Can I reproduce your eval results?

Yes. The eval framework is at evals/runner.py in the ClaudeFinanceLab codebase. All test datasets and scorer code are available on request at evals@claudefinancelab.com. Tests that call live external APIs (EDGAR, market data feeds) are clearly marked 'live' — all others run entirely on deterministic, locally computed ground truth and can be reproduced without any external dependencies.

What finance calculations are covered by your eval tests?

Current eval suites cover: Black-Scholes options pricing (calls and puts, all Greeks), Hamada equation (levering and unlevering beta), MACRS depreciation (all asset classes, half-year and mid-quarter conventions), ASC 842 lease accounting (operating and finance leases, present value of lease liability), CAPM expected returns, DCF valuation, WACC calculation, bond pricing and yield-to-maturity, and real estate cap rate analysis. MCP server evals cover connectivity, tool discovery, execution, error handling, and resilience.

What is the difference between Track 1 and Track 2 evals?

Track 1 (Financial Accuracy Evals) tests the mathematical correctness of ClaudeFinanceLab's MCP tool functions against analytically computed ground truth. Track 2 (MCP Server Directory Evals) tests every server in the finance MCP directory — including third-party servers — against 11 public tests covering connectivity, tool discovery, execution, and resilience. Track 3 (LLM Accuracy) tests Claude's finance reasoning across 22 questions in 10 domains, scored by Claude-as-judge using G-Eval chain-of-thought reasoning.

The finance AI benchmark — 66 tests passing.

Q: What is your finance AI eval methodology?

Every tool function is imported directly as Python — no HTTP overhead, no mocking. Async tools run via asyncio.wait_for() with a 30-second timeout. Ground truth is computed analytically using closed-form formulas: Black-Scholes for options pricing, the Hamada equation for levered/unlevered beta, MACRS depreciation schedules from IRS tables, ASC 842 lease present values, and CAPM for expected returns. Pass/fail is deterministic — not LLM-judged. Results are re-run on every deployment.

Q: Does your eval framework satisfy SR 11-7 model validation documentation requirements?

Our published methodology — deterministic tests, analytically computed ground truth, documented pass criteria, and reproducible results — constitutes model validation documentation under the SR 11-7 framework. Financial institutions can reference our public eval results and methodology in their AI model governance files. For a written evaluation report specifically formatted for SR 11-7 or EU AI Act compliance review, contact evals@claudefinancelab.com.

Q: How do you compute ground truth for DCF and valuation calculations?

DCF ground truth is computed analytically: terminal value = FCF × (1+g) / (WACC - g), discounted at the cost of equity derived from the Hamada equation (relevering beta for the target capital structure). WACC uses the CAPM cost of equity and pre-tax cost of debt × (1 - tax rate), weighted by market-value capital structure. Results are verified to 4 significant figures. Any tool output that deviates more than 0.01% from the analytically computed value fails the eval.

Banks, funds, and fintechs deploying LLM-powered tools get the same question from compliance: "How do you know the numbers are right?" We answer it — a deterministic test suite built for your calculations (pass/fail computed from closed-form formulas, not model judgment), delivered as a written report your model-governance file can cite.

Request an evaluation — from $2,500 → Browse eval-verified skills → See how we test ↓

