2. About Page Scrape (LLM fallback)
    │       ├─► Try: /about, /about-us, /about/, /about-us/,
    │       │        /corporate/about, /company/about
    │       ├─► Extract page text (up to 5000 chars)
    │       └─► LLM extracts official name from about page
    │           ├─► "apnews" → "The Associated Press"
    │           ├─► "foxnews" → "Fox News"
    │           └─► "nytimes" → "The New York Times"
    │
    └─► OUTPUT: Official outlet name (string)

research_all(outlet_name, domain)  — uses batch_search() when supported
    │
    ├─► research_history(outlet_name, domain)
    │       ├─► Tier 1: Direct about page scrape
    │       ├─► Tier 2: Search (via backend)
    │       ├─► Tier 3: Wikipedia fallback
    │       ├─► Tier 4: Domain-based search
    │       └─► LLM Extraction → HistoryLLMOutput
    │
    ├─► research_ownership(outlet_name, domain)
    │       ├─► Search: "{outlet_name}" ownership owner parent company
    │       └─► LLM Extraction → OwnershipLLMOutput (owner, funding_model, headquarters)
    │
    └─► research_external_analysis(outlet_name, domain)
            ├─► Targets: MBFC, Ad Fontes, NewsGuard, CJR, Nieman Lab
            └─► LLM Extraction → ExternalAnalysisLLMOutput

SEARCH BLACKLIST (filtered from all results):
    facebook.com, twitter.com, x.com, instagram.com,
    tiktok.com, pinterest.com, linkedin.com, reddit.com, youtube.com

# Run full benchmark
python run_benchmark.py

# Run specific modes
python run_benchmark.py --modes system hybrid llm

# Extract underperforming subset
python extract_subset.py --mode extract

# Compare all 6 modes
python extract_subset.py --mode compare

# Clone repository
git clone https://github.com/MirasBaisbay/media-profiling.git
cd media-profiling

# Create virtual environment
python -m venv venv
source venv/bin/activate  # Linux/Mac
# or: venv\Scripts\activate  # Windows

# Install dependencies
pip install -r requirements.txt

# Set API key (for LLM-based analyzers)
export OPENAI_API_KEY="your-api-key"

streamlit>=1.30.0          # Web interface
langchain-openai>=0.1.0    # LLM structured output
duckduckgo-search>=5.0.0   # Default search backend
openai>=1.30.0             # OpenAI search backend + LLM
pydantic>=2.0.0            # Type-safe schemas
beautifulsoup4>=4.12.0     # Web scraping
python-whois>=0.9.0        # Domain age lookup
evaluate>=0.4.0            # Benchmark metrics (METEOR, ROUGE)
numpy>=1.24.0              # Numerical computation
tqdm>=4.65.0               # Progress bars
tabulate>=0.9.0            # Benchmark result tables

# Ensure datasets/ folder contains SemEval 2020 Task 11 data
python train_pipeline.py

# Run locally
streamlit run app.py

# Opens at http://localhost:8501

# Option A: Clone the empty Space repo and copy files into it
git clone https://huggingface.co/spaces/<your-username>/media-profiler
cd media-profiler

# Copy all project files into this directory

# Option B: Add the Space as a remote to your existing repo
cd /path/to/your/media-profiling
git remote add space https://huggingface.co/spaces/<your-username>/media-profiler

# If using Option A (cloned Space repo):
git add .
git commit -m "Deploy Media Profiler to HuggingFace Spaces"
git push origin main

# If using Option B (added as remote):
git push space main

---
title: Media Profiler
emoji: 📰
colorFrom: blue
colorTo: indigo
sdk: streamlit
sdk_version: "1.30.0"
app_file: app.py
pinned: false
---

# Run the CLI profiler
python main_pipeline.py https://www.bbc.com

# Force re-analysis (ignore cache)
python main_pipeline.py https://www.bbc.com --refresh

from research import MediaProfiler

# Initialize profiler (default: DuckDuckGo search backend)
profiler = MediaProfiler()

# Or with hybrid search backend
from search_backends import HybridSearchBackend
profiler = MediaProfiler(search_backend=HybridSearchBackend())

