QWED verification: specification guide

┌─────────────────────────────────────────────────────────────┐
│                    CORRECT USAGE                            │
├─────────────────────────────────────────────────────────────┤
│  Developer provides:  expected_value = "100000 * 1.05**10"  │
│  LLM generates:       answer = "150000"                     │
│  QWED verifies:       150000 ≠ 162889.46 → REJECTED         │
└─────────────────────────────────────────────────────────────┘

┌─────────────────────────────────────────────────────────────┐
│                    INCORRECT USAGE                          │
├─────────────────────────────────────────────────────────────┤
│  LLM generates:       expected_value = "100000 * 1.05**5"   │
│  LLM generates:       answer = "127628"                     │
│  QWED verifies:       127628 ≈ 127628.16 → VERIFIED ❌      │
│  (Hallucination verified because spec was also hallucinated)│
└─────────────────────────────────────────────────────────────┘

from qwed_new.core.verifier import VerificationEngine

engine = VerificationEngine()

# ✅ CORRECT: Developer provides formula
result = engine.verify_compound_interest(
    principal=100000,      # Developer knows this
    rate=0.05,             # Developer knows this  
    time=10,               # Developer knows this
    n=1,                   # Developer knows this
    expected=150000        # LLM claimed this
)
# Result: is_correct=False, calculated=162889.46

# ❌ WRONG: Letting LLM provide the formula
# If LLM says "use rate=0.03" and you trust it, QWED can't help

from qwed_new.core.sql_verifier import SQLVerifier

verifier = SQLVerifier()

# ✅ CORRECT: Developer defines allowed tables
result = verifier.verify_query(
    query="SELECT * FROM users",           # LLM generated
    allowed_tables=["users", "orders"],    # Developer defines
    allowed_columns=["id", "name", "email"] # Developer defines
)

# ❌ WRONG: Letting LLM define what tables are allowed

from qwed_new.core.code_verifier import CodeSecurityVerifier

verifier = CodeSecurityVerifier()

# ✅ CORRECT: Developer defines forbidden patterns
result = verifier.analyze_code(
    code="eval(user_input)",  # LLM generated
    # Verifier has built-in dangerous pattern detection
)
# Result: UNSAFE - eval detected

# The patterns (eval, exec, __import__) are defined by QWED, not LLM

from qwed_new.core.logic_verifier import LogicVerifier

verifier = LogicVerifier()

# ✅ CORRECT: Developer provides premises
result = verifier.verify_conclusion(
    premises=["All humans are mortal", "Socrates is human"],  # Developer
    conclusion="Socrates is mortal"  # LLM claimed
)

# ❌ WRONG: Letting LLM provide premises
# If LLM says "All humans are immortal", QWED will verify wrong logic