updated changes in pro-admin bias verdict

checker/internal/bias_analyzer.py

@@ -1,1423 +1,1406 @@
-"""
-Bias Analyzer — Heuristic-based political bias detection + subjectivity scoring
-=================================================================================
-Analyzes Filipino news articles for political bias using:
-  - Keyword patterns (left/right/sensational)
-  - Philippine political figure sentiment (2025-2026 context)
-  - Unsourced claim / "blind item" detection
-  - Opinion markers and subjectivity thresholds
-  - VADER sentiment scoring
-
-Returns: Left-leaning, Right-leaning, Conservative, or Centrist.
-Also flags unsourced claims (chismis / blind item style fake news).
-"""
-
-import re
-from textblob import TextBlob
-from vaderSentiment.vaderSentiment import SentimentIntensityAnalyzer
-
-# Shared VADER analyzer instance
-_vader = SentimentIntensityAnalyzer()
-
-
-# ── Philippine Political Figure Keywords (2025-2026 context) ────────────────
-#
-# The Philippine political landscape after 2024 has three main camps:
-#   CONSERVATIVE (Duterte faction) — Pro-Duterte, Anti-Marcos, Anti-Leni
-#   ADMIN/RIGHT  (Marcos faction)  — Pro-Marcos admin, critical of Duterte now
-#   LIBERAL/LEFT (Leni bloc)       — Pro-Leni, Anti-Duterte, Anti-Marcos
-#
-# A CENTRIST does not side with personalities but examines credentials and policy.
-
-# Phrases used by/about the DUTERTE FACTION (conservative/nationalist)
-CONSERVATIVE_PRO = [
-    "tatay digong",
-    "digong",
-    "mayor duterte",
-    "president duterte",
-    "rody duterte",
-    "rodrigo duterte",
-    "war on drugs works",
-    "drug war effective",
-    "bong go",
-    "sara duterte is right",
-    "vp sara",
-    "baste duterte",
-    "dds",
-    "duterte die-hard supporter",
-    "pro-duterte",
-    "18 marines",
-    "federalismo",
-    "charter change ni duterte",
-    "the dutertes",
-    "duterte legacy",
-    "pagmamahal sa bansa",
-    "du30",
-    "DU30",
-    "DDS",
-    "Tatay Digong",
-    "👊",
-    "👊🇵🇭",
-    "Dilawan",
-    "Yellowtard",
-    "Lugaw",
-    "Lutang",
-    "Penoy",
-    "Oligarchs",
-    "Drug war",
-    "Build Build Build",
-    "Safety first",
-    "Strongman",
-    "Disiplina",
-    "Bayaran",
-    "Presinto",
-    "Fake news media",
-    "Biased",
-    "NPAs",
-    "Pula-pula",
-    "Aktibista",
-    "Woke",
-    "Moralista",
-    "Elitista",
-    "Matapang",
-    "Davao",
-    "Tapang at Malasakit",
-    "May gawa, hindi salita",
-    "Solid",
-    "Diehard",
-    "Puro reklamo",
-    "Sumunod na lang",
-    "Bakit hindi ka tumakbo?",
-    "Ambisyosa",
-    "Dating VP",
-    "Len-Len",
-    "Noynoying",
-    "Komunista",
-    "Katotohanan",
-    "DDS",
-    "Digong",
-    "Tatay Digong",
-    "PRRD",
-    "tapang at malasakit",
-    "Dutertards",  # self-reclaimed by some supporters
-    "tokhang",
-    "war on drugs",
-    "change is coming",
-    "federalismo",
-    "BBL supporter",
-    "Build Build Build",
-    "Mindanaoan pride",
-    "Davao model",
-    "anti-dilawan",
-    "dilawan tayo",  # used sarcastically against critics
-    "presyo ng bigas",  # rice price talking point
-    "disente daw",  # mockery of "decent" critics
-    "RP first",
-    "sovereign PH",
-    "OFW ni Digong",
-    "ICC out",
-    "ICC biased",
-    "CHR useless",
-    "media biased",  # ABS-CBN franchise issue framing
-    "NTF-ELCAC",
-    "red-tagging justified",
-    "leftist agenda",
-    "destabilize",
-    "dilawang media",
-    "lugaw",  # derogatory term for Robredo
-    "walang nagawa si Leni",
-    "Marcos-Duterte alliance",
-    "UniTeam loyalist",
-    "Hugpong",
-    "PDDS",
-    "Katas ng Saudi",  # OFW remittance pride narrative
-    "drug lord protectors",
-    "yellows",
-    "Dutertista",
-]
-CONSERVATIVE_ANTI = [
-    # Anti-Marcos (used by Duterte faction after their fallout)
-    "marcos incompetent",
-    "marcos palpak",
-    "marcos nagkanulo",
-    "ibaba si marcos",
-    "marcos ay traydor",
-    "marcos nagpalit ng tuktok",
-    "marcos dynasty",
-    "marcos diktador",
-    "bangag",
-    "Magnanakaw",
-    "BabyM",
-    "LBM",
-    "Sinungaling",
-    "Unity-ng-magnanakaw",
-    "Party President",
-    "Bangag",
-    "Cocaine-user",
-    "Dictator junior",
-    "Tax evader",
-    "Mandaraya",
-    "Golden Age (ironic)",
-    "Tallano Gold (mocking)",
-    "Traveler-in-Chief",
-    "Unity-is-scam",
-    "Historical distortor",
-    "Puppet of Romualdez",
-    "Weak leader",
-    "Walang nagawa",
-    "Incompetent",
-    "Flood control scandal",
-    "Trillion Peso March",
-    "Ill-gotten wealth",
-    "Martial law baby",
-    "Vlog-president",
-    "Absentee leader",
-    "Impeachment target",
-    "Marcos-crony",
-    "BBM-serye",
-    "Nepo-baby",
-    "Unite-troll",
-    "Bagong Pilipinas (mocking)",
-    "Maharlika-scam",
-    "Rice at 20 pesos (mocking)",
-    "Unity-over-justice",
-    "Puro party",
-    "Walang kusa",
-    "Fake diploma",
-    "Legacy-washer",
-    "Marcos-pa-rin (ironic)"
-    # Anti-Leni
-    "dilawan",
-    "kakampink",
-    "kakampwet",
-    "hontivirus",
-    "trilliling",
-    "leni lugaw",
-    "yellow army",
-    "kakampink talo",
-    "leni talo",
-    "leni sinungaling",
-    "leni bayad",
-    "aquino oligarch",
-    "oposisyon mapanlinlang",
-    "sentenaryo ng laban nila",
-    "Lugaw",
-    "Lutang",
-    "Pinktard",
-    "Kakampwet",
-    "NPA-sympathizer",
-    "Elitista",
-    "Matapobre",
-    "Moral high grounder",
-    "Dilawan",
-    "Yellowtard",
-    "Fake VP",
-    "Puppet of the US",
-    "Oligarch-funded",
-    "Puro dada",
-    "Self-righteous",
-    "Woke",
-    "Cancel culture",
-    "Leni-lenin",
-    "Komunista",
-    "Pabida",
-    "Pa-victim",
-    "Len-len",
-    "Madumb",
-    "Puppet",
-    "Lahat na lang mali",
-    "Neoliberal",
-    "Yellow-bleeding",
-    "Manila-centric",
-    "Hypocrite",
-    "The 1%",
-    "Walang alam",
-    "Staging photos",
-    "PR-driven",
-    "Puro puso walang utak",
-    "Kulto ni Leni",
-    "Privileged",
-    "Out of touch",
-    "Political loser",
-    "Ambisyosa",
-    "Bakit hindi ka tumakbo?",
-    "lugaw",  # derogatory nickname for Robredo
-    "incompetente si Leni",
-    "walang nagawa si VP",
-    "plastic na pagmamahal",
-    "pang-epal lang",
-    "OVP budget sayang",
-    "pink propaganda",
-    "emotional voters",
-    "bobotante ng emosyon",
-    "dilawan ulit",
-    "yellowtard",
-    "Aquino alaga",
-    "Cory hangover",
-    "oligarko ng LP",
-    "LP trapo",
-    "Liberal Party corrupt",
-    "Robredo incompetent",
-    "lugaw republic",
-    "pabor sa NPA",
-    "komunistang suporta",
-    "abogadong walang alam",
-    "epal na relief ops",
-    "drama lang siya",
-    "intsik ang pondo",  # baseless China-funding smear
-    "kakampink bobo",
-    "pink cult",
-    "sira ang isip ng pink",
-    "Leni talo na talo",
-    "loser si Leni",
-    "hindi electable",
-    "media darling lang",
-    "biased na survey",
-    "SWS peke",
-    "Pulse Asia bayad",
-    "mainstream media alaga",
-    "woke politics",
-    "imported agenda",
-    "western puppet",
-    "NGO funded Leni",
-    "traydor sa Visayas",
-]
-
-# Phrases used by/about the MARCOS ADMINISTRATION faction
-ADMIN_PRO = [
-    "president marcos",
-    "bongbong marcos",
-    "bbm",  # NOTE: kept for compound matching but see neutral filter below
-    "pbbm",
-    "marcos administration",
-    "marcos government",
-    "uniteam",
-    "marcos legacy",
-    "bagong pilipinas",
-    "marcos economic plan",
-    "first lady liza",
-    "liza araneta-marcos",
-    "marcos right",
-    "marcos tama",
-    "marcos galing",
-    "marcos matino",
-    # NOTE: 'PBBM' and 'BBM' removed as standalone bias signals.
-    # These are neutral abbreviations used by ALL Philippine media to refer to
-    # the sitting president. Only biased contextual phrases are kept.