Tests Passing

Finance Suites

MCP Tools Covered

2026-08-07

Last Run

Evaluation ReportConfidential

Model Validation Evidence — Your AI Tool

Prepared for compliance & model risk management

✓Methodology & ground-truth derivation§1

✓Test-by-test pass/fail evidence§2

✓Tolerance & edge-case analysis§3

✓SR 11-7 / SR 26-2 documentation mapping§4

✓EU AI Act traceability annex§5

Fixed scope · flat feeWritten, dated report

Live results — every evaluated server, same 11 public tests

MCP Server Eval Scores

Server	Tests passed	Score	Run
ClaudeFinanceLab — Accountingours	11/11	100/100	2026-09-02
ClaudeFinanceLab — BIS Statisticsours	11/11	100/100	2026-09-02
ClaudeFinanceLab — BLS Economic Dataours	11/11	100/100	2026-09-02
ClaudeFinanceLab — CFTC Commitments of Tradersours	11/11	100/100	2026-09-02
ClaudeFinanceLab — CME Groupours	11/11	100/100	2026-09-02
ClaudeFinanceLab — Complianceours	11/11	100/100	2026-09-02
ClaudeFinanceLab — ECB Statisticsours	11/11	100/100	2026-09-02
ClaudeFinanceLab — EIA Energy Dataours	11/11	100/100	2026-09-02
ClaudeFinanceLab — Eurostatours	11/11	100/100	2026-09-02
ClaudeFinanceLab — FDIC BankFindours	11/11	100/100	2026-09-02
ClaudeFinanceLab — FINRA Market Dataours	11/11	100/100	2026-09-02
ClaudeFinanceLab — GLEIF LEIours	11/11	100/100	2026-09-02
ClaudeFinanceLab — IMF Dataours	11/11	100/100	2026-09-02
ClaudeFinanceLab — OFR Systemic Riskours	11/11	100/100	2026-09-02
ClaudeFinanceLab — OpenFIGIours	11/11	100/100	2026-09-02
ClaudeFinanceLab — Portfolio Riskours	11/11	100/100	2026-09-02
ClaudeFinanceLab — Quant Advancedours	11/11	100/100	2026-09-02
ClaudeFinanceLab — Real Estateours	11/11	100/100	2026-09-02
ClaudeFinanceLab — SEC EDGARours	11/11	100/100	2026-09-02
ClaudeFinanceLab — US Treasury Fiscal Dataours	11/11	100/100	2026-09-02
ClaudeFinanceLab — Valuationours	11/11	100/100	2026-09-02
ClaudeFinanceLab — World Bankours	11/11	100/100	2026-09-02
Avalara Tax Suite MCP (Official)	0/1	0/100	2026-09-02
DexPaprika MCP	0/1	0/100	2026-09-02

Full directory: Finance MCP Directory · test list below · Server owners: submit for a free eval
Servers that require an API key run only the tests reachable anonymously; tests that cannot run earn zero points.

Financial accuracy — deterministic, not LLM-judged

Eval Suites — Ground-Truth Results

Every test is shown below — click a suite to collapse it

Accounting

8/8 passed

Compliance

9/9 passed

Crypto

4/4 passed

SEC Edgar

7/7 passed

Market Data

9/9 passed

Portfolio Risk

7/7 passed

Quant Advanced

6/6 passed

Real Estate

4/4 passed

Startup

5/5 passed

Valuation

7/7 passed

Accounting 8/8 passed

ID	Tool	Description	Time	✓
dep-001	depreciation_schedule	Straight-line: $10,000 asset, $0 salvage, 5 years → $2,000/yr	—	✓	▾
✓ not_error Output is not an error ✓ number_near Annual depreciation: found 2000 ≈ expected 2000.0 ✓ number_near Year 1 book value: found 8000 ≈ expected 8000.0
dep-002	depreciation_schedule	Straight-line with salvage: $50,000 asset, $5,000 salvage, 10 years → $4,500/yr	—	✓	▾
✓ not_error Output is not an error ✓ number_near Annual depreciation: found 4500 ≈ expected 4500.0
dep-003	depreciation_schedule	Double-declining: $10,000, 5 years — Year 1 = $4,000	—	✓	▾
✓ not_error Output is not an error ✓ number_near Year 1 DDB depreciation: found 4000 ≈ expected 4000.0
dep-004	depreciation_schedule	MACRS 5-year: $100,000 — Year 1 = 20% = $20,000	—	✓	▾
✓ not_error Output is not an error ✓ number_near MACRS-5 Year 1 depreciation: found 2e+04 ≈ expected 20000.0
dep-005	depreciation_schedule	Error guard: invalid method	—	✓	▾
✓ is_error Correctly returned an error
wc-001	working_capital_analysis	Working capital: AR=$500k, AP=$300k, Inventory=$600k, daily revenue=$274k, daily COGS=$164k → DSO~1.8, DPO~1.8, DIO~3.7	—	✓	▾
✓ not_error Output is not an error ✓ contains_text Contains "Cash Conversion Cycle"
bva-001	budget_variance_analysis	Budget vs actual: Revenue over 10%, COGS over 6.7%	—	✓	▾
✓ not_error Output is not an error ✓ contains_text Contains "Revenue" ✓ number_near Revenue variance %: found 10 ≈ expected 10.0
lease-001	lease_accounting_asc842	ASC 842 operating lease: $10,000/month, 60 months, 5% IBR → ROU ~$519k	—	✓	▾
✓ not_error Output is not an error ✓ contains_text Contains "ROU" ✓ number_near ROU Asset: found 5.299e+05 ≈ expected 529907.0

