AI tools make claims.
ClauseGrade checks them.
ClauseGrade is Precision Advisers' independent accuracy-testing service for AI tools. We test models against deterministic answer keys computed in code — not against another model's opinion — and return verdicts that anyone can reproduce. Built for teams whose decisions ride on whether the model's numbers are right.
A vendor demo is not evidence.
Every AI tool aimed at serious work arrives with the same proof: a polished demo, a benchmark the vendor chose, and a promise. For tools whose numbers feed credit decisions, actuarial models, claims, and valuations, that is not evidence. It is marketing. The question that matters — does this tool get our work right? — is almost never the question that gets answered.
ClauseGrade answers it directly. We build the test from the work itself, compute the correct answers in code before any model runs, and score the tool against them. What comes back is not an impression. It is a verdict with a method behind it.
Tasks. Keys. Verdicts.
Tasks drawn from the real work
Test tasks are built from the actual work of the vertical: covenant and borrowing-base calculations, actuarial computations, policy and claims figures, rent-roll and valuation math. Not trivia. Not prompts designed to flatter a model.
Answer keys computed in code
Every task's correct answer is computed deterministically in code before any model is run. The key does not come from another model, a crowd vote, or an expert's afternoon. It is arithmetic, and it can be re-run by anyone.
Verdicts anyone can reproduce
The tool under test runs the same tasks under the same conditions. Its answers are scored against the key, and the result is a number with a method behind it — reproducible by anyone, anywhere, from the same task set.
Why deterministic? Because most AI evaluation is one model grading another, or a person skimming outputs and nodding. Both drift. A deterministic key cannot drift: the borrowing base is what the formula says it is, today and in a year, no matter who asks. That is what makes a ClauseGrade verdict evidence instead of commentary.
Where the numbers have consequences.
Private Credit
Covenant compliance, borrowing-base math, and the financial-figure extraction that lenders are already handing to AI tools — tested against keys computed from the documents and the formulas, not from a model's memory of either.
Actuarial
Computation-heavy actuarial work, where a small numeric error is a large professional problem. Tasks and keys built for the standard the profession already holds itself to.
Insurance
Policy, claims, and underwriting figures tested against keys computed from the policy language and the math. If a tool reads coverage wrong, the test should catch it before a claim does.
Commercial Real Estate
Rent rolls, lease abstractions, and valuation arithmetic — the spreadsheet work AI tools are quietly being handed across the industry, scored against keys that do not get tired.
Custom Benchmarks
Benchmarks built for professional societies, agencies, and standards bodies that need a defensible answer to one question: does this tool get our work right? Your tasks, your standard, our deterministic keys.
ClauseGrade runs on its own site: clausegrade.com. It is a service of Precision Advisers LLC.
Put your AI tool on trial.
Tell us the tool and the work it is being trusted with. We will describe the test we would build, the keys we would compute, and the verdict you would get — before anything is signed.