Independent AI testing for private credit
Independent tests of whether credit AI gets the covenant right.
ClauseGrade tests AI tools on the work a credit team does every period: covenant tests, add-back caps, borrowing base and NAV facility checks. We score the call, the number, the escalation and the clause it cites, against answer keys computed by code and re-checked independently.
What we test
Operational credit work, not general finance trivia
The suite covers the tasks a portfolio or operations team performs each reporting period, where a wrong reading has a cost.
Maintenance covenants
- Leverage, interest coverage, fixed-charge coverage, liquidity and capex
- Cure periods and near-threshold results
- Step-down schedules and amended versus closing definitions
EBITDA add-backs
- Add-back caps, including grower baskets (the greater of a fixed amount and a percentage of EBITDA)
- Disallowed run-rate items
Borrowing base
- Overadvance checks
- Aging, cross-aging, concentration limits and dilution
- Advance-rate step-downs and inventory sublimits
NAV facilities
- Loan-to-value tests and LTV step-downs
- Eligible-NAV concentration
- Pro forma draw requests
Amendment chains
- Which amendments materially changed the covenant terms, and which were clean-up
Reporting deadlines
- Late and missing deliverables, counted in days
Every borrower, facility and financial statement in the suite is invented. Covenant wording follows formulations common in credit agreements filed publicly with the SEC, written for the suite rather than copied from any single agreement. No lender or borrower data is used to build the tests.
How the tests are built
Built to catch the misreadings that matter
- Traps in both directions.
Many tasks are built so a common misreading gives a wrong answer, such as reading only the fixed leg of a grower basket or the wrong row of a step-down table. Fourteen of the 37 tasks contain no problem at all, so a tool that flags everything is marked down too.
- A real choice of clause.
Each package carries a clause index that includes sections that do not govern the task, so citing the right section is a decision, not a formality.
- Answer keys computed by code.
Expected answers come from our calculation engine, written in ordinary code with exact decimal arithmetic, before any AI system sees the task. No language model writes or grades an answer, so the same response always gets the same score.
- An independent re-check.
A second checker, written separately and sharing no code with the engine, recomputes every key using exact fractions. The two must agree on every task, or the suite is not used. Today they agree on all 37.
- Nothing leaks to the tool under test.
Expected answers are never sent to the system being tested.
How scoring works
Four questions on every answer
A tool can spot a breach and still get the number or the clause wrong. We score each part separately and report it by task and by category.
| Axis | Question | Scoring |
|---|---|---|
| Detect | Did the tool correctly say whether there is a problem: a breach, an over-cap add-back, a missed report, a material amendment? | 1 or 0 |
| Quantify | Did it compute the number within the task's stated tolerance? | 1 or 0 |
| Escalate | Did it correctly decide whether a person needs to be alerted? | 1 or 0 |
| Cite | Did it point to the governing clause? | 1 exact, 0.5 right section but wrong subsection, 0 wrong or missing |
A task's score is the average of the four axes. A run passes only when all four are fully correct. Each task runs three times by default. The scorecard shows each task's pass rate across runs, the spread of its answers and a consistency flag, plus a suite-wide pass rate with a 95% confidence interval.
Worked example: a NAV facility draw request
One of the 37 tasks. The borrower is fictional.
The task
Larkspur Secondaries Fund II SPV asks for a $30.0M draw on its NAV facility. The tool must test the loan-to-value ratio after the draw, say whether the draw can be funded as requested, decide whether to escalate, and cite the governing clause.
- Outstanding loans
- $230.0M
- Net asset value
- $500.0M
- Requested draw
- $30.0M
- Maximum LTV
- 50%
Section 2.03(b), Conditions to each Borrowing: No Loan shall be made if, after giving pro forma effect to such Loan and the use of proceeds, the Loan-to-Value Ratio would exceed the maximum permitted under Section 7.15.
The package also contains Section 7.15, the LTV covenant itself, along with other sections that do not govern this question.
The answer key
- Before the draw:
230.0 / 500.0 = 46.0%, inside the 50% limit. - After the draw:
(230.0 + 30.0) / 500.0 = 52.0%, over the limit. - The draw cannot be funded as requested. The most that can be drawn is
$20.0M. - Escalate: yes.
- Governing clause:
Section 2.03(b). Tolerance on the ratio is ±0.5 percentage points.
| Response | Detect | Quantify | Escalate | Cite | Task score |
|---|---|---|---|---|---|
| Tests today's balance only: 46.0%, approve the draw, cites Section 7.15 | 0 | 0 | 0 | 0 | 0.00, fail |
| 52.0%, refuse and escalate, cites Section 7.15 | 1 | 1 | 1 | 0 | 0.75, fail |
| 52.0%, refuse and escalate, cites Section 2.03(b) | 1 | 1 | 1 | 1 | 1.00, pass |
The responses are illustrations, scored by the same code that scores real runs.
Who it's for
Evidence both sides of the table can use
AI vendors
When a lender asks how accurate your tool is and how you know, a scorecard from an outside lab answers with task-level detail.
- See where your tool fails, and why, before a customer finds it
- Re-test after a model, prompt or pipeline change and compare against your saved baseline on the same suite version
- Scorecards are private to you. We publish nothing about a named vendor without written consent
Lenders and credit funds
If your team relies on an AI tool for covenant, borrowing base or NAV work, an independent test record shows you checked it.
- Ask a vendor for an independent scorecard, or commission a test of a tool you use or are evaluating
- See results by axis and category, not one headline accuracy figure
- Re-test when the vendor changes the model behind the tool
And don’t wait for regulators to hold your hands through this shift, because you may receive guidance that is both too limited and too late.Brian Daly, Director of the SEC Division of Investment Management, speaking about advisers adopting AI (his remarks use the term “SI”), remarks at the Stanford University Private Capital Roundtable, October 16, 2026. Given in his official capacity; not necessarily the views of the Commission.
Pilot
A fixed-fee pilot on the full suite
- Scope.
We agree which configuration of your tool is tested: model, prompts and pipeline. We can connect to an OpenAI- or Anthropic-compatible API endpoint, or to a command that takes each task as JSON and returns its answer as JSON.
- Run.
Your tool works through all 37 tasks, three runs each. It never sees the expected answers.
- Scorecard.
You receive a private scorecard recording the suite version, model identifiers, run count and timestamps, every expected answer, your tool's answer, and the reason for each miss.
- Re-test.
When your model, prompt or pipeline changes, we can re-run the same suite version and show what moved against your baseline.
How we stay independent
- Fixed feesNo success fees, equity, referral or revenue-share arrangements with any vendor we test.
- Private by defaultScorecards go only to the party that commissioned them.
- No paid rankingWe do not sell placement, badges or favourable wording, and we do not publish named league tables.
- No coaching to passWe do not consult for vendors on how to pass our tests. Scores change only if we find an error in our own answer key or test harness, and any such correction is logged.
What a scorecard is not: our tasks are synthetic and cleaner than real document sets, with no scans or OCR errors, so they measure reasoning and calculation on stated terms rather than document parsing. A scorecard describes one configuration at one point in time. It is evidence, not a warranty, certification or regulatory approval.
Contact
Talk to us about a pilot
Tell us what your tool does and which configuration you would like tested. We will reply with a proposed scope.