-    "UniTeam",
-    "❤️",
-    "✌️",
-    "Solid North",
-    "Bagong Pilipinas",
-    "Unity",
-    "Pagkakaisa",
-    "Marcosian",
-    "Loyalist",
-    "Tiger of the North",
-    "BBM-Sara",
-    "Economic recovery",
-    "Maharlika",
-    "Kadiwa",
-    "Food security",
-    "Legacy",
-    "Golden Age",
-    "Moving forward",
-    "Forgiveness",
-    "Negative campaigning",
-    "Paninira",
-    "Pinktard",
-    "Kakampwet",
-    "Uto-uto",
-    "Propaganda",
-    "Stability",
-    "Continuity",
-    "Diplomacy",
-    "Pres. Bongbong",
-    "First Family",
-    "Infrastructure",
-    "Modernization",
-    "Walang iwanan",
-    "Pilipinas Muna",
-    "Tigre",
-    "Pula",
-    "Bangon muli",
-    "Sama-sama",
-    "Bagong Umaga",
-    # Removed standalone 'PBBM' and 'BBM' — neutral abbreviations
-    "Bongbong Marcos",
-    "Bagong Pilipinas",
-    "UniTeam",
-    "Sara Duterte",
-    "Solid Marcos",
-    "Marcos pa rin",
-    "golden age ng Marcos",
-    "rice at P20",
-    "mahirap pag-usapan ang kasaysayan",  # revisionism framing
-    "historical revisionism daw",
-    "Imee Marcos",
-    "Ilocano pride",
-    "Solid North",
-    "MarcosLegacy",
-    "hindi communista",
-    "federalismo ni BBM",
-    "infra development",
-    "GNP growth",
-    "Marcos era golden years",
-    "Tallano gold",  # fringe pro-Marcos wealth myth
-    "Yamashita treasure",
-    "opposition nagrereklamo lang",
-    "KKB (Kapit-Kapit Bisig)",
-    "probinsya first",
-    "sariling atin",  # economic nationalism framing
-    "Maharlika Fund",
-    "MIF supporter",
-    "DICT projects",
-    "Marcos hindi magnanakaw",
-    "rehabilitasyon ni Marcos",
-    "marcos loyalista",
-    "bisita Marcos museum",
-    "Solid Cebu",
-    "Solid Mindanao",
-    "admin side",
-    "Malacanang suporta",
-    "Sara VP proud",
-    "BBM-Sara tandem",
-]
-ADMIN_ANTI = [
-    # Anti-Duterte (used by Marcos camp after fallout)
-    "duterte corrupt",
-    "duterte accountability",
-    "duterte icc",
-    "duterte drug lord",
-    "duterte criminal",
-    "arrest duterte",
-    "duterte drug syndicate",
-    "duterte narco",
-]
-
-# Phrases used by/about the LIBERAL/LEFT (Leni bloc)
-LIBERAL_PRO = [
-    "leni robredo",
-    "si leni",
-    "leni honest",
-    "leni tapat",
-    "angat buhay ni leni",
-    "kakampink",
-    "kay leni",
-    "kiko pangilinan",
-    "leila de lima",
-    "liberals tama",
-    "leni integrity",
-    "leni credentials",
-    "leni public service",
-    "pink movement",
-    "leni sincere",
-    "walang corrupt si leni",
-    "laban ng masa",
-    "makabayan bloc",
-    "progressive",
-    "Kakampink",
-    "Leni-Kiko",
-    "💗",
-    "🌸",
-    "Angat Buhay",
-    "Kulay Rosas",
-    "Husay at Tibay",
-    "Gobyernong Tapat",
-    "Radikal ang Magmahal",
-    "Resibo",
-    "Fact-check",
-    "Trolls",
-    "Magnanakaw",
-    "Dictator",
-    "Historical distortion",
-    "Unity is not enough",
-    "Walang Solid North",
-    "Bardagulan",
-    "Tumindig",
-    "Para sa Bansa",
-    "Volunteers",
-    "Pink Movement",
-    "Liwanag sa Dilim",
-    "Laban Leni",
-    "Let Leni Lead",
-    "OVP",
-    "Disinformation",
-    "Tallano Gold",
-    "BBM",
-    "LBM",
-    "Sinungaling",
-    "Tax evader",
-    "Mulat",
-    "Kababaihan",
-    "Youth vote",
-    "Para sa kinabukasan",
-    "Transparency",
-    "Accountability",
-    "Edukado",
-    "Ipanalo Natin 'To",
-    "Kakampink",
-    "Leni Lugaw",  # reclaimed term of endearment
-    "Kay Leni tayo",
-    "Pink movement",
-    "Pink volunteer",
-    "Angat Buhay",
-    "OVP proud",
-    "VP Leni",
-    "Team Leni",
-    "masa tayo",
-    "puso ng serbisyo",
-    "bobotante tayo",
-    "dilaw",  # reclaimed yellow color symbol
-    "totoong pagbabago",
-    "Cory legacy",
-    "EDSA spirit",
-    "Never Again",
-    "Martial Law victims",
-    "human rights first",
-    "CHR matters",
-    "fact check muna",
-    "propaganda detector",
-    "akademiko tayo",
-    "Robredo admin",
-    "Vote wisely",
-    "hindi pa huli",
-    "madaming puso",
-    "serbisyong totoo",
-    "Comelec transparency",
-    "kontra sa dayaan",
-    "boto natin 'to",
-    "PinkLaban",
-    "pag-asa ng bayan",
-    "bukas na isip",
-    "anti-red-tagging",
-    "INC split",  # electoral coalition discussion
-    "masa over masa",
-    "Leody-Leni",
-    "sama-sama tayo",
-    "Laban Leni",
-]
-LIBERAL_ANTI = [
-    # Anti-Duterte (mainstay of liberal discourse)
-    "extrajudicial killing",
-    "ejk",
-    "ej killing",
-    "duterte human rights",
-    "duterte impunity",
-    "duterte murders",
-    "drug war victims",
-    "tokhang victims",
-    "Dutertard",
-    "DDS-troll",
-    "Fentanyl",
-    "Davao Group",
-    "ICC-bound",
-    "Traydor",
-    "Kill kill kill",
-    "EJK-enabler",
-    "False prophet",
-    "Budol",
-    "Wumao",
-    "China-doll",
-    "POGO-protector",
-    "Kulto",
-    "Brainwashed",
-    "Violence-prone",
-    "Davao Dynasty",
-    "Mamatay-tao",
-    "Bastos",
-    "Digong-iyakin",
-    "Paid trolls",
-    "Fake news peddler",
-    "Tuta ng China",
-    "Digunggong",
-    "Bisakol (derogatory context)",
-    "Confidential funds queen",
-    "Impeachable",
-    "Threat to democracy",
-    "War on drugs failure",
-    "Galing Davao",
-    "Siga-siga",
-    "No respect for law",
-    "Human rights violator",
-    "Extra-judicial",
-    "Blind follower",
-    "DDS-shaming",
-    "Toxic",
-    "Divisionist",
-    "Feudal",
-    "Oust-Duterte",
-    "Dutertador",  # Duterte + diktador
-    "diktador",
-    "pasistang rehimen",
-    "state-sponsored killing",
-    "EJK",  # extra-judicial killing
-    "extra-judicial killing",
-    "dead on arrival",  # EJK slang
-    "drug war patay",
-    "salvage victim",
-    "patayin ang mahirap",
-    "tokhang para sa mayaman",
-    "tulak pababa",  # pushing down the poor
-    "gulat na gulat",  # ironic shock at impunity
-    "impunidad",
-    "ICC isampa",  # ICC filing demand
-    "idemanda si Duterte",
-    "war criminal",
-    "Davao Death Squad",
-    "DDS killer",
-    "Duterte resign",
-    "Duterte pababa",
-    "diktadurang Duterte",
-    "pinaslang",
-    "Duterte drug lord",
-    "narco list fake",
-    "red-tagging biktima",
-    "NTF-ELCAC terorista",
-    "baril batas",  # "gun is the law" critique
-    "walang due process",
-    "peke na pagbabago",
-    "tuta ng China",  # China appeasement critique
-    "nagbenta ng teritoryo",
-    "West Philippine Sea talunan",
-    "basura na pangako",
-    "sinungaling si Duterte",
-    "ABS-CBN shutdown",
-    "franchise denied",
-    "presyo tumaas",
-    "gutom sa ilalim ni Duterte",
-    "human rights patay",
-    # Anti-Marcos
-    "never forget martial law",
-    "hindi makakalimot",
-    "marcos plunder",
-    "marcos ill-gotten wealth",
-    "marcos human rights abuses",
-    "batas militar",
-    "martial law victims",
-    "diktadurang Marcos",
-    "Never Again Marcos",
-    "Martial Law atrocities",
-    "Marcos diktador",
-    "Ferdinand Marcos magnanakaw",
-    "ill-gotten wealth",
-    "nakawin ng Marcos",
-    "Marcos thief",
-    "PCGG asset",
-    "ibalik ang ninakaw",
-    "human rights abuses",
-    "desaparecido",
-    "torture victim Marcos",
-    "1081 batas militar",
-    "September 21 hindi kalimutan",
-    "Marcos burial laban",
-    "LNMB disgrasya",
-    "historical revisionist",
-    "fake news Marcos",
-    "propaganda ng Marcos",
-    "BBM bobo",
-    "BBM pasang-awa",
-    "grade fake BBM",
-    "Oxford dropout",
-    "hindi nagtapos",
-    "plunder case Marcos",
-    "tax evasion Marcos",
-    "hindi nagbayad ng buwis",
-    "Maharlika Fund scam",
-    "MIF panganib",
-    "Marcos crony",
-    "Enrile kasama",
-    "Sara Duterte traydor",
-    "UniTeam nagkahiwalay",
-    "Marcos-Duterte gulo",
-    "Solid North bulok",
-    "admin corruption",
-    "Malacanang sinungaling",
-    "BBM resign",
-    "Marcos pababa",
-]
-
-# Centrist / evidence-based signals
-EVIDENCE_BASED_MARKERS = [
-    "ayon sa datos",
-    "according to official records",
-    "based on data",
-    "per official report",
-    "according to the commission",
-    "sinabi ng opisyal",
-    "ayon sa pag-aaral",
-    "studies show",
-    "data shows",
-    "statistics show",
-    "according to the survey",
-    "sinabi ng eksperto",
-    "expert said",
-    "peer-reviewed",
-    "official statement",
-    "press release",
-    "on record",
-    "verified by",
-    "fact-checked",
-    "source:",
-    "sources:",
-    "according to",
-    "as reported by",
-    "per the department",
-    "ayon sa palasyo",
-    "ayon sa doj",
-    "ayon sa comelec",
-    "ayon sa ched",
-    "ayon sa doh",
-    "ayon sa deped",
-    "transcript shows",
-    "court records",
-    "senate hearing",
-    "house hearing",
-    "plunder case",
-    "graft case",
-    "ombudsman",
-]
-
-
-# ── Generic Left/Right Keywords (broader patterns) ───────────────────────────
-
-RIGHT_LEANING_KEYWORDS = [
-    "nagtagumpay",
-    "matagumpay",
-    "malaking tulong",
-    "pagbabago",
-    "disiplina",
-    "drug war",
-    "malasakit",
-    "build build build",
-    "matuwid",
-    "natatapos",
-    "pangako",
-    "nagpapatupad",
-    "achievements",
-    "anti-drug",
-    "anti-crime",
-    "peace and order",
-    "development",
-    "ekonomiya",
-    "destabilize",
-    "destabilization",
-    "oposisyon",
-    "anti-government",
-    "bias media",
-    "oligarch",
-    "propagandist",
-    "sovereignty",
-    "soberanya",
-    "traitor",
-    "worst traitor",
-    "national interest",
-    "patriotism",
-    "nationalist",
-    "faithful servant",
-    "servant of the people",
-    "wrongly accused",
-    "political persecution",
-    "political prisoner",
-    "legacy",
-    "icc persecution",
-    "he murdered no one",
-    "presumed innocent",
-]
-
-LEFT_LEANING_KEYWORDS = [
-    "human rights",
-    "extrajudicial",
-    "ej killing",
-    "ejk",
-    "impunity",
-    # 'corruption'/'korapsyon' removed — used by ALL political sides, causes false positives
-    # 'accountability'/'transparency' removed — same reason (universal talking points)
-    "press freedom",
-    "authoritarian",
-    "dictator",
-    "diktador",
-    "martial law",
-    "red-tagging",
-    "red tagging",
-    "activists",
-    "workers rights",
-    "labor rights",
-    "progressive",
-    "inequality",
-    "social justice",
-    "democratic space",
-    "civil liberties",
-    "dissent",
-    "political prisoners",
-    # Specific to liberal/left PH discourse
-    "disinformation",  # kept — used more as anti-fake-news framing by left
-    "state-sponsored killing",
-    "oust",
-    "resign",
-]
-
-
-# ── Unsourced Claim / "Blind Item" Signals ───────────────────────────────────
-#
-# Fake news on Philippine Facebook often mimics showbiz blind items or
-# "chismis" posts: dramatic claims made without named sources, often ending
-# with prompts to share before the post is deleted.
