Most contract-analysis products do something that, on reflection, should bother anyone who has to rely on them: they invent their own clause categories, train an AI model against those categories, and then publish accuracy numbers — sometimes the famous "95% accuracy" — without ever showing how the categories were defined, how the ground truth was constructed, or what "accuracy" actually measured. The result is a black-box system citing self-graded results against a private rubric.
RevU was built on the opposite premise. Our clause-classification engine uses the CUAD taxonomy — a 41-clause schema developed in 2021 by The Atticus Project, a Stanford-led legal NLP research effort, and published as a public dataset of expertly-annotated commercial contracts. This post explains what CUAD is, why we chose it as the foundation, what limits it has, and what we built on top of it to fit NIL contracts specifically.
What CUAD is
CUAD stands for Contract Understanding Atticus Dataset. It is a corpus of 510 commercial contracts, expert-annotated for 41 categories of clauses that contract attorneys care about most. The categories were chosen by a team of attorneys at Atticus, refined over multiple rounds of annotation, and validated against industry usage. The dataset and benchmarks were published openly in 2021 (arXiv 2103.06268) and have become the de facto standard against which legal-NLP clause-classification systems are evaluated.
The 41 categories are not exotic. They are the contract clauses that show up over and over in commercial agreements: anti-assignment, audit rights, cap on liability, change of control, competitive restriction, covenant not to sue, effective date, exclusivity, expiration date, governing law, ip ownership assignment, insurance, joint ip ownership, license grant, liquidated damages, minimum commitment, most favored nation, no-solicit of customers, no-solicit of employees, non-compete, non-disparagement, non-transferable license, notice period to terminate renewal, post-termination services, price restrictions, renewal term, revenue/profit sharing, rofr/rofo/rofn, source code escrow, termination for convenience, third party beneficiary, uncapped liability, unlimited/all-you-can-eat-license, volume restriction, warranty duration — and a handful of others.
Why a public taxonomy matters
When the underlying taxonomy is published openly, several things follow that don't follow with a proprietary one.
First, the categories are reviewable. Every clause RevU's engine extracts is mapped to a CUAD category, and the category itself has a public definition. If you want to know what "non-compete" means inside our system, you can read the CUAD documentation rather than relying on us to define it consistently.
Second, accuracy is benchmarkable. CUAD has a public test set with annotated ground truth. Multiple academic and commercial systems have published F1 scores against it. Our model's clause-classification accuracy can be evaluated by anyone who wants to do the work, against the same benchmark everyone else in the field uses.
Third, training data is reusable. Our model was fine-tuned on CUAD plus additional NIL-specific contracts annotated to the same schema. If a future research group publishes an updated CUAD release, our taxonomy can extend cleanly — we are not stuck with a proprietary schema that no external work uses.
Why we chose CUAD specifically
Of the academic taxonomies available, CUAD is the largest, most expert-annotated, and most clause-focused. Earlier datasets like LEDGAR are organized at the document level (what kind of contract is this?) rather than the clause level (what does this specific paragraph do?). Datasets focused on specific industries (real estate, employment) are too narrow for general contract analysis. CUAD is industry-agnostic, written by lawyers, and granular enough to drive a clause-by-clause review.
The other