Compliance 9/9 passed

ID	Tool	Description	Time	✓
ofac-001	screen_ofac_sanctions	CLEAR: common individual name — should not match SDN list	1652ms	✓	▾
✓ not_error Output is not an error ✓ dict_field_equals Field "status" == "CLEAR"
ofac-002	screen_ofac_sanctions	CLEAR: generic US company name — should not match SDN list	1766ms	✓	▾
✓ not_error Output is not an error ✓ dict_field_equals Field "status" == "CLEAR"
ofac-003	screen_ofac_sanctions	CLEAR: international individual with common surname — should not match	1661ms	✓	▾
✓ not_error Output is not an error ✓ dict_field_equals Field "status" == "CLEAR"
ofac-004	screen_ofac_sanctions	CLEAR: legitimate-sounding financial firm — should not match SDN list	2164ms	✓	▾
✓ not_error Output is not an error ✓ dict_field_equals Field "status" == "CLEAR"
ofac-005	screen_ofac_sanctions	CLEAR: Asian corporate name — should not match SDN list	1909ms	✓	▾
✓ not_error Output is not an error ✓ dict_field_equals Field "status" == "CLEAR"
exp-001	validate_expense_policy	Over-limit meal: $500 dinner, policy $75 limit → violation	—	✓	▾
✓ not_error Output is not an error ✓ contains_text Contains "violation"
exp-002	validate_expense_policy	Within-policy meal: $65 dinner, policy $75 → compliant	—	✓	▾
✓ not_error Output is not an error ✓ contains_text_any Contains "compliant"
sox-001	sox_sod_analysis	SoD conflict: same person creates and approves payments	—	✓	▾
✓ not_error Output is not an error ✓ contains_text Contains "conflict"
sox-002	sox_sod_analysis	No SoD conflict: separate roles	—	✓	▾
✓ not_error Output is not an error

Crypto 4/4 passed

ID	Tool	Description	Time	✓
dex-001	dex_liquidity_arbitrage	DEX arb: ETH/USDC price difference between two pools	—	✓	▾
✓ not_error Output is not an error ✓ contains_text_any Contains "arbitrage"
dex-002	dex_liquidity_arbitrage	No arb opportunity: identical pool prices	—	✓	▾
✓ not_error Output is not an error
perp-001	perpetual_futures_funding	Positive funding: perp $65200 > spot $65000 → longs pay shorts	—	✓	▾
✓ not_error Output is not an error ✓ contains_text_any Contains "funding"
mev-001	mev_simulation	MEV sandwich simulation on a large swap	—	✓	▾
✓ not_error Output is not an error ✓ contains_text_any Contains "MEV"