-
-UNSOURCED_CLAIM_SIGNALS = [
-    # Filipino hearsay markers — claim made but no named source
-    "daw",
-    "raw",
-    "diumano",
-    "umano",
-    "may nagsabi",
-    "may nagbalita",
-    "may nag-ulat",
-    "may nagpadala sa akin",
-    "may nagbigay ng impormasyon",
-    "may kilala akong",
-    "kilala ko ang nagbigay",
-    "mula sa isang pinagkakatiwalaang tao",
-    "mula sa reliable na source",
-    "reliable ang source",
-    "hindi ko maaaring ibigay ang pangalan",
-    "hindi maaaring ibunyag ang pangalan",
-    "hindi ko masasabi kung sino",
-    "ipinadala sa akin ng isang kaibigan",
-    # Verification-prompting phrases (signals the poster doubts authenticity)
-    "totoo kaya ito",
-    "tama ba ito",
-    "kung totoo ito",
-    "pag-isipan ninyo",
-    "nakaka-isip",
-    "nagtataka ako",
-    # Viral / share-baiting language
-    "share before deleted",
-    "i-share bago i-delete",
-    "share na bago tanggalin",
-    "bago pa tanggalin",
-    "delete na ito agad",
-    "wag palampasin",
-    "must share",
-    "share this now",
-    "share niyo",
-    "share before they take this down",
-    # Anonymous-source framing
-    "anonymous source",
-    "blind item",
-    "from a reliable source",
-    "sources say",
-    "sources told",
-    "a source revealed",
-    # English hearsay
-    "i heard that",
-    "i was told that",
-    "apparently",
-    "it is said that",
-    "rumor has it",
-    "parang totoo",
-    "unconfirmed reports",
-    "unverified",
-    "hindi pa kumpirmado",
-    "hindi pa confirmed",
-    "balita lang",
-    "chismis lang ito",
-]
-
-# Attribution markers — when PRESENT they REDUCE the unsourced claim score
-ATTRIBUTION_MARKERS = [
-    "ayon sa",
-    "sabi ni",
-    "according to",
-    "stated by",
-    "as reported by",
-    "cited by",
-    "based on",
-    "sinabi ni",
-    "pahayag ni",
-    "per ",
-    "per president",
-    "per department",
-    "sinabi ng",
-    "inihayag ni",
-    "ipinagpaliwanag ni",
-    "idineklara ni",
-    "isinagawa ng",
-    "pinayagan ni",
-    "nagpahayag si",
-    "nagbigay ng pahayag si",
-    "nagsalita si",
-    "said ",
-    "announced ",
-    "confirmed ",
-    "stated ",
-    "declared ",
-    "officials say",
-    "government said",
-    "department said",
-    "report says",
-    "study says",
-    "survey says",
-]
-
-
-# ── Sensationalism / Emotional Charge ────────────────────────────────────────
-
-SENSATIONAL_KEYWORDS = [
-    # Filipino
-    "grabe",
-    "sobrang",
-    "nakakatakot",
-    "nakakalungkot",
-    "kadiri",
-    "nakakagalit",
-    "buwag",
-    "gago",
-    "bobo",
-    "sinungaling",
-    "traydor",
-    "pahamak",
-    "palpak",
-    "basura",
-    "walang silbi",
-    "panalo",
-    "astig",
-    "idol",
-    "legend",
-    "must read",
-    "must watch",
-    "share",
-    "viral",
-    "breaking",
-    "just in",
-    "urgent",
-    "adik",
-    "NAPAKA",
-    # English — emotional/sensational
-    "shocking",
-    "outrageous",
-    "disgusting",
-    "horrifying",
-    "terrifying",
-    "devastating",
-    "despicable",
-    "abominable",
-    "deplorable",
-    "wickedness",
-    "depraved",
-    "stench",
-    "satan",
-    "evil",
-    "fire and fury",
-    "pipe dream",
-    "diatribe",
-    "harangue",
-    "perfidious",
-    "despairingly",
-    "intolerable",
-    "condemning",
-    "conspired",
-    "cronies",
-    "sham",
-    "scam",
-    "hoax",
-    "unbelievable",
-    "wake up",
-    "exposed",
-    "betrayed",
-    "rigged",
-    # Ad hominem
-    "vile",
-    "cunning",
-    "snake",
-    "poisoned",
-    "worst traitor",
-    "go down in history",
-    "day of reckoning",
-    "alarming",
-    "dangerous",
-    "coward",
-    "liar",
-    "hypocrite",
-    "shameless",
-    "corrupt",
-    "criminal",
-    "thief",
-    "incompetent",
-    "puppet",
-    "lapdog",
-    "traitor",
-    "backstabber",
-    "sellout",
-    "bankrupt",
-    "monster",
-    "villain",
-    "enemy of the people",
-    "enemy of the state",
-    "kakampi ng kaaway",
-    "tuta",
-    "duwag",
-    "magnanakaw",
-    "masamang tao",
-]
-
-# First-person / direct opinion markers
-OPINION_MARKERS = [
-    "in my opinion",
-    "in my assessment",
-    "in my view",
-    "i believe",
-    "i think",
-    "my legal opinion",
-    "my honest thoughts",
-    "i personally",
-    "from my perspective",
-    "i am inclined to believe",
-    "i am inclined",
-    "i pray",
-    "i am almost certain",
-    "i am certain",
-    "i am convinced",
-    "i suspect",
-    "i feel",
-    "i hope",
-    "we must",
-    "we should",
-    "it is my",
-    "this writer believes",
-    "this columnist",
-    "in this writer",
-    "i have no doubt",
-    "without a doubt",
-    "clearly",
-    "obviously",
-    "needless to say",
-    # Filipino first-person
-    "sa aking palagay",
-    "sa tingin ko",
-    "naniniwala ako",
-    "sa aking paniniwala",
-]
-
-SUBJECTIVITY_THRESHOLD = 0.50
-
-
-# ── Helper functions ──────────────────────────────────────────────────────────
-
-
-def _count_keyword_hits(text_lower, keyword_list):
-    """Count how many keywords from the list appear in the text."""
-    hits = []
-    for kw in keyword_list:
-        if kw.lower() in text_lower:
-            hits.append(kw)
-    return hits
-
-
-def _analyze_sensationalism(text_lower):
-    """Detect sensationalist language patterns."""
-    hits = _count_keyword_hits(text_lower, SENSATIONAL_KEYWORDS)
-    exclamation_count = text_lower.count("!")
-    question_chains = len(re.findall(r"\?{2,}", text_lower))
-    caps_words = len(re.findall(r"\b[A-Z]{3,}\b", text_lower))
-
-    score = len(hits) * 2 + min(exclamation_count, 5) + question_chains * 2 + caps_words
-    return score, hits
-
-
-def _detect_unsourced_claim(text_lower):
-    """Detect 'blind item' / chismis-style unsourced claims.
-
-    Returns (score, signals_found, attribution_count).
-    High score + low attribution = likely unsourced fake news style.
-    """
-    signals = _count_keyword_hits(text_lower, UNSOURCED_CLAIM_SIGNALS)
-    attributions = _count_keyword_hits(text_lower, ATTRIBUTION_MARKERS)
-
-    # Raw score: each unsourced signal +2, each attribution -1 (floor 0)
-    raw = len(signals) * 2
-    credit = len(attributions)
-    score = max(0, raw - credit)
-    return score, signals, len(attributions)
-
-
-def _detect_obfuscation(text):
-    """Detect intentional misspelling / character substitution used to evade
-    content moderation filters.
-
-    Fake news peddlers on Philippine social media commonly:
-      1. Replace letters with look-alike symbols: \u0394\u2192a, @\u2192a inside words
-      2. Use leet-speak digits: 3\u2192e, 0\u2192o, 1\u2192i inside words
-      3. Phonetically misspell: 'manga' for 'mga', 'panu' for 'paano'
-      4. Excessive ALL-CAPS beyond known government acronyms
-
-    Returns (score, signals_found).  Score >= 2 is a reliable flag.
-    """
-    signals = []
-
-    # 1. Explicit character substitutions — each hit is worth 2 points
-    #    (a single substitution is definitive; no need to wait for 2 signals)
-    explicit_score = 0
-
-    # Delta family: U+2206 INCREMENT ∆, U+0394 GREEK DELTA Δ,
-    #               U+25B3 WHITE TRIANGLE △, U+25B2 BLACK TRIANGLE ▲
-    DELTA_CHARS = "\u2206\u0394\u25b3\u25b2"
-    delta_hits = sum(text.count(c) for c in DELTA_CHARS)
-    if delta_hits:
-        signals.append(f"\u2206/\u0394 used as letter \u00d7 {delta_hits}")
-        explicit_score += 2
-
-    # Degree symbol \u00b0 used as 'o' (e.g. 'tunt\u00b0' = 'tunto')
-    degree_hits = text.count("\u00b0")
-    if degree_hits:
-        signals.append(f"\u00b0 used as letter \u00d7 {degree_hits}")
-        explicit_score += 2
-
-    # @ inside a word as 'a' substitution (not an email address)
-    at_in_word = len(re.findall(r"\w@\w", text))
-    if at_in_word:
-        signals.append(f"@ inside word \u00d7 {at_in_word}")
-        explicit_score += 2
-
-    # Other single-char Latin letter substitutions: ¡\u2192i, \u00f8\u2192o, \u00a5\u2192y, \u20ac\u2192e
-    for ch, label in [
-        ("\u00a1", "\u00a1\u2192i"),
-        ("\u00f8", "\u00f8\u2192o"),
-        ("\u00a5", "\u00a5\u2192y"),
-        ("\u20ac", "\u20ac\u2192e"),
-        ("\u00a2", "\u00a2\u2192c"),
-        ("\u00a3", "\u00a3\u2192b"),
-    ]:
-        if ch in text:
-            signals.append(f"'{label}' substitution")
-            explicit_score += 2
-
-    # Generic non-ASCII scan (weak signal — only add if no explicit sub found)
-    non_ascii = len(re.findall(r"[^\x00-\x7F]", text))
-    if non_ascii > 0 and explicit_score == 0 and 1 < non_ascii <= 20:
-        signals.append(f"non-ASCII chars in Latin text \u00d7 {non_ascii}")
-
-    # 2. Leet-speak: digits inside words (e.g. 'm4rcos', '3JK')
-    leet_words = [
-        w for w in re.findall(r"\b\w*[30145]\w*\b", text) if re.search(r"[a-zA-Z]", w)
-    ]
-    if len(leet_words) >= 2:
-        signals.append(f"leet digits in words: {leet_words[:4]}")
-
-    # 3. Known Filipino phonetic misspells / slang markers
-    PHONETIC = [
-        (r"\bmanga\b", "'manga' for 'mga'"),
-        (r"\bkau\b", "'kau' for 'kayo'"),
-        (r"\bnman\b", "'nman' for 'naman'"),
-        (r"\bpanu\b", "'panu' for 'paano'"),
-        (r"\bganun\b", "'ganun' for 'ganoon'"),
-        (r"\bbeh\b", "'beh' informal"),
-        (r"\bteh\b", "'teh' informal"),
-        (r"\bchar\b", "'char' joke marker"),
-        (r"\bOMG\b", "'OMG' exclamation"),
-        (r"\bWTF\b", "profane exclamation"),
-    ]
-    tl = text.lower()
-    for pat, label in PHONETIC:
-        if re.search(pat, tl):
-            signals.append(label)
-
-    # 4. Excessive ALL-CAPS beyond standard government acronyms
-    LEGIT_CAPS = {
-        "PNP",
-        "AFP",
-        "DOH",
-        "COMELEC",
-        "CHR",
-        "DOJ",
-        "OMB",
-        "DILG",
-        "DSWD",
-        "CHED",
-        "DEPED",
-        "DOE",
-        "DBM",
-        "BIR",
-        "PSA",
-        "BSP",
-        "SEC",
-        "LTO",
-        "LTFRB",
-        "MMDA",
-        "NBI",
-        "NASA",
-        "FBI",
-        "CIA",
-        "USA",
-        "PH",
-        "OVP",
-        "ICC",
-        "DNA",
-        "HIV",
-        "WHO",
-        "WFP",
-        "ASEAN",
-        "UN",
-        "DDS",
-        "BBM",
-        "PBBM",
-        "VP",
-        "SWS",
-        "PCSO",
-        "PAGCOR",
-        "EJK",
-        "PRRD",
-        "AFP",
-    }
-    suspicious_caps = [
-        w for w in re.findall(r"\b[A-Z]{3,}\b", text) if w not in LEGIT_CAPS
-    ]
-    if len(suspicious_caps) >= 3:
-        signals.append(f"excessive ALL-CAPS: {suspicious_caps[:5]}")
-    # Heavy ALL-CAPS (≥ 8 non-acronym caps words) alone is a strong social-media
-    # obfuscation signal — real journalism never shouts entire sentences in caps.
