QWED Finance: deterministic verification for financial AI

When an LLM told a customer his Chase card had “$12,889” in rewards, QWED-Finance would have caught the hallucination before it caused a lawsuit.

• ISO 20022 business rules read amounts and currencies from the parsed XML tree, not from raw text (GHSA-mrrj-6m2q-jch9)
• Sanctions screening folds diacritics and look-alike letters and compares separator-free names, so E.V.I.L. CORP or ÉVIL CORP no longer clears an EVIL CORP entry (GHSA-mv2c-jwm9-pfrq)
• QueryGuard rejects MySQL /*! ... */ and MariaDB /*M! ... */ executable comments (GHSA-q8r4-6gpp-5fx2)

​

Why QWED Finance?

• Problem: LLM says “IRR is 12%” (when it’s actually 11.8%)
• Solution: QWED calculates the exact IRR symbolically and either validates or corrects the LLM.

​

Key features

• 11 specialized guards: Compliance, Calendar, Derivatives, Messages, Query, Cross, Bond, FX, Risk, ISO, Trading.
• GitHub Action: CI/CD verifier with optional SARIF output for security dashboards.
• Audit trails: Verification receipts for cross-guard and integration flows. Receipts are unsigned by default: call get_signature(key) and store the returned HMAC-SHA256 signature separately to get a tamper-evident record.
• Deterministic verification: When deterministic engines apply, results are exact rather than probabilistic.

​

The 4 pillars of banking verification

​

Quick example

from qwed_finance import ComplianceGuard

guard = ComplianceGuard()

# Verify AML flagging decision
result = guard.verify_aml_flag(
    amount=15000,        # Over $10k threshold
    country_code="US",
    llm_flagged=True     # LLM flagged it
)

print(result.compliant)  # True ✅
print(result.proof)      # "Z3: LLM decision matches regulatory requirement"

​

Architecture

​

High-level flow

​

Guard selection flow

​

Verification engine stack

​

Payment verification sequence

​

Why not just trust the LLM?

• Hallucinate numbers ($12,889 instead of$2.88)
• Miss compliance thresholds (CTR at $10,000.01)
• Generate malformed XML (rejected by SWIFT)
• Create dangerous SQL (DROP TABLE)

​

Regulatory alignment

• RBI FREE-AI Framework (India 2025)
• BSA/FinCEN (AML/CTR thresholds)
• OFAC (Sanctions screening)
• ISO 20022 (Payment messaging)

“Accuracy alone is not sufficient - transparency, auditability, and defensible decision logic are required.” — India AI Governance Guidelines

​

FinanceVerifier: fail-closed inputs

​

verify_compound_interest — accepted compounding frequencies

from qwed_finance import FinanceVerifier

verifier = FinanceVerifier()

# ✅ Valid frequency
result = verifier.verify_compound_interest(
    principal=10000,
    rate=0.05,
    periods=10,
    compounding="monthly",
    llm_output="$16,470.09",
)

# ❌ Unknown frequency — fail-closed
verifier.verify_compound_interest(
    principal=10000,
    rate=0.05,
    periods=10,
    compounding="weekly",  # not supported
    llm_output="$16,470.09",
)
# raises ValueError:
#   Unknown compounding frequency: 'weekly'.
#   Accepted values: annual, daily, monthly, quarterly, semi-annual

​

Next steps

• The 11 guards - Deep dive into each verification guard
• Compliance and auditing - Receipts and regulatory proof
• QWED UCP: transaction verification for AI commerce - Connect finance verification to commerce transactions
• QWED Open Responses: verified tool calls for AI agents - Verify finance-related agent actions before execution