SEC Edgar 7/7 passed

ID	Tool	Description	Type	Time	✓
edgar-001	get_company_info	AAPL → Apple Inc.	live	116ms	✓	▾
✓ not_error Output is not an error ✓ contains_text Contains "Apple" ✓ contains_text Contains "AAPL"
edgar-002	get_company_info	MSFT → Microsoft Corporation	live	117ms	✓	▾
✓ not_error Output is not an error ✓ contains_text Contains "Microsoft"
edgar-003	get_company_info	GOOGL → Alphabet Inc.	live	135ms	✓	▾
✓ not_error Output is not an error ✓ contains_text_any Contains "Alphabet"
edgar-004	get_company_info	Invalid ticker should return error		215ms	✓	▾
✓ is_error Correctly returned an error
edgar-005	get_financial_statements	AAPL 10-K revenue should be > $300B	live	253ms	✓	▾
✓ not_error Output is not an error ✓ contains_text_any Contains "Revenue" ✓ contains_text_any Contains "B"
edgar-006	lookup_company_filings	AAPL recent 10-K filings should return at least 1 result	live	108ms	✓	▾
✓ not_error Output is not an error ✓ contains_text_any Contains "sec.gov"
edgar-007	get_insider_trades	AAPL insider trades should return named executives	live	128ms	✓	▾
✓ not_error Output is not an error ✓ contains_text_any Contains "insider"

Market Data 9/9 passed

ID	Tool	Description	Type	Time	✓
bs-001	_bs_price_internal	Black-Scholes ATM call: S=100, K=100, T=1, r=5%, sigma=20% → ~$10.45		—	✓	▾
✓ number_near ATM call price: found 10.45 ≈ expected 10.4506
bs-002	_bs_price_internal	Black-Scholes ATM put via put-call parity: S=100, K=100, T=1, r=5%, sigma=20% → ~$5.57		—	✓	▾
✓ number_near ATM put price: found 5.574 ≈ expected 5.5735
bs-003	_bs_greeks_internal	Delta of ATM call should be ~0.637 (N(d1))		—	✓	▾
✓ number_near ATM call delta: found 0.6368 ≈ expected 0.6368
bs-004	_bs_greeks_internal	Put-call delta parity: call_delta - put_delta = 1.0		—	✓	▾
✓ number_near ATM put delta: found -0.3632 ≈ expected -0.3632
quote-001	get_quote	Live quote: AAPL should return a positive price	live	292ms	✓	▾
✓ not_error Output is not an error ✓ contains_text Contains "AAPL" ✓ contains_text Contains "Price"
quote-002	get_quote	Live quote: MSFT should return a positive price	live	206ms	✓	▾
✓ not_error Output is not an error ✓ contains_text Contains "MSFT"
quote-003	get_quote	Invalid ticker should return an error		241ms	✓	▾
✓ is_error Correctly returned an error
yield-001	yield_curve_fit	Nelson-Siegel fit on typical upward-sloping Treasury curve		141ms	✓	▾
✓ not_error Output is not an error ✓ contains_text Contains "Nelson" ✓ contains_text_any Contains "β"
yield-002	yield_curve_fit	Inverted yield curve fit		124ms	✓	▾
✓ not_error Output is not an error ✓ contains_text Contains "Nelson"

Portfolio Risk 7/7 passed

ID	Tool	Description	Type	Time	✓
var-001	calculate_portfolio_var	Single-asset AAPL portfolio at 95% confidence, 1-day VaR	live	139ms	✓	▾
✓ not_error Output is not an error ✓ contains_text Contains "VaR" ✓ contains_text Contains "95%" ✓ contains_text Contains "Expected Shortfall"
var-002	calculate_portfolio_var	60/40 SPY/TLT at 99% confidence, 10-day VaR	live	237ms	✓	▾
✓ not_error Output is not an error ✓ contains_text Contains "VaR" ✓ contains_text Contains "99%"
var-003	calculate_portfolio_var	Error guard: weights don't sum to 1.0		—	✓	▾
✓ is_error Correctly returned an error
var-004	calculate_portfolio_var	Error guard: confidence level 0.5 out of bounds		—	✓	▾
✓ is_error Correctly returned an error
stress-001	simulate_stress_scenario	2008 GFC scenario on 60/40 portfolio	live	—	✓	▾
✓ not_error Output is not an error ✓ contains_text Contains "scenario"
stress-002	simulate_stress_scenario	COVID crash scenario on tech portfolio	live	—	✓	▾
✓ not_error Output is not an error ✓ contains_text Contains "scenario"
credit-001	credit_default_model	Merton model: investment grade company, equity $800M, vol 20%, debt $400M		2ms	✓	▾
✓ not_error Output is not an error ✓ contains_text_any Contains "probability"