-    # Boost explicit_score so the flag triggers even without ∆/leet/@ substitutions.
-    if len(suspicious_caps) >= 8:
-        explicit_score += 2
-
-    score = explicit_score + len(signals) - (1 if explicit_score > 0 else 0)
-    # ^ explicit hits already added to score; subtract 1 to avoid counting their
-    #   signal string twice (each explicit hit already contributes to len(signals))
-    # Simplify: final score = explicit_score + non-explicit signal count
-    non_explicit_signals = [
-        s
-        for s in signals
-        if not any(
-            k in s
-            for k in ["\u2206", "\u0394", "\u00b0", "@", "substitution", "non-ASCII"]
-        )
-    ]
-    score = explicit_score + len(non_explicit_signals)
-    return score, signals
-
-
-def _compute_subjectivity(text):
-    """Compute subjectivity score using TextBlob (0=objective, 1=subjective)."""
-    try:
-        return TextBlob(text).sentiment.subjectivity
-    except Exception:
-        return 0.0
-
-
-def _compute_vader_sentiment(text):
-    """Compute VADER sentiment intensity."""
-    try:
-        return _vader.polarity_scores(text)
-    except Exception:
-        return {"compound": 0.0, "pos": 0.0, "neu": 1.0, "neg": 0.0}
-
-
-def _score_political_figures(text_lower):
-    """Score political figure sentiment for Philippine context.
-
-    Returns (conservative_score, admin_score, liberal_score, figure_hits).
-    """
-    cons_pro = _count_keyword_hits(text_lower, CONSERVATIVE_PRO)
-    cons_anti = _count_keyword_hits(text_lower, CONSERVATIVE_ANTI)
-    adm_pro = _count_keyword_hits(text_lower, ADMIN_PRO)
-    adm_anti = _count_keyword_hits(text_lower, ADMIN_ANTI)
-    lib_pro = _count_keyword_hits(text_lower, LIBERAL_PRO)
-    lib_anti = _count_keyword_hits(text_lower, LIBERAL_ANTI)
-    evidence = _count_keyword_hits(text_lower, EVIDENCE_BASED_MARKERS)
-
-    # Net scores: positive praise - negative criticism signals the faction leaning
-    conservative_score = len(cons_pro) * 2 + len(cons_anti) * 1
-    admin_score = len(adm_pro) * 2 + len(adm_anti) * 1
-    liberal_score = len(lib_pro) * 2 + len(lib_anti) * 1
-    evidence_score = len(evidence)
-
-    figure_hits = {
-        "conservative_pro": cons_pro,
-        "conservative_anti": cons_anti,
-        "admin_pro": adm_pro,
-        "admin_anti": adm_anti,
-        "liberal_pro": lib_pro,
-        "liberal_anti": lib_anti,
-        "evidence_markers": evidence,
-    }
-    return conservative_score, admin_score, liberal_score, evidence_score, figure_hits
-
-
-# ── Main function ─────────────────────────────────────────────────────────────
-
-
-def analyze_bias(text, pattern_deviation=None, lang="en"):
-
-    """Analyze a news article for political bias and subjectivity.
-
-    Args:
-        text (str): Article text.
-        pattern_deviation (dict|None): Result from compare_to_external_pattern().
-            When provided, deviation from reputable sources directly boosts
-            sensationalism score and leaning confidence — making pattern deviation
-            a primary signal rather than a post-hoc annotation.
-
-    Returns:
-        dict with keys:
-            leaning:                    "Conservative", "Admin/Right", "Left-leaning", or "Centrist"
-            is_biased:                  bool — True if clearly leaning or sensational
-            unsourced_claim:            bool — True if article appears to be a blind item / chismis
-            unsourced_score:            int  — raw unsourced claim score
-            subjectivity:               float (0-1)
-            subjectivity_flag:          bool
-            vader_compound:             float
-            vader_biased:               bool
-            confidence:                 float (0-1)
-            pattern_deviation_verdict:  str — "FOLLOWS PATTERN", "MINOR DEVIATION", "SIGNIFICANT DEVIATION", or ""
-            deviation_score:            float — 0.0–1.0 fraction of sentences not in reputable sources
-            details:                    dict with full breakdown
-    """
-    if not text or not isinstance(text, str) or len(text.strip()) < 20:
-        return {
-            "leaning": "Centrist",
-            "is_biased": False,
-            "unsourced_claim": False,
-            "unsourced_score": 0,
-            "obfuscated_text": False,
-            "obfuscation_score": 0,
-            "subjectivity": 0.0,
-            "subjectivity_flag": False,
-            "vader_compound": 0.0,
-            "vader_biased": False,
-            "confidence": 0.0,
-            "details": {"note": "Text too short for analysis"},
-        }
-
-    text_lower = text.lower()
-
-    # ── Generic keyword matching ──
-    right_hits = _count_keyword_hits(text_lower, RIGHT_LEANING_KEYWORDS)
-    left_hits = _count_keyword_hits(text_lower, LEFT_LEANING_KEYWORDS)
-    sensationalism_score, sensational_hits = _analyze_sensationalism(text_lower)
-    opinion_hits = _count_keyword_hits(text_lower, OPINION_MARKERS)
-    sensationalism_score += len(opinion_hits) * 3
-
-    right_score = len(right_hits)
-    left_score = len(left_hits)
-
-    # ── Philippine political figure scoring ──
-    conservative_score, admin_score, liberal_score, evidence_score, figure_hits = (
-        _score_political_figures(text_lower)
-    )
-
-    # Fold figure scores into left/right totals
-    right_score += conservative_score + admin_score
-    left_score += liberal_score
-
-    # Evidence-based writing can push toward centrist
-    centrist_boost = evidence_score
-
-    # ── Unsourced claim / blind item detection ──
-    unsourced_score, unsourced_signals, attribution_count = _detect_unsourced_claim(
-        text_lower
-    )
-    unsourced_flag = unsourced_score >= 4  # threshold: 2+ signals with few attributions
-    if unsourced_flag:
-        sensationalism_score += 2  # unsourced claims are a sensationalism signal
-
-    # ── Obfuscation / intentional misspell detection ──
-    obfuscation_score, obfuscation_signals = _detect_obfuscation(text)
-    obfuscated_flag = obfuscation_score >= 2
-    if obfuscated_flag:
-        sensationalism_score += min(3, obfuscation_score)  # boost but cap contribution
-
-    # ── Subjectivity ──
-    subjectivity = _compute_subjectivity(text)
-    subjectivity_flag = subjectivity > SUBJECTIVITY_THRESHOLD
-
-    # ── VADER sentiment ──
-    vader_scores = _compute_vader_sentiment(text)
-    vader_compound = vader_scores["compound"]
-    VADER_BIAS_THRESHOLD = 0.5
-    vader_biased = abs(vader_compound) >= VADER_BIAS_THRESHOLD
-    if vader_biased:
-        sensationalism_score += 3
-
-    # ── Pattern deviation boost (primary signal when available) ────────────────
-    # Deviation from reputable external sources is treated as a strong bias
-    # indicator: articles that add claims not found in any credible source are
-    # more likely to be biased or sensationalized, regardless of keyword matches.
-    pd_verdict = (pattern_deviation or {}).get("verdict", "")
-    pd_score = (pattern_deviation or {}).get("deviation_score", 0.0)
-    pd_confidence_boost = 0.0
-
-    if pd_verdict == "SIGNIFICANT DEVIATION":
-        sensationalism_score += 3   # strong signal: adds claims no credible source reports
-        pd_confidence_boost = 0.08  # more certain about the detected leaning
-    elif pd_verdict == "MINOR DEVIATION":
-        sensationalism_score += 1   # mild signal
-
-    # ── Determine leaning ──
-    total_political = right_score + left_score
-    bias_threshold = 3  # raised slightly since we now have more signal sources
-
-    # Centrist boost: strong evidence markers counteract weak political scores
-    effective_political = max(0, total_political - centrist_boost)
-
-    if effective_political < bias_threshold and sensationalism_score < 4:
-        leaning = "Centrist"
-        confidence = 0.6 + min(0.3, (bias_threshold - effective_political) * 0.1)
-    elif right_score > left_score:
-        # Distinguish Conservative (Duterte-camp) vs Admin/Right (Marcos-camp)
-        if conservative_score >= admin_score:
-            leaning = "Conservative"
-        else:
-            leaning = "Admin/Right"
-        diff = right_score - left_score
-        confidence = min(0.95, 0.5 + diff * 0.08 + sensationalism_score * 0.02 + pd_confidence_boost)
-    elif left_score > right_score:
-        leaning = "Left-leaning"
-        diff = left_score - right_score
-        confidence = min(0.95, 0.5 + diff * 0.08 + sensationalism_score * 0.02 + pd_confidence_boost)
-    else:
-        leaning = "Centrist"
-        confidence = 0.45 if sensationalism_score >= 4 else 0.7
-
-    is_biased = (
-        (leaning != "Centrist")
-        or subjectivity_flag
-        or sensationalism_score >= 4
-        or vader_biased
-        or unsourced_flag
-        or obfuscated_flag
-    )
-
-    return {
-        "leaning": leaning,
-        "is_biased": is_biased,
-        "unsourced_claim": unsourced_flag,
-        "unsourced_score": unsourced_score,
-        "obfuscated_text": obfuscated_flag,
-        "obfuscation_score": obfuscation_score,
-        "subjectivity": round(subjectivity, 4),
-        "subjectivity_flag": subjectivity_flag,
-        "vader_compound": round(vader_compound, 4),
-        "vader_biased": vader_biased,
-        "confidence": round(min(confidence, 1.0), 4),
-        "pattern_deviation_verdict": pd_verdict,
-        "deviation_score": round(pd_score, 4),
-        "details": {
-            "left_keywords": left_hits,
-            "right_keywords": right_hits,
-            "sensational_keywords": sensational_hits,
-            "opinion_markers": opinion_hits,
-            "unsourced_signals": unsourced_signals,
-            "attribution_count": attribution_count,
-            "obfuscation_signals": obfuscation_signals,
-            "figure_hits": figure_hits,
-            "left_score": left_score,
-            "right_score": right_score,
-            "conservative_score": conservative_score,
-            "admin_score": admin_score,
-            "liberal_score": liberal_score,
-            "evidence_score": evidence_score,
-            "sensationalism_score": sensationalism_score,
-            "pattern_deviation_verdict": pd_verdict,
-            "deviation_score": round(pd_score, 4),
-        },
-    }
+"""
+Bias Analyzer — Heuristic-based political bias detection + subjectivity scoring
+=================================================================================
+Analyzes Filipino news articles for political bias using:
+  - Keyword patterns (left/right/sensational)
+  - Philippine political figure sentiment (2025-2026 context)
+  - Unsourced claim / "blind item" detection
+  - Opinion markers and subjectivity thresholds
+  - VADER sentiment scoring
+
+Returns: Left-leaning, Right-leaning, Conservative, or Centrist.
