How the benchmark actually works

This page is the reference for the ClaudeFinanceLab finance AI benchmark: what is tested, how ground truth is computed, how pass and fail are decided, and what the results do and do not prove. It is written to be citable — if you are documenting an AI model for review, you can link this URL and the live results directly.

The three tracks

Track 1 · Financial Accuracy
Are the tool calculations mathematically correct?
66 tests across 10 suites (accounting, valuation, portfolio risk, compliance, market data, crypto, quant, real estate, startup, SEC EDGAR). Each tool function is imported directly as Python — no HTTP layer, no mocking — called with fixed inputs, and its output compared to an analytically computed reference value. Async tools run under asyncio.wait_for() with a 30-second timeout. Pass/fail is deterministic.
Track 2 · Server Directory
Do the MCP servers in the directory work?
Every server in the finance MCP directory — including third-party servers — is run against the same 11 public tests: connectivity (SSE handshake, tools/list within 5 s), discovery (schema completeness, valid JSON Schema), execution (first call returns content within 15 s), and resilience (bad arguments return a structured error). This is the suite behind the certification badge.
Track 3 · LLM Accuracy
Can the model reason about finance?
22 questions across 10 domains, scored by Claude-as-judge using G-Eval-style chain-of-thought reasoning against a written rubric. Pass = overall ≥ 7.0/10. This track is judged, not deterministic — it is reported separately and never mixed into the Track 1 pass count.

How ground truth is computed

Track 1 never checks one model's answer against another model's answer. Every reference value comes from a closed-form formula or a published table:

DomainGround-truth basis
Options pricingBlack–Scholes closed form; Greeks by analytic derivative
Cost of capital / betaHamada equation for relevering; CAPM for cost of equity
DCF & valuationTerminal value = FCF·(1+g)/(WACC−g), discounted analytically; verified to 4 significant figures
DepreciationMACRS schedules from IRS tables
Lease accountingASC 842 present-value calculation
Portfolio riskParametric and historical VaR from the return series

Tests that must call a live external API (EDGAR filings, market-data feeds) are explicitly labelled live. Everything else runs on locally computed ground truth and reproduces with no network access.

Pass, fail, and tolerance

A Track 1 test passes only if the tool output is within tolerance of the reference value. Tolerance is per-test and documented in the scorer: exact-match for integer or categorical outputs (share counts, schedule rows, sanctions hits); a relative band — typically 0.01% — for continuous financial quantities. Any deviation beyond the band is a fail, with the expected and actual values recorded. There is no partial credit and no manual override.

When the benchmark re-runs

The full Track 1 suite runs on every deployment; the current results page always reflects the deployed code. Track 2 re-runs weekly against every directory server so a third-party regression shows up within seven days. The framework lives at evals/runner.py; datasets and scorer code are available on request at evals@claudefinancelab.com.

What the results prove — and what they do not

A green Track 1 board means the tool functions compute the covered calculations correctly, within documented tolerance, reproducibly. It does not prove that a language model calling those tools will select the right tool, pass the right arguments, or interpret the result correctly — that is Track 3's job, and Track 3 is judged, not proven. It also does not cover calculations outside the 66 tests. For a written evaluation of your model or workflow, scoped and mapped to the regime you answer to, see Enterprise validation.

Citing this in a governance file

The published methodology — deterministic tests, analytically-computed ground truth, documented pass criteria, reproducible results — is model-validation evidence under the SR 11-7 framework and maps to the traceability expectations in the EU AI Act. Institutions can reference the public results and this methodology in an AI model governance file. For a report formatted specifically for SR 11-7, SR 26-2, or EU AI Act review, contact us.

See the current numbers: Live benchmark →
Get your tool certified: Certification →

How consistent is this in practice, run to run? See our reproducibility pilot study — real data from repeating the same prompts multiple times.

FEEDBACK