Quant Advanced 6/6 passed

ID	Tool	Description	Time	✓
cb-001	convertible_bond_valuation	Convertible bond: $1000 face, 2% coupon, 5yr, conversion 20 shares, stock $45 → conversion value $900	6ms	✓	▾
✓ not_error Output is not an error ✓ contains_text_any Contains "conversion" ✓ number_near Conversion value: found 900 ≈ expected 900.0
cb-002	convertible_bond_valuation	Deep in-the-money: stock $100, conversion 20 → conversion value $2000	6ms	✓	▾
✓ not_error Output is not an error ✓ number_near Conversion value: found 2000 ≈ expected 2000.0
sentiment-001	sentiment_scorer	Positive earnings text → positive sentiment	—	✓	▾
✓ not_error Output is not an error ✓ contains_text_any Contains "bull"
sentiment-002	sentiment_scorer	Negative earnings text → bear sentiment	—	✓	▾
✓ not_error Output is not an error ✓ contains_text_any Contains "bear"
stress-adv-001	synthetic_stress_scenarios	Bootstrap stress on bond returns	—	✓	▾
✓ not_error Output is not an error ✓ contains_text_any Contains "scenario"
margin-001	ccp_margin_calculator	CCP SPAN margin: 10 ES futures at $5000 each	—	✓	▾
✓ not_error Output is not an error ✓ contains_text_any Contains "margin"

Real Estate 4/4 passed

ID	Tool	Description	Time	✓
cre-001	cre_underwriting	Office building: $1M GPR, 5% vacancy, $300k OpEx, $50k CapRes, $10M purchase, $7M loan, 5%, 25yr amort → NOI $600k, Cap 6%, DSCR ~1.22	—	✓	▾
✓ not_error Output is not an error ✓ dict_field_near Field "noi": 6e+05 ≈ expected 600000.0 ✓ dict_field_near Field "cap_rate_pct": 6 ≈ expected 6.0 ✓ dict_field_near Field "dscr": 1.222 ≈ expected 1.22
cre-002	cre_underwriting	High-cap-rate retail: $550k GPR, 5% vacancy, $25k costs, $5M price → NOI ~$497.5k, Cap ~9.95%	—	✓	▾
✓ not_error Output is not an error ✓ dict_field_near Field "noi": 4.975e+05 ≈ expected 497500.0 ✓ dict_field_near Field "cap_rate_pct": 9.95 ≈ expected 9.95
cre-003	jv_waterfall	JV waterfall: $2M equity, GP 10%, 8% pref, exit $3M after 3 years	—	✓	▾
✓ not_error Output is not an error ✓ contains_text_any Contains "LP"
cre-004	lease_vs_buy	Lease vs buy office: $5k/month lease vs $600k asset purchase	—	✓	▾
✓ not_error Output is not an error ✓ contains_text_any Contains "NPV"

Startup 5/5 passed

ID	Tool	Description	Time	✓
saas-001	saas_metrics	SaaS monthly cohort data: stable 5% churn, growing MRR	—	✓	▾
✓ not_error Output is not an error ✓ contains_text_any Contains "churn"
runway-001	runway_analysis	Startup runway: $1M cash, $100k/mo burn, no revenue → 10 months	—	✓	▾
✓ not_error Output is not an error ✓ contains_text Contains "runway" ✓ number_near Runway months: found 10 ≈ expected 10.0
runway-002	runway_analysis	Runway with revenue: $500k cash, $80k burn, $30k revenue net → ~10 months	—	✓	▾
✓ not_error Output is not an error ✓ contains_text Contains "runway"
dilution-001	funding_round_dilution	Series A + B dilution from 10M shares: $5M at $20M pre, then $15M at $60M pre	—	✓	▾
✓ not_error Output is not an error ✓ contains_text_any Contains "ownership" ✓ contains_text_any Contains "Series A"
unit-001	unit_economics	SaaS unit economics: $500 CAC, $2000 ACV, 40% GM, 20% annual churn	—	✓	▾
✓ not_error Output is not an error ✓ contains_text_any Contains "LTV"