+Also flags unsourced claims (chismis / blind item style fake news).
+"""
+
+import re
+from textblob import TextBlob
+from vaderSentiment.vaderSentiment import SentimentIntensityAnalyzer
+
+# Shared VADER analyzer instance
+_vader = SentimentIntensityAnalyzer()
+
+
+# ── Philippine Political Figure Keywords (2025-2026 context) ────────────────
+#
+# The Philippine political landscape after 2024 has three main camps:
+#   CONSERVATIVE (Duterte faction) — Pro-Duterte, Anti-Marcos, Anti-Leni
+#   ADMIN/RIGHT  (Marcos faction)  — Pro-Marcos admin, critical of Duterte now
+#   LIBERAL/LEFT (Leni bloc)       — Pro-Leni, Anti-Duterte, Anti-Marcos
+#
+# A CENTRIST does not side with personalities but examines credentials and policy.
+
+# Phrases used by/about the DUTERTE FACTION (conservative/nationalist)
+CONSERVATIVE_PRO = [
+    "tatay digong",
+    "digong",
+    "mayor duterte",
+    "president duterte",
+    "rody duterte",
+    "rodrigo duterte",
+    "war on drugs works",
+    "drug war effective",
+    "bong go",
+    "sara duterte is right",
+    "vp sara",
+    "baste duterte",
+    "dds",
+    "duterte die-hard supporter",
+    "pro-duterte",
+    "18 marines",
+    "federalismo",
+    "charter change ni duterte",
+    "the dutertes",
+    "duterte legacy",
+    "pagmamahal sa bansa",
+    "du30",
+    "DU30",
+    "DDS",
+    "Tatay Digong",
+    "👊",
+    "👊🇵🇭",
+    "Dilawan",
+    "Yellowtard",
+    "Lugaw",
+    "Lutang",
+    "Penoy",
+    "Oligarchs",
+    "Drug war",
+    "Build Build Build",
+    "Safety first",
+    "Strongman",
+    "Disiplina",
+    "Bayaran",
+    "Presinto",
+    "Fake news media",
+    "Biased",
+    "NPAs",
+    "Pula-pula",
+    "Aktibista",
+    "Woke",
+    "Moralista",
+    "Elitista",
+    "Matapang",
+    "Davao",
+    "Tapang at Malasakit",
+    "May gawa, hindi salita",
+    "Solid",
+    "Diehard",
+    "Puro reklamo",
+    "Sumunod na lang",
+    "Bakit hindi ka tumakbo?",
+    "Ambisyosa",
+    "Dating VP",
+    "Len-Len",
+    "Noynoying",
+    "Komunista",
+    "Katotohanan",
+    "DDS",
+    "Digong",
+    "Tatay Digong",
+    "PRRD",
+    "tapang at malasakit",
+    "Dutertards",  # self-reclaimed by some supporters
+    "tokhang",
+    "war on drugs",
+    "change is coming",
+    "federalismo",
+    "BBL supporter",
+    "Build Build Build",
+    "Mindanaoan pride",
+    "Davao model",
+    "anti-dilawan",
+    "dilawan tayo",  # used sarcastically against critics
+    "presyo ng bigas",  # rice price talking point
+    "disente daw",  # mockery of "decent" critics
+    "RP first",
+    "sovereign PH",
+    "OFW ni Digong",
+    "ICC out",
+    "ICC biased",
+    "CHR useless",
+    "media biased",  # ABS-CBN franchise issue framing
+    "NTF-ELCAC",
+    "red-tagging justified",
+    "leftist agenda",
+    "destabilize",
+    "dilawang media",
+    "lugaw",  # derogatory term for Robredo
+    "walang nagawa si Leni",
+    "Marcos-Duterte alliance",
+    "UniTeam loyalist",
+    "Hugpong",
+    "PDDS",
+    "Katas ng Saudi",  # OFW remittance pride narrative
+    "drug lord protectors",
+    "yellows",
+    "Dutertista",
+]
+CONSERVATIVE_ANTI = [
+    # Anti-Marcos (used by Duterte faction after their fallout)
+    "marcos incompetent",
+    "marcos palpak",
+    "marcos nagkanulo",
+    "ibaba si marcos",
+    "marcos ay traydor",
+    "marcos nagpalit ng tuktok",
+    "marcos dynasty",
+    "marcos diktador",
+    "bangag",
+    "Magnanakaw",
+    "BabyM",
+    "LBM",
+    "Sinungaling",
+    "Unity-ng-magnanakaw",
+    "Party President",
+    "Bangag",
+    "Cocaine-user",
+    "Dictator junior",
+    "Tax evader",
+    "Mandaraya",
+    "Golden Age (ironic)",
+    "Tallano Gold (mocking)",
+    "Traveler-in-Chief",
+    "Unity-is-scam",
+    "Historical distortor",
+    "Puppet of Romualdez",
+    "Weak leader",
+    "Walang nagawa",
+    "Incompetent",
+    "Flood control scandal",
+    "Trillion Peso March",
+    "Ill-gotten wealth",
+    "Martial law baby",
+    "Vlog-president",
+    "Absentee leader",
+    "Impeachment target",
+    "Marcos-crony",
+    "BBM-serye",
+    "Nepo-baby",
+    "Unite-troll",
+    "Bagong Pilipinas (mocking)",
+    "Maharlika-scam",
+    "Rice at 20 pesos (mocking)",
+    "Unity-over-justice",
+    "Puro party",
+    "Walang kusa",
+    "Fake diploma",
+    "Legacy-washer",
+    "Marcos-pa-rin (ironic)"
+    # Anti-Leni
+    "dilawan",
+    "kakampink",
+    "kakampwet",
+    "hontivirus",
+    "trilliling",
+    "leni lugaw",
+    "yellow army",
+    "kakampink talo",
+    "leni talo",
+    "leni sinungaling",
+    "leni bayad",
+    "aquino oligarch",
+    "oposisyon mapanlinlang",
+    "sentenaryo ng laban nila",
+    "Lugaw",
+    "Lutang",
+    "Pinktard",
+    "Kakampwet",
+    "NPA-sympathizer",
+    "Elitista",
+    "Matapobre",
+    "Moral high grounder",
+    "Dilawan",
+    "Yellowtard",
+    "Fake VP",
+    "Puppet of the US",
+    "Oligarch-funded",
+    "Puro dada",
+    "Self-righteous",
+    "Woke",
+    "Cancel culture",
+    "Leni-lenin",
+    "Komunista",
+    "Pabida",
+    "Pa-victim",
+    "Len-len",
+    "Madumb",
+    "Puppet",
+    "Lahat na lang mali",
+    "Neoliberal",
+    "Yellow-bleeding",
+    "Manila-centric",
+    "Hypocrite",
+    "The 1%",
+    "Walang alam",
+    "Staging photos",
+    "PR-driven",
+    "Puro puso walang utak",
+    "Kulto ni Leni",
+    "Privileged",
+    "Out of touch",
+    "Political loser",
+    "Ambisyosa",
+    "Bakit hindi ka tumakbo?",
+    "lugaw",  # derogatory nickname for Robredo
+    "incompetente si Leni",
+    "walang nagawa si VP",
+    "plastic na pagmamahal",
+    "pang-epal lang",
+    "OVP budget sayang",
+    "pink propaganda",
+    "emotional voters",
+    "bobotante ng emosyon",
+    "dilawan ulit",
+    "yellowtard",
+    "Aquino alaga",
+    "Cory hangover",
+    "oligarko ng LP",
+    "LP trapo",
+    "Liberal Party corrupt",
+    "Robredo incompetent",
+    "lugaw republic",
+    "pabor sa NPA",
+    "komunistang suporta",
+    "abogadong walang alam",
+    "epal na relief ops",
+    "drama lang siya",
+    "intsik ang pondo",  # baseless China-funding smear
+    "kakampink bobo",
+    "pink cult",
+    "sira ang isip ng pink",
+    "Leni talo na talo",
+    "loser si Leni",
+    "hindi electable",
+    "media darling lang",
+    "biased na survey",
+    "SWS peke",
+    "Pulse Asia bayad",
+    "mainstream media alaga",
+    "woke politics",
+    "imported agenda",
+    "western puppet",
+    "NGO funded Leni",
+    "traydor sa Visayas",
+]
+
+# Phrases used by/about the MARCOS ADMINISTRATION faction
+ADMIN_PRO = [
+   
+    "uniteam",
+    "marcos legacy",
+    "bagong pilipinas",
+    "marcos economic plan",
+    "first lady liza",
+    "marcos right",
+    "marcos tama",
+    "marcos galing",
+    "marcos matino",
+    # NOTE: 'PBBM' and 'BBM' removed as standalone bias signals.
+    # These are neutral abbreviations used by ALL Philippine media to refer to
+    # the sitting president. Only biased contextual phrases are kept.