# Sample articles (in practice, use scraper.py to collect these)
articles = [
    {
        "title": "Climate Change Policy Faces Opposition",
        "text": "The administration's new climate policy has drawn criticism..."
    },
    {
        "title": "Healthcare Reform Debate Intensifies",
        "text": "As healthcare costs continue to rise, lawmakers are divided..."
    },
]

# Profile an outlet
report = profiler.profile("https://www.bbc.com", articles)

# Generate text report
report_text = profiler.generate_report_text(report)
print(report_text)

======================================================================
MEDIA BIAS/FACT CHECK REPORT: BBC
======================================================================
URL: https://www.bbc.com
Analysis Date: 2025-01-15

QUICK SUMMARY
----------------------------------------
  Bias Rating:        Left-Center
  Factuality Rating:  High
  Credibility:        High Credibility
  Media Type:         TV
  Traffic:            HIGH
  Domain Age:         28.5 years
  Freedom Rating:     Good Situation

HISTORY
----------------------------------------
  Founded: 1922
  Founder(s): John Reith
  Key Events:
    - First TV broadcasts 1936
    - Charter renewal 2017

  The BBC was founded in 1922 by John Reith...

FUNDED BY / OWNERSHIP
----------------------------------------
  Owner: British Public
  Funding: License Fee (public funding)
  Headquarters: London, United Kingdom

BIAS ANALYSIS
----------------------------------------
  Overall Bias: Left-Center (score: -2.5)
  Economic: -2.0 | Social: -3.0 | News Reporting: -2.5 | Editorial: -2.5
  Uses Loaded Language: No

FACTUALITY ANALYSIS
----------------------------------------
  Factuality Rating: High (score: 1.8/10)

  Fact Check Search Results:
    Total Fact Checks Found: 3
    Failed Fact Checks: 0

  Sourcing Quality:
    Score: 2.0/10
    Has Primary Sources: Yes
    Has Wire Services: Yes

  Transparency: 1.0/10
  One-Sidedness: 1.5/10

PSEUDOSCIENCE CHECK
----------------------------------------
  Promotes Pseudoscience: No
  Respects Scientific Consensus: Yes

======================================================================
Articles Analyzed: 2
Generated by Media Profiling System
======================================================================