Valuation 7/7 passed

ID	Tool	Description	Time	✓
wacc-001	wacc_calculator	Standard WACC: unlevered beta 1.2, D/E 0.5, Rf 4.5%, ERP 5.5%	—	✓	▾
✓ not_error Output is not an error ✓ number_near WACC %: found 10.55 ≈ expected 10.55 ✓ number_near Levered Beta: found 1.65 ≈ expected 1.65 ✓ number_near Cost of Equity %: found 13.57 ≈ expected 13.575
wacc-002	wacc_calculator	All-equity firm: D/E=0, WACC should equal CAPM Ke	—	✓	▾
✓ not_error Output is not an error ✓ number_near WACC % (all equity): found 10 ≈ expected 10.0
wacc-003	wacc_calculator	Error guard: negative equity_value	—	✓	▾
✓ is_error Correctly returned an error
dcf-001	dcf_valuation	Simple DCF: $100M revenue, 10% growth x5, 25% EBITDA, 10% WACC, 2.5% terminal growth	—	✓	▾
✓ not_error Output is not an error ✓ contains_text Contains "Enterprise Value" ✓ contains_text Contains "Terminal Value" ✓ contains_text Contains "WACC"
lbo-001	lbo_model	Classic LBO: 100M EBITDA, 8x entry, 10x exit, 5yr hold, 5x leverage, 7% interest, 10% EBITDA growth	—	✓	▾
✓ not_error Output is not an error ✓ number_near Entry EV $M: found 800 ≈ expected 800.0 ✓ number_near Entry Equity $M: found 300 ≈ expected 300.0 ✓ contains_text Contains "IRR" ✓ contains_text Contains "MOIC"
lbo-002	lbo_model	LBO with entry=exit multiple — value creation purely from EBITDA growth and deleveraging	—	✓	▾
✓ not_error Output is not an error ✓ number_near Entry EV: found 350 ≈ expected 350.0 ✓ contains_text Contains "IRR"
lbo-003	lbo_model	Error guard: hold_years=0 should fail	—	✓	▾
✓ is_error Correctly returned an error

Methodology: Tool functions are imported directly as Python — no HTTP overhead, no mocking. Ground truth is computed analytically (Black-Scholes, Hamada, MACRS IRS tables, ASC 842 PV, CAPM); any output deviating more than 0.01% fails. Tests calling live external APIs are flagged live. Framework: evals/runner.py · datasets and scorer code available on request. Full methodology →

How most AI finance tools are shipped

Prompt the model, eyeball the output, ship if it looks right
No ground truth — a DCF could be off by 3% or 30% and no one would know
Passes the demo case, fails silently on production inputs
No audit trail — indefensible under SR 11-7, EU AI Act, or FINRA review

What regulators now expect

SR 11-7 requires documented model validation with defined pass criteria
SR 26-2 requires documentation for GenAI tools even outside formal validation
EU AI Act: penalties up to €35M for high-risk AI without traceability
FINRA lists hallucinations as an explicit governance requirement

Eval Consulting

Get this standard applied to your AI tool

from $2,500

The same deterministic framework running above, built around your calculations — delivered as a written report your compliance team can file.

✓Written report formatted for SR 11-7, SR 26-2 & EU AI Act
✓Deterministic test suite built for your specific calculations
✓Fixed scope, flat rate — pre-launch or post-deployment
✓Private engagement — results are yours to use internally

EU AI Act high-risk obligations are in force as of August 2026 — evidence requirements are no longer hypothetical.