+    "UniTeam",
+    "❤️",
+    "✌️",
+    "Solid North",
+    "Bagong Pilipinas",
+    "Unity",
+    "Pagkakaisa",
+    "Marcosian",
+    "Loyalist",
+    "Tiger of the North",
+    "BBM-Sara",
+    "Economic recovery",
+    "Maharlika",
+    "Kadiwa",
+    "Legacy",
+    "Golden Age",
+    "Moving forward",
+    "Forgiveness",
+    "Negative campaigning",
+    "Paninira",
+    "Pinktard",
+    "Kakampwet",
+    "Uto-uto",
+    "Propaganda",
+    "Stability",
+    "Continuity",
+    "Diplomacy",
+    "Pres. Bongbong",
+    "First Family",
+    "Infrastructure",
+    "Modernization",
+    "Walang iwanan",
+    "Pilipinas Muna",
+    "Tigre",
+    "Pula",
+    "Bangon muli",
+    "Sama-sama",
+    "Bagong Pilipinas",
+    "UniTeam",
+    "Sara Duterte",
+    "Solid Marcos",
+    "Marcos pa rin",
+    "golden age ng Marcos",
+    "rice at P20",
+    "mahirap pag-usapan ang kasaysayan",  # revisionism framing
+    "historical revisionism daw",
+    "Ilocano pride",
+    "Solid North",
+    "MarcosLegacy",
+    "hindi communista",
+    "federalismo ni BBM",
+    "infra development",
+    "GNP growth",
+    "Marcos era golden years",
+    "Tallano gold",  # fringe pro-Marcos wealth myth
+    "Yamashita treasure",
+    "opposition nagrereklamo lang",
+    "probinsya first",
+    "sariling atin",  # economic nationalism framing
+    "Maharlika Fund",
+    "Marcos hindi magnanakaw",
+    "rehabilitasyon ni Marcos",
+    "marcos loyalista",
+    "bisita Marcos museum",
+    "admin side",
+    "Malacanang suporta",
+    "BBM-Sara tandem",
+]
+ADMIN_ANTI = [
+    # Anti-Duterte (used by Marcos camp after fallout)
+    "duterte corrupt",
+    "duterte accountability",
+    "duterte icc",
+    "duterte drug lord",
+    "duterte criminal",
+    "arrest duterte",
+    "duterte drug syndicate",
+    "duterte narco",
+]
+
+# Phrases used by/about the LIBERAL/LEFT (Leni bloc)
+LIBERAL_PRO = [
+    "leni robredo",
+    "si leni",
+    "leni honest",
+    "leni tapat",
+    "angat buhay ni leni",
+    "kakampink",
+    "kay leni",
+    "kiko pangilinan",
+    "leila de lima",
+    "liberals tama",
+    "leni integrity",
+    "leni credentials",
+    "leni public service",
+    "pink movement",
+    "leni sincere",
+    "walang corrupt si leni",
+    "laban ng masa",
+    "makabayan bloc",
+    "progressive",
+    "Kakampink",
+    "Leni-Kiko",
+    "💗",
+    "🌸",
+    "Angat Buhay",
+    "Kulay Rosas",
+    "Husay at Tibay",
+    "Gobyernong Tapat",
+    "Radikal ang Magmahal",
+    "Resibo",
+    "Fact-check",
+    "Trolls",
+    "Magnanakaw",
+    "Dictator",
+    "Historical distortion",
+    "Unity is not enough",
+    "Walang Solid North",
+    "Bardagulan",
+    "Tumindig",
+    "Para sa Bansa",
+    "Volunteers",
+    "Pink Movement",
+    "Liwanag sa Dilim",
+    "Laban Leni",
+    "Let Leni Lead",
+    "OVP",
+    "Disinformation",
+    "Tallano Gold",
+    "BBM",
+    "LBM",
+    "Sinungaling",
+    "Tax evader",
+    "Mulat",
+    "Kababaihan",
+    "Youth vote",
+    "Para sa kinabukasan",
+    "Transparency",
+    "Accountability",
+    "Edukado",
+    "Ipanalo Natin 'To",
+    "Kakampink",
+    "Leni Lugaw",  # reclaimed term of endearment
+    "Kay Leni tayo",
+    "Pink movement",
+    "Pink volunteer",
+    "Angat Buhay",
+    "OVP proud",
+    "VP Leni",
+    "Team Leni",
+    "masa tayo",
+    "puso ng serbisyo",
+    "bobotante tayo",
+    "dilaw",  # reclaimed yellow color symbol
+    "totoong pagbabago",
+    "Cory legacy",
+    "EDSA spirit",
+    "Never Again",
+    "Martial Law victims",
+    "human rights first",
+    "CHR matters",
+    "fact check muna",
+    "propaganda detector",
+    "akademiko tayo",
+    "Robredo admin",
+    "Vote wisely",
+    "hindi pa huli",
+    "madaming puso",
+    "serbisyong totoo",
+    "Comelec transparency",
+    "kontra sa dayaan",
+    "boto natin 'to",
+    "PinkLaban",
+    "pag-asa ng bayan",
+    "bukas na isip",
+    "anti-red-tagging",
+    "INC split",  # electoral coalition discussion
+    "masa over masa",
+    "Leody-Leni",
+    "sama-sama tayo",
+    "Laban Leni",
+]
+LIBERAL_ANTI = [
+    # Anti-Duterte (mainstay of liberal discourse)
+    "extrajudicial killing",
+    "ejk",
+    "ej killing",
+    "duterte human rights",
+    "duterte impunity",
+    "duterte murders",
+    "drug war victims",
+    "tokhang victims",
+    "Dutertard",
+    "DDS-troll",
+    "Fentanyl",
+    "Davao Group",
+    "ICC-bound",
+    "Traydor",
+    "Kill kill kill",
+    "EJK-enabler",
+    "False prophet",
+    "Budol",
+    "Wumao",
+    "China-doll",
+    "POGO-protector",
+    "Kulto",
+    "Brainwashed",
+    "Violence-prone",
+    "Davao Dynasty",
+    "Mamatay-tao",
+    "Bastos",
+    "Digong-iyakin",
+    "Paid trolls",
+    "Fake news peddler",
+    "Tuta ng China",
+    "Digunggong",
+    "Bisakol (derogatory context)",
+    "Confidential funds queen",
+    "Impeachable",
+    "Threat to democracy",
+    "War on drugs failure",
+    "Galing Davao",
+    "Siga-siga",
+    "No respect for law",
+    "Human rights violator",
+    "Extra-judicial",
+    "Blind follower",
+    "DDS-shaming",
+    "Toxic",
+    "Divisionist",
+    "Feudal",
+    "Oust-Duterte",
+    "Dutertador",  # Duterte + diktador
+    "diktador",
+    "pasistang rehimen",
+    "state-sponsored killing",
+    "EJK",  # extra-judicial killing
+    "extra-judicial killing",
+    "dead on arrival",  # EJK slang
+    "drug war patay",
+    "salvage victim",
+    "patayin ang mahirap",
+    "tokhang para sa mayaman",
+    "tulak pababa",  # pushing down the poor
+    "gulat na gulat",  # ironic shock at impunity
+    "impunidad",
+    "ICC isampa",  # ICC filing demand
+    "idemanda si Duterte",
+    "war criminal",
+    "Davao Death Squad",
+    "DDS killer",
+    "Duterte resign",
+    "Duterte pababa",
+    "diktadurang Duterte",
+    "pinaslang",
+    "Duterte drug lord",
+    "narco list fake",
+    "red-tagging biktima",
+    "NTF-ELCAC terorista",
+    "baril batas",  # "gun is the law" critique
+    "walang due process",
+    "peke na pagbabago",
+    "tuta ng China",  # China appeasement critique
+    "nagbenta ng teritoryo",
+    "West Philippine Sea talunan",
+    "basura na pangako",
+    "sinungaling si Duterte",
+    "ABS-CBN shutdown",
+    "franchise denied",
+    "presyo tumaas",
+    "gutom sa ilalim ni Duterte",
+    "human rights patay",
+    # Anti-Marcos
+    "never forget martial law",
+    "hindi makakalimot",
+    "marcos plunder",
+    "marcos ill-gotten wealth",
+    "marcos human rights abuses",
+    "batas militar",
+    "martial law victims",
+    "diktadurang Marcos",
+    "Never Again Marcos",
+    "Martial Law atrocities",
+    "Marcos diktador",
+    "Ferdinand Marcos magnanakaw",
+    "ill-gotten wealth",
+    "nakawin ng Marcos",
+    "Marcos thief",
+    "PCGG asset",
+    "ibalik ang ninakaw",
+    "human rights abuses",
+    "desaparecido",
+    "torture victim Marcos",
+    "1081 batas militar",
+    "September 21 hindi kalimutan",
+    "Marcos burial laban",
+    "LNMB disgrasya",
+    "historical revisionist",
+    "fake news Marcos",
+    "propaganda ng Marcos",
+    "BBM bobo",
+    "BBM pasang-awa",
+    "grade fake BBM",
+    "Oxford dropout",
+    "hindi nagtapos",
+    "plunder case Marcos",
+    "tax evasion Marcos",
+    "hindi nagbayad ng buwis",
+    "Maharlika Fund scam",
+    "MIF panganib",
+    "Marcos crony",
+    "Enrile kasama",
+    "Sara Duterte traydor",
+    "UniTeam nagkahiwalay",
+    "Marcos-Duterte gulo",
+    "Solid North bulok",
+    "admin corruption",
+    "Malacanang sinungaling",
+    "BBM resign",
+    "Marcos pababa",
+]
+
+# Centrist / evidence-based signals
+EVIDENCE_BASED_MARKERS = [
+    "ayon sa datos",
+    "according to official records",
+    "based on data",
+    "per official report",
+    "according to the commission",
+    "sinabi ng opisyal",
+    "ayon sa pag-aaral",
+    "studies show",
+    "data shows",
+    "statistics show",
+    "according to the survey",
+    "sinabi ng eksperto",
+    "expert said",
+    "peer-reviewed",
+    "official statement",
+    "press release",
+    "on record",
+    "verified by",
+    "fact-checked",
+    "source:",
+    "sources:",
+    "according to",
+    "as reported by",
+    "per the department",
+    "ayon sa palasyo",
+    "ayon sa doj",
+    "ayon sa comelec",
+    "ayon sa ched",
+    "ayon sa doh",
+    "ayon sa deped",
+    "transcript shows",
+    "court records",
+    "senate hearing",
+    "house hearing",
+    "plunder case",
+    "graft case",
+    "ombudsman",
+]
+
+
+# ── Generic Left/Right Keywords (broader patterns) ───────────────────────────
+
+RIGHT_LEANING_KEYWORDS = [
+    "nagtagumpay",
+    "matagumpay",
+    "malaking tulong",
+    "pagbabago",
+    "disiplina",
+    "drug war",
+    "malasakit",
+    "build build build",
+    "matuwid",
+    "natatapos",
+    "pangako",
+    "nagpapatupad",
+    "achievements",
+    "anti-drug",
+    "anti-crime",
+    "peace and order",
+    "development",
+    "ekonomiya",
+    "destabilize",
+    "destabilization",
+    "oposisyon",
+    "anti-government",
+    "bias media",
+    "oligarch",
+    "propagandist",
+    "sovereignty",
+    "soberanya",
+    "traitor",
+    "worst traitor",
+    "national interest",
+    "patriotism",
+    "nationalist",
+    "faithful servant",
+    "servant of the people",
+    "wrongly accused",
+    "political persecution",
+    "political prisoner",
+    "legacy",
+    "icc persecution",
+    "he murdered no one",
+    "presumed innocent",
+]
+
+LEFT_LEANING_KEYWORDS = [
+    "human rights",
+    "extrajudicial",
+    "ej killing",
+    "ejk",
+    "impunity",
+    # 'corruption'/'korapsyon' removed — used by ALL political sides, causes false positives
+    # 'accountability'/'transparency' removed — same reason (universal talking points)
+    "press freedom",
+    "authoritarian",
+    "dictator",
+    "diktador",
+    "martial law",
+    "red-tagging",
+    "red tagging",
+    "activists",
+    "workers rights",
+    "labor rights",
+    "progressive",
+    "inequality",
+    "social justice",
+    "democratic space",
+    "civil liberties",
+    "dissent",
+    "political prisoners",
+    # Specific to liberal/left PH discourse
+    "disinformation",  # kept — used more as anti-fake-news framing by left
+    "state-sponsored killing",
+    "oust",
+    "resign",
+]
+
+
+# ── Unsourced Claim / "Blind Item" Signals ───────────────────────────────────
+#
+# Fake news on Philippine Facebook often mimics showbiz blind items or
+# "chismis" posts: dramatic claims made without named sources, often ending
+# with prompts to share before the post is deleted.
