examples.py

"""Demo examples — real, labeled rows (single source of truth for app.py and the verifier).

Each entry is [premise, claim, expected]. English cases first; then multilingual cases where the English
model is fooled but the multilingual model is right. These are verified to match the models' verdicts by
scripts/verify_examples path (see the repo), so the shown "expected" is what the model actually returns.
"""

_PREM_INCOME = (
    "CONSOLIDATED STATEMENTS OF INCOME - USD ($) shares in Thousands, $ in Thousands | 12 Months Ended\n"
    "Nov. 01, 2025 | Nov. 02, 2024 | Oct. 28, 2023\n"
    "Revenue | $ 11,019,707 | $ 9,427,157 | $ 12,305,539\n"
    "Cost of sales | 4,246,229 | 4,045,814 | 4,428,321\n"
    "Gross margin | 6,773,478 | 5,381,343 | 7,877,218\n"
    "Operating income | 2,932,496 | 2,032,798 | 3,823,112\n"
    "Net income | $ 2,267,342 | $ 1,635,273 | $ 3,314,579"
)
_PREM_EARNINGS = (
    "Consolidated Statements of Earnings (Parenthetical) - USD ($) $ in Millions | 12 Months Ended\n"
    "Sep. 30, 2025 | Sep. 30, 2024 | Sep. 30, 2023\n"
    "Interest income | $ 150 | $ 148 | $ 227\n"
    "Discontinued operation taxes | $ (4) | $ 85 | $ 2,969"
)

EXAMPLES = [
    [_PREM_INCOME, "For 2025-11-01, the revenue is $ 11,019,707.",
     "✅ grounded (matches the 2025 revenue)"],
    [_PREM_INCOME, "For 2025-11-01, the revenue is $ 9,427,157.",
     "❌ not grounded (that's the 2024 column — wrong period)"],
    [_PREM_EARNINGS, "For 2024, the period end is Sep. 30, 2024.",
     "✅ grounded"],
    [_PREM_EARNINGS, "For 2024, the period end is Sep. 30, 2025.",
     "❌ not grounded (Sep 30, 2025 is the wrong year's column)"],
    [_PREM_EARNINGS, "For 2024, the period end is Sep. 38, 2024.",
     "❌ not grounded (Sep 38 is not a real date)"],
    ["Drake đã phát hành album của mình Thank Me Hi vào năm 2010, nó được phát hành tại số một tại "
     "U.S.board 200 và chẳng bao lâu sau đó bởi Hiệp hội công nghiệp Hoa Kỳ (RIAA).",
     "Anh đã phát hành một album vào năm 2010.",
     "✅ grounded · Vietnamese (premise: album released in 2010)"],
    ["8 января 2017 года CW продлил шоу на четвёртый сезон.",
     "У Флэша много сезонов.",
     "✅ grounded · Russian (renewed for a 4th season → many seasons)"],
    ["IBMの発明には、自動トレーラーマシン(ATM)、PC、フロッピディスク、ハードディスクドライブ、磁気ストリップカード、"
     "関係データベース、SQLプログラミング言語、UPCバーコード、およびダイナミックランダムアクセスメモリ(DRAM)が含まれます。",
     "1998年にIBMがATMを発明した。",
     "❌ not grounded · Japanese (premise gives no year — 1998 is invented)"],
    ["Jennifer Garner. Garner 2005 yılında Ben Affleck ile evlendi, üç çocuğu var; 2015 yılında "
     "ayrıldılar ve Nisan 2017'de boşanma davası açtılar.",
     "Jennifer Garner sıfır ilişki yaşadı.",
     "❌ not grounded · Turkish (premise describes a marriage and children)"],
    ["杰夫·布里奇斯(Jeff Bridges)因在2009年电影《疯狂之心》中扮演Otis `Bad' Blake的角色而获最佳演员学院奖。",
     "Jeff Bridges赢了一美元",
     "❌ not grounded · Chinese (he won an Oscar, not 'one dollar')"],
]