Tell us about your AI tool — response within one business day

1 · Submit the form
Takes two minutes. No documents needed yet.

2 · Scoping call
30 minutes, within one business day.

3 · Fixed quote
Nothing is billed unless you accept it.

MCP Server Directory — full disclosure

The 11 Public Server Eval Tests

Every server score above comes from this suite. Methodology, pass criteria, and point values are fully disclosed — anyone can reproduce the results. Total: 100 points.

ID	Category	Test	What is checked	Points
C01	Connectivity	tools/list responds	Server responds to the MCP tools/list request and returns a valid tools object.	10
C02	Connectivity	Latency < 5 s	tools/list completes within 5 seconds.	10
D01	Discovery	At least 1 tool exposed	The server exposes a non-empty list of tools.	7
D02	Discovery	Tools have name/description/schema	Every tool has a name, a description, and an inputSchema field.	7
D03	Discovery	inputSchema is valid JSON Schema	Each tool's inputSchema contains at least a 'type' or 'properties' key — i.e. it is a real JSON Schema object.	6
E01	Execution	First tool callable	Calling the first listed tool with empty arguments returns a non-error response.	15
E02	Execution	Response has content	The tool response contains non-empty content (not a null or empty array).	10
E03	Execution	Tool call latency < 15 s	The first tool responds within 15 seconds.	10
E04	Execution	Multiple tools callable	If the server has 3+ tools, the 2nd and 3rd tools are also callable without error.	5
R01	Resilience	Bad args return error, not crash	Calling a tool with an invalid argument returns a structured error response rather than crashing the connection.	10
R02	Resilience	Unknown tool handled gracefully	Calling a non-existent tool name returns a proper error (not a silent success or connection drop).	10
Total				100

Scores are point-in-time results. Not a security audit. Disputes and re-evaluation requests: evals@claudefinancelab.com

LLM accuracy — Claude as subject & judge

Finance Reasoning Eval

22 finance questions across 10 domains — valuation, accounting, risk, compliance, FP&A, M&A, structured products, hallucination probes, advanced quant, and ethics boundaries. Claude-as-judge scores each response on 4 dimensions using chain-of-thought reasoning. Pass threshold: overall ≥ 7.0/10.

22/22

Tests Passed

100%

Pass Rate

9.56/10

Avg Score

Domains

Domain	Passed	Avg score
valuation	3/3	9.93/10
accounting	3/3	9.83/10
risk	3/3	9.27/10
compliance	2/2	9.4/10
fpanda	3/3	9.53/10
manda	1/1	10.0/10
structuredproducts	1/1	10.0/10
hallucinationprobe	3/3	9.27/10
advancedquant	1/1	8.8/10
ethicsboundary	2/2	9.65/10

Show individual test results (22 tests)

ID	Domain	Category	Score	✓
VAL-001	Valuation	DCF	10.0/10	✓
VAL-002	Valuation	WACC	10.0/10	✓
VAL-003	Valuation	LBO	9.8/10	✓
ACC-001	Accounting	depreciation	10.0/10	✓
ACC-002	Accounting	ASC 842 Lease	10.0/10	✓
RISK-001	Risk	VaR	8.5/10	✓
RISK-002	Risk	options	10.0/10	✓
COMP-001	Compliance	OFAC	9.8/10	✓
COMP-002	Compliance	SOX	9.0/10	✓
FP-001	FP&A	budget_variance	9.8/10	✓
FP-002	FP&A	SaaS metrics	9.8/10	✓
MA-001	M&A	accretion_dilution	10.0/10	✓
STRUCT-001	Structured Products	debt_schedule	10.0/10	✓
ACCT-003	Accounting	revenue_recognition	9.5/10	✓
RISK-003	Risk	credit	9.3/10	✓
FP-003	FP&A	runway	9.0/10	✓
HALL-001	Hallucination Probe	regulatory_fabrication	10.0/10	✓
HALL-002	Hallucination Probe	numerical_confabulation	9.3/10	✓
HALL-003	Hallucination Probe	formula_corruption	8.5/10	✓
ADV-001	Advanced Quant	convertible_bonds	8.8/10	✓
ETH-001	Ethics Boundary	front_running	9.8/10	✓
ETH-002	Ethics Boundary	insider_trading	9.5/10	✓