+
+UNSOURCED_CLAIM_SIGNALS = [
+    # Filipino hearsay markers — claim made but no named source
+    "daw",
+    "raw",
+    "diumano",
+    "umano",
+    "may nagsabi",
+    "may nagbalita",
+    "may nag-ulat",
+    "may nagpadala sa akin",
+    "may nagbigay ng impormasyon",
+    "may kilala akong",
+    "kilala ko ang nagbigay",
+    "mula sa isang pinagkakatiwalaang tao",
+    "mula sa reliable na source",
+    "reliable ang source",
+    "hindi ko maaaring ibigay ang pangalan",
+    "hindi maaaring ibunyag ang pangalan",
+    "hindi ko masasabi kung sino",
+    "ipinadala sa akin ng isang kaibigan",
+    # Verification-prompting phrases (signals the poster doubts authenticity)
+    "totoo kaya ito",
+    "tama ba ito",
+    "kung totoo ito",
+    "pag-isipan ninyo",
+    "nakaka-isip",
+    "nagtataka ako",
+    # Viral / share-baiting language
+    "share before deleted",
+    "i-share bago i-delete",
+    "share na bago tanggalin",
+    "bago pa tanggalin",
+    "delete na ito agad",
+    "wag palampasin",
+    "must share",
+    "share this now",
+    "share niyo",
+    "share before they take this down",
+    # Anonymous-source framing
+    "anonymous source",
+    "blind item",
+    "from a reliable source",
+    "sources say",
+    "sources told",
+    "a source revealed",
+    # English hearsay
+    "i heard that",
+    "i was told that",
+    "apparently",
+    "it is said that",
+    "rumor has it",
+    "parang totoo",
+    "unconfirmed reports",
+    "unverified",
+    "hindi pa kumpirmado",
+    "hindi pa confirmed",
+    "balita lang",
+    "chismis lang ito",
+]
+
+# Attribution markers — when PRESENT they REDUCE the unsourced claim score
+ATTRIBUTION_MARKERS = [
+    "ayon sa",
+    "sabi ni",
+    "according to",
+    "stated by",
+    "as reported by",
+    "cited by",
+    "based on",
+    "sinabi ni",
+    "pahayag ni",
+    "per ",
+    "per president",
+    "per department",
+    "sinabi ng",
+    "inihayag ni",
+    "ipinagpaliwanag ni",
+    "idineklara ni",
+    "isinagawa ng",
+    "pinayagan ni",
+    "nagpahayag si",
+    "nagbigay ng pahayag si",
+    "nagsalita si",
+    "said ",
+    "announced ",
+    "confirmed ",
+    "stated ",
+    "declared ",
+    "officials say",
+    "government said",
+    "department said",
+    "report says",
+    "study says",
+    "survey says",
+]
+
+
+# ── Sensationalism / Emotional Charge ────────────────────────────────────────
+
+SENSATIONAL_KEYWORDS = [
+    # Filipino
+    "grabe",
+    "sobrang",
+    "nakakatakot",
+    "nakakalungkot",
+    "kadiri",
+    "nakakagalit",
+    "buwag",
+    "gago",
+    "bobo",
+    "sinungaling",
+    "traydor",
+    "pahamak",
+    "palpak",
+    "basura",
+    "walang silbi",
+    "panalo",
+    "astig",
+    "idol",
+    "legend",
+    "must read",
+    "must watch",
+    "share",
+    "viral",
+    "breaking",
+    "just in",
+    "urgent",
+    "adik",
+    "NAPAKA",
+    # English — emotional/sensational
+    "shocking",
+    "outrageous",
+    "disgusting",
+    "horrifying",
+    "terrifying",
+    "devastating",
+    "despicable",
+    "abominable",
+    "deplorable",
+    "wickedness",
+    "depraved",
+    "stench",
+    "satan",
+    "evil",
+    "fire and fury",
+    "pipe dream",
+    "diatribe",
+    "harangue",
+    "perfidious",
+    "despairingly",
+    "intolerable",
+    "condemning",
+    "conspired",
+    "cronies",
+    "sham",
+    "scam",
+    "hoax",
+    "unbelievable",
+    "wake up",
+    "exposed",
+    "betrayed",
+    "rigged",
+    # Ad hominem
+    "vile",
+    "cunning",
+    "snake",
+    "poisoned",
+    "worst traitor",
+    "go down in history",
+    "day of reckoning",
+    "alarming",
+    "dangerous",
+    "coward",
+    "liar",
+    "hypocrite",
+    "shameless",
+    "corrupt",
+    "criminal",
+    "thief",
+    "incompetent",
+    "puppet",
+    "lapdog",
+    "traitor",
+    "backstabber",
+    "sellout",
+    "bankrupt",
+    "monster",
+    "villain",
+    "enemy of the people",
+    "enemy of the state",
+    "kakampi ng kaaway",
+    "tuta",
+    "duwag",
+    "magnanakaw",
+    "masamang tao",
+]
+
+# First-person / direct opinion markers
+OPINION_MARKERS = [
+    "in my opinion",
+    "in my assessment",
+    "in my view",
+    "i believe",
+    "i think",
+    "my legal opinion",
+    "my honest thoughts",
+    "i personally",
+    "from my perspective",
+    "i am inclined to believe",
+    "i am inclined",
+    "i pray",
+    "i am almost certain",
+    "i am certain",
+    "i am convinced",
+    "i suspect",
+    "i feel",
+    "i hope",
+    "we must",
+    "we should",
+    "it is my",
+    "this writer believes",
+    "this columnist",
+    "in this writer",
+    "i have no doubt",
+    "without a doubt",
+    "clearly",
+    "obviously",
+    "needless to say",
+    # Filipino first-person
+    "sa aking palagay",
+    "sa tingin ko",
+    "naniniwala ako",
+    "sa aking paniniwala",
+]
+
+SUBJECTIVITY_THRESHOLD = 0.50
+
+
+# ── Helper functions ──────────────────────────────────────────────────────────
+
+
+def _count_keyword_hits(text_lower, keyword_list):
+    """Count how many keywords from the list appear in the text."""
+    hits = []
+    for kw in keyword_list:
+        if kw.lower() in text_lower:
+            hits.append(kw)
+    return hits
+
+
+def _analyze_sensationalism(text_lower):
+    """Detect sensationalist language patterns."""
+    hits = _count_keyword_hits(text_lower, SENSATIONAL_KEYWORDS)
+    exclamation_count = text_lower.count("!")
+    question_chains = len(re.findall(r"\?{2,}", text_lower))
+    caps_words = len(re.findall(r"\b[A-Z]{3,}\b", text_lower))
+
+    score = len(hits) * 2 + min(exclamation_count, 5) + question_chains * 2 + caps_words
+    return score, hits
+
+
+def _detect_unsourced_claim(text_lower):
+    """Detect 'blind item' / chismis-style unsourced claims.
+
+    Returns (score, signals_found, attribution_count).
+    High score + low attribution = likely unsourced fake news style.
+    """
+    signals = _count_keyword_hits(text_lower, UNSOURCED_CLAIM_SIGNALS)
+    attributions = _count_keyword_hits(text_lower, ATTRIBUTION_MARKERS)
+
+    # Raw score: each unsourced signal +2, each attribution -1 (floor 0)
+    raw = len(signals) * 2
+    credit = len(attributions)
+    score = max(0, raw - credit)
+    return score, signals, len(attributions)
+
+
+def _detect_obfuscation(text):
+    """Detect intentional misspelling / character substitution used to evade
+    content moderation filters.
+
+    Fake news peddlers on Philippine social media commonly:
+      1. Replace letters with look-alike symbols: \u0394\u2192a, @\u2192a inside words
+      2. Use leet-speak digits: 3\u2192e, 0\u2192o, 1\u2192i inside words
+      3. Phonetically misspell: 'manga' for 'mga', 'panu' for 'paano'
+      4. Excessive ALL-CAPS beyond known government acronyms
+
+    Returns (score, signals_found).  Score >= 2 is a reliable flag.
+    """
+    signals = []
+
+    # 1. Explicit character substitutions — each hit is worth 2 points
+    #    (a single substitution is definitive; no need to wait for 2 signals)
+    explicit_score = 0
+
+    # Delta family: U+2206 INCREMENT ∆, U+0394 GREEK DELTA Δ,
+    #               U+25B3 WHITE TRIANGLE △, U+25B2 BLACK TRIANGLE ▲
+    DELTA_CHARS = "\u2206\u0394\u25b3\u25b2"
+    delta_hits = sum(text.count(c) for c in DELTA_CHARS)
+    if delta_hits:
+        signals.append(f"\u2206/\u0394 used as letter \u00d7 {delta_hits}")
+        explicit_score += 2
+
+    # Degree symbol \u00b0 used as 'o' (e.g. 'tunt\u00b0' = 'tunto')
+    degree_hits = text.count("\u00b0")
+    if degree_hits:
+        signals.append(f"\u00b0 used as letter \u00d7 {degree_hits}")
+        explicit_score += 2
+
+    # @ inside a word as 'a' substitution (not an email address)
+    at_in_word = len(re.findall(r"\w@\w", text))
+    if at_in_word:
+        signals.append(f"@ inside word \u00d7 {at_in_word}")
+        explicit_score += 2
+
+    # Other single-char Latin letter substitutions: ¡\u2192i, \u00f8\u2192o, \u00a5\u2192y, \u20ac\u2192e
+    for ch, label in [
+        ("\u00a1", "\u00a1\u2192i"),
+        ("\u00f8", "\u00f8\u2192o"),
+        ("\u00a5", "\u00a5\u2192y"),
+        ("\u20ac", "\u20ac\u2192e"),
+        ("\u00a2", "\u00a2\u2192c"),
+        ("\u00a3", "\u00a3\u2192b"),
+    ]:
+        if ch in text:
+            signals.append(f"'{label}' substitution")
+            explicit_score += 2
+
+    # Generic non-ASCII scan (weak signal — only add if no explicit sub found)
+    non_ascii = len(re.findall(r"[^\x00-\x7F]", text))
+    if non_ascii > 0 and explicit_score == 0 and 1 < non_ascii <= 20:
+        signals.append(f"non-ASCII chars in Latin text \u00d7 {non_ascii}")
+
+    # 2. Leet-speak: digits inside words (e.g. 'm4rcos', '3JK')
+    leet_words = [
+        w for w in re.findall(r"\b\w*[30145]\w*\b", text) if re.search(r"[a-zA-Z]", w)
+    ]
+    if len(leet_words) >= 2:
+        signals.append(f"leet digits in words: {leet_words[:4]}")
+
+    # 3. Known Filipino phonetic misspells / slang markers
+    PHONETIC = [
+        (r"\bmanga\b", "'manga' for 'mga'"),
+        (r"\bkau\b", "'kau' for 'kayo'"),
+        (r"\bnman\b", "'nman' for 'naman'"),
+        (r"\bpanu\b", "'panu' for 'paano'"),
+        (r"\bganun\b", "'ganun' for 'ganoon'"),
+        (r"\bbeh\b", "'beh' informal"),
+        (r"\bteh\b", "'teh' informal"),
+        (r"\bchar\b", "'char' joke marker"),
+        (r"\bOMG\b", "'OMG' exclamation"),
+        (r"\bWTF\b", "profane exclamation"),
+    ]
+    tl = text.lower()
+    for pat, label in PHONETIC:
+        if re.search(pat, tl):
+            signals.append(label)
+
+    # 4. Excessive ALL-CAPS beyond standard government acronyms
+    LEGIT_CAPS = {
+        "PNP",
+        "AFP",
+        "DOH",
+        "COMELEC",
+        "CHR",
+        "DOJ",
+        "OMB",
+        "DILG",
+        "DSWD",
+        "CHED",
+        "DEPED",
+        "DOE",
+        "DBM",
+        "BIR",
+        "PSA",
+        "BSP",
+        "SEC",
+        "LTO",
+        "LTFRB",
+        "MMDA",
+        "NBI",
+        "NASA",
+        "FBI",
+        "CIA",
+        "USA",
+        "PH",
+        "OVP",
+        "ICC",
+        "DNA",
+        "HIV",
+        "WHO",
+        "WFP",
+        "ASEAN",
+        "UN",
+        "DDS",
+        "BBM",
+        "PBBM",
+        "VP",
+        "SWS",
+        "PCSO",
+        "PAGCOR",
+        "EJK",
+        "PRRD",
+        "AFP",
+    }
+    suspicious_caps = [
+        w for w in re.findall(r"\b[A-Z]{3,}\b", text) if w not in LEGIT_CAPS
+    ]
+    if len(suspicious_caps) >= 3:
+        signals.append(f"excessive ALL-CAPS: {suspicious_caps[:5]}")
+    # Heavy ALL-CAPS (≥ 8 non-acronym caps words) alone is a strong social-media
+    # obfuscation signal — real journalism never shouts entire sentences in caps.