media-profiling/
│
├── config.py                    # Configuration constants and scoring scales
│   ├── CREDIBILITY_POINTS (factual, bias, traffic, freedom penalties)
│   ├── FACTUALITY_RANGES / BIAS_RANGES (MBFC label thresholds)
│   ├── FREEDOM_LABELS (5-tier freedom rating system)
│   ├── COUNTRY_NAME_ALIASES (HQ location → country matching)
│   ├── ECONOMIC_SCALE / SOCIAL_SCALE / NEWS_REPORTING_SCALE
│   └── Propaganda techniques (14 classes)
│
├── schemas.py                   # Pydantic v2 schemas for structured LLM outputs
│   ├── Article Classification (ArticleType, ArticleClassification)
│   ├── Media Type (MediaType, MediaTypeClassification)
│   ├── Traffic/Longevity (TrafficTier, TrafficData)
│   ├── Fact Check (FactCheckVerdict, FactCheckFinding, FactCheckAnalysisResult)
│   ├── Sourcing (SourceQuality, SourceAssessment, SourcingAnalysisResult)
│   ├── Editorial Bias (BiasDirection, PolicyPosition, EditorialBiasResult)
│   ├── Pseudoscience (PseudoscienceCategory, PseudoscienceIndicator, PseudoscienceAnalysisResult)
│   ├── Transparency (TransparencyLLMOutput, TransparencyResult)
│   ├── One-Sidedness (OneSidednessLLMOutput, OneSidednessResult)
│   ├── ComprehensiveReportData (complete output with all subscores)
│   └── MBFCTargetSchema / EvidenceSource (benchmark evaluation schema)
│
├── methodology.py               # MBFC 2025 scoring rubric + anti-contamination
│   ├── SCORING_INSTRUCTIONS (master methodology prompt)
│   ├── ANTI_CONTAMINATION_SYSTEM_PROMPT
│   ├── Mode-specific data source instructions
│   └── build_system_prompt() (combines all prompts per mode)
│
├── refactored_analyzers.py      # LLM-based analyzers with structured output
│   ├── OpinionAnalyzer (article type classification)
│   ├── TrafficLongevityAnalyzer (hybrid Tranco + WHOIS + LLM)
│   ├── MediaTypeAnalyzer (hybrid lookup + LLM)
│   ├── FactCheckSearcher (multi-site search + LLM parsing)
│   ├── SourcingAnalyzer (link extraction + LLM quality assessment)
│   ├── EditorialBiasAnalyzer (4 weighted bias subscores)
│   ├── PseudoscienceAnalyzer (LLM-based pseudoscience detection)
│   ├── TransparencyAnalyzer (ownership/funding/location transparency)
│   └── OneSidednessAnalyzer (propaganda/one-sidedness detection)
│
├── research.py                  # Web research and profiling orchestrator
│   ├── MediaResearcher (pluggable search backend, batch support)
│   │   ├── resolve_outlet_name() (URL heuristics + LLM about page fallback)
│   │   ├── _scrape_about_page() (direct /about page fetch, cached)
│   │   ├── research_history() (tiered: about page → search → Wikipedia)
│   │   ├── research_ownership() (search + domain fallback)
│   │   ├── research_external_analysis() (media watchdog search)
│   │   └── research_all() (batch search when backend supports it)
│   ├── MediaProfiler (9-analyzer orchestrator, freedom rating lookup)
│   └── Convenience functions (research_outlet, profile_outlet)
│
├── search_backends.py           # Pluggable search backend system
│   ├── DDGSearchBackend (DuckDuckGo, default)
│   ├── OpenAISearchBackend (OpenAI Responses API web_search)
│   └── HybridSearchBackend (routes site: queries to DDG, others to OpenAI)
│
├── evaluators.py                # 6-mode evaluation framework
│   ├── LLM-only, articles, search, MBC, system, hybrid runners
│   ├── CoTWrapper (Chain-of-Thought schema)
│   └── RUNNER_MAP / get_runner() (unified interface)
│
├── app.py                       # Streamlit web interface (HuggingFace Spaces ready)
│
├── main_pipeline.py             # CLI entry point: scrape → profile → generate → save
│
├── report_generator.py          # LLM-based MBFC prose report generation
│   └── ReportGenerator (GPT-4o synthesis with structured prompting)
│
├── storage.py                   # Persistence layer (30-day cache)
│   └── StorageManager (JSON + Markdown report caching)
│
├── requirements.txt             # Python dependencies for deployment
│
├── scraper.py                   # Web scraping for articles and metadata
│   ├── MediaScraper
│   ├── Article dataclass
│   └── SiteMetadata dataclass
│
├── metrics.py                   # Benchmark evaluation metrics
│   └── MetricsCalculator (FACTScore, METEOR, ROUGE-L, MAE, FC Detection)
│
├── run_benchmark.py             # Grid evaluation framework (models × modes × samples)
│
├── extract_subset.py            # Mode comparison and subset extraction
│
├── article_cache.py             # Article caching for benchmark fairness
│   └── ArticleCache (scrape-once, reuse-everywhere)
│
├── dataset_builder.py           # Stratified golden dataset builder from MBFC data
│
├── clean_mbfc.py                # MBFC dataset cleaning (removes fact-check URLs)
│
├── local_detector.py            # DeBERTa inference pipeline
│   └── LocalPropagandaDetector
│
├── train_pipeline.py            # DeBERTa fine-tuning pipeline
│
├── parser.py                    # MBFC website parser
│
├── known_media_types.csv        # Pre-classified media outlet types
│
├── tranco_top1m.csv             # Tranco top 1M domains (auto-downloaded)
│
├── 2025.csv                     # RSF Press Freedom Index dataset (181 countries)
│
├── datasets/                    # SemEval 2020 Task 11 data
│   ├── train/
│   ├── dev/
│   └── test/
│
└── propaganda_models/           # Trained model outputs
    ├── si_model/
    └── tc_model/