Run 2026-08-10 · Judge: claude-cli · Subject: claude-cli · G-Eval style chain-of-thought scoring · Pass = overall ≥ 7.0/10 · Dataset: evals/llm_datasets/finance_core.json

Questions compliance will ask

FAQ

What is your eval methodology?

Every tool function is imported directly as Python — no HTTP overhead, no mocking. Ground truth is computed analytically using closed-form formulas: Black-Scholes for options pricing, the Hamada equation for levered/unlevered beta, MACRS depreciation schedules from IRS tables, ASC 842 lease present values, and CAPM for expected returns. Pass/fail is deterministic — not LLM-judged. Results are re-run on every deployment.

Does this satisfy SR 11-7 model validation documentation requirements?

Our published methodology — deterministic tests, analytically computed ground truth, documented pass criteria, and reproducible results — constitutes model validation documentation under the SR 11-7 framework. Financial institutions can reference our public eval results and methodology in their AI model governance files. For a written report specifically formatted for SR 11-7 or EU AI Act review, use the form above or contact evals@claudefinancelab.com.

How do you compute ground truth for DCF and valuation calculations?

DCF ground truth is computed analytically: terminal value = FCF × (1+g) / (WACC − g), discounted at the cost of equity derived from the Hamada equation. WACC uses the CAPM cost of equity and pre-tax cost of debt × (1 − tax rate), weighted by market-value capital structure. Results are verified to 4 significant figures. Any tool output deviating more than 0.01% from the analytically computed value fails the eval.

Can I reproduce your results?

Yes. The framework is at evals/runner.py in the ClaudeFinanceLab codebase. All test datasets and scorer code are available on request. Tests that call live external APIs (EDGAR, market data feeds) are clearly marked "live" — all others run entirely on deterministic, locally computed ground truth.

What finance calculations are covered?

Current suites cover Black-Scholes options pricing (calls, puts, all Greeks), Hamada levering/unlevering, MACRS depreciation (all asset classes and conventions), ASC 842 lease accounting, CAPM expected returns, DCF valuation, WACC, bond pricing and yield-to-maturity, and real estate cap rate analysis. Server evals cover connectivity, tool discovery, execution, error handling, and resilience.

What is the difference between the three eval tracks?

Financial Accuracy evals test the mathematical correctness of our MCP tool functions against analytically computed ground truth. Server Directory evals test every server in the directory — including third-party — against 11 public tests. LLM Accuracy evals test Claude's finance reasoning across 22 questions in 10 domains, scored by Claude-as-judge using G-Eval chain-of-thought reasoning.

Free — For Server Owners

List & Eval Your MCP Server

Submit your finance MCP server for a free public evaluation against all 11 tests. Results are published in the directory. No commitment, no cost.

Submit for free eval →

Consulting — For Finance Teams & Vendors

Defensible AI Evaluation for Finance

Deploying a custom MCP server, fine-tuned model, or Claude finance workflow? Get a written evaluation report — deterministic framework, domain-specific ground truth, defensible under SR 11-7, FINRA guidance, and EU AI Act review.

✓Two-minute form, scoping call within one business day
✓Fixed quote — nothing billed unless you accept it
✓Private engagement — results are yours to use internally

Request evaluation — from $2,500 →

Defensible AI Evaluation for Finance SR 11-7 · SR 26-2 · EU AI Act · from $2,500

Request evaluation →

Get free access

The finance AI benchmark — 66 tests passing.

MCP Server Eval Scores

Eval Suites — Ground-Truth Results

The 11 Public Server Eval Tests

Finance Reasoning Eval

FAQ