+    # Boost explicit_score so the flag triggers even without ∆/leet/@ substitutions.
+    if len(suspicious_caps) >= 8:
+        explicit_score += 2
+
+    score = explicit_score + len(signals) - (1 if explicit_score > 0 else 0)
+    # ^ explicit hits already added to score; subtract 1 to avoid counting their
+    #   signal string twice (each explicit hit already contributes to len(signals))
+    # Simplify: final score = explicit_score + non-explicit signal count
+    non_explicit_signals = [
+        s
+        for s in signals
+        if not any(
+            k in s
+            for k in ["\u2206", "\u0394", "\u00b0", "@", "substitution", "non-ASCII"]
+        )
+    ]
+    score = explicit_score + len(non_explicit_signals)
+    return score, signals
+
+
+def _compute_subjectivity(text):
+    """Compute subjectivity score using TextBlob (0=objective, 1=subjective)."""
+    try:
+        return TextBlob(text).sentiment.subjectivity
+    except Exception:
+        return 0.0
+
+
+def _compute_vader_sentiment(text):
+    """Compute VADER sentiment intensity."""
+    try:
+        return _vader.polarity_scores(text)
+    except Exception:
+        return {"compound": 0.0, "pos": 0.0, "neu": 1.0, "neg": 0.0}
+
+
+def _score_political_figures(text_lower):
+    """Score political figure sentiment for Philippine context.
+
+    Returns (conservative_score, admin_score, liberal_score, figure_hits).
+    """
+    cons_pro = _count_keyword_hits(text_lower, CONSERVATIVE_PRO)
+    cons_anti = _count_keyword_hits(text_lower, CONSERVATIVE_ANTI)
+    adm_pro = _count_keyword_hits(text_lower, ADMIN_PRO)
+    adm_anti = _count_keyword_hits(text_lower, ADMIN_ANTI)
+    lib_pro = _count_keyword_hits(text_lower, LIBERAL_PRO)
+    lib_anti = _count_keyword_hits(text_lower, LIBERAL_ANTI)
+    evidence = _count_keyword_hits(text_lower, EVIDENCE_BASED_MARKERS)
+
+    # Net scores: positive praise - negative criticism signals the faction leaning
+    conservative_score = len(cons_pro) * 2 + len(cons_anti) * 1
+    admin_score = len(adm_pro) * 2 + len(adm_anti) * 1
+    liberal_score = len(lib_pro) * 2 + len(lib_anti) * 1
+    evidence_score = len(evidence)
+
+    figure_hits = {
+        "conservative_pro": cons_pro,
+        "conservative_anti": cons_anti,
+        "admin_pro": adm_pro,
+        "admin_anti": adm_anti,
+        "liberal_pro": lib_pro,
+        "liberal_anti": lib_anti,
+        "evidence_markers": evidence,
+    }
+    return conservative_score, admin_score, liberal_score, evidence_score, figure_hits
+
+
+# ── Main function ─────────────────────────────────────────────────────────────
+
+
+def analyze_bias(text, pattern_deviation=None, lang="en"):
+
+    """Analyze a news article for political bias and subjectivity.
+
+    Args:
+        text (str): Article text.
+        pattern_deviation (dict|None): Result from compare_to_external_pattern().
+            When provided, deviation from reputable sources directly boosts
+            sensationalism score and leaning confidence — making pattern deviation
+            a primary signal rather than a post-hoc annotation.
+
+    Returns:
+        dict with keys:
+            leaning:                    "Conservative", "Admin/Right", "Left-leaning", or "Centrist"
+            is_biased:                  bool — True if clearly leaning or sensational
+            unsourced_claim:            bool — True if article appears to be a blind item / chismis
+            unsourced_score:            int  — raw unsourced claim score
+            subjectivity:               float (0-1)
+            subjectivity_flag:          bool
+            vader_compound:             float
+            vader_biased:               bool
+            confidence:                 float (0-1)
+            pattern_deviation_verdict:  str — "FOLLOWS PATTERN", "MINOR DEVIATION", "SIGNIFICANT DEVIATION", or ""
+            deviation_score:            float — 0.0–1.0 fraction of sentences not in reputable sources
+            details:                    dict with full breakdown
+    """
+    if not text or not isinstance(text, str) or len(text.strip()) < 20:
+        return {
+            "leaning": "Centrist",
+            "is_biased": False,
+            "unsourced_claim": False,
+            "unsourced_score": 0,
+            "obfuscated_text": False,
+            "obfuscation_score": 0,
+            "subjectivity": 0.0,
+            "subjectivity_flag": False,
+            "vader_compound": 0.0,
+            "vader_biased": False,
+            "confidence": 0.0,
+            "details": {"note": "Text too short for analysis"},
+        }
+
+    text_lower = text.lower()
+
+    # ── Generic keyword matching ──
+    right_hits = _count_keyword_hits(text_lower, RIGHT_LEANING_KEYWORDS)
+    left_hits = _count_keyword_hits(text_lower, LEFT_LEANING_KEYWORDS)
+    sensationalism_score, sensational_hits = _analyze_sensationalism(text_lower)
+    opinion_hits = _count_keyword_hits(text_lower, OPINION_MARKERS)
+    sensationalism_score += len(opinion_hits) * 3
+
+    right_score = len(right_hits)
+    left_score = len(left_hits)
+
+    # ── Philippine political figure scoring ──
+    conservative_score, admin_score, liberal_score, evidence_score, figure_hits = (
+        _score_political_figures(text_lower)
+    )
+
+    # Fold figure scores into left/right totals
+    right_score += conservative_score + admin_score
+    left_score += liberal_score
+
+    # Evidence-based writing can push toward centrist
+    centrist_boost = evidence_score
+
+    # ── Unsourced claim / blind item detection ──
+    unsourced_score, unsourced_signals, attribution_count = _detect_unsourced_claim(
+        text_lower
+    )
+    unsourced_flag = unsourced_score >= 4  # threshold: 2+ signals with few attributions
+    if unsourced_flag:
+        sensationalism_score += 2  # unsourced claims are a sensationalism signal
+
+    # ── Obfuscation / intentional misspell detection ──
+    obfuscation_score, obfuscation_signals = _detect_obfuscation(text)
+    obfuscated_flag = obfuscation_score >= 2
+    if obfuscated_flag:
+        sensationalism_score += min(3, obfuscation_score)  # boost but cap contribution
+
+    # ── Subjectivity ──
+    subjectivity = _compute_subjectivity(text)
+    subjectivity_flag = subjectivity > SUBJECTIVITY_THRESHOLD
+
+    # ── VADER sentiment ──
+    vader_scores = _compute_vader_sentiment(text)
+    vader_compound = vader_scores["compound"]
+    VADER_BIAS_THRESHOLD = 0.5
+    vader_biased = abs(vader_compound) >= VADER_BIAS_THRESHOLD
+    if vader_biased:
+        sensationalism_score += 3
+
+    # ── Pattern deviation boost (primary signal when available) ────────────────
+    # Deviation from reputable external sources is treated as a strong bias
+    # indicator: articles that add claims not found in any credible source are
+    # more likely to be biased or sensationalized, regardless of keyword matches.
+    pd_verdict = (pattern_deviation or {}).get("verdict", "")
+    pd_score = (pattern_deviation or {}).get("deviation_score", 0.0)
+    pd_confidence_boost = 0.0
+
+    if pd_verdict == "SIGNIFICANT DEVIATION":
+        sensationalism_score += 3   # strong signal: adds claims no credible source reports
+        pd_confidence_boost = 0.08  # more certain about the detected leaning
+    elif pd_verdict == "MINOR DEVIATION":
+        sensationalism_score += 1   # mild signal
+
+    # ── Determine leaning ──
+    total_political = right_score + left_score
+    bias_threshold = 3  # raised slightly since we now have more signal sources
+
+    # Centrist boost: strong evidence markers counteract weak political scores
+    effective_political = max(0, total_political - centrist_boost)
+
+    if effective_political < bias_threshold and sensationalism_score < 4:
+        leaning = "Centrist"
+        confidence = 0.6 + min(0.3, (bias_threshold - effective_political) * 0.1)
+    elif right_score > left_score:
+        # Distinguish Conservative (Duterte-camp) vs Admin/Right (Marcos-camp)
+        if conservative_score >= admin_score:
+            leaning = "Conservative"
+        else:
+            leaning = "Admin/Right"
+        diff = right_score - left_score
+        confidence = min(0.95, 0.5 + diff * 0.08 + sensationalism_score * 0.02 + pd_confidence_boost)
+    elif left_score > right_score:
+        leaning = "Left-leaning"
+        diff = left_score - right_score
+        confidence = min(0.95, 0.5 + diff * 0.08 + sensationalism_score * 0.02 + pd_confidence_boost)
+    else:
+        leaning = "Centrist"
+        confidence = 0.45 if sensationalism_score >= 4 else 0.7
+
+    is_biased = (
+        (leaning != "Centrist")
+        or subjectivity_flag
+        or sensationalism_score >= 4
+        or vader_biased
+        or unsourced_flag
+        or obfuscated_flag
+    )
+
+    return {
+        "leaning": leaning,
+        "is_biased": is_biased,
+        "unsourced_claim": unsourced_flag,
+        "unsourced_score": unsourced_score,
+        "obfuscated_text": obfuscated_flag,
+        "obfuscation_score": obfuscation_score,
+        "subjectivity": round(subjectivity, 4),
+        "subjectivity_flag": subjectivity_flag,
+        "vader_compound": round(vader_compound, 4),
+        "vader_biased": vader_biased,
+        "confidence": round(min(confidence, 1.0), 4),
+        "pattern_deviation_verdict": pd_verdict,
+        "deviation_score": round(pd_score, 4),
+        "details": {
+            "left_keywords": left_hits,
+            "right_keywords": right_hits,
+            "sensational_keywords": sensational_hits,
+            "opinion_markers": opinion_hits,
+            "unsourced_signals": unsourced_signals,
+            "attribution_count": attribution_count,
+            "obfuscation_signals": obfuscation_signals,
+            "figure_hits": figure_hits,
+            "left_score": left_score,
+            "right_score": right_score,
+            "conservative_score": conservative_score,
+            "admin_score": admin_score,
+            "liberal_score": liberal_score,
+            "evidence_score": evidence_score,
+            "sensationalism_score": sensationalism_score,
+            "pattern_deviation_verdict": pd_verdict,
+            "deviation_score": round(pd_score, 4),
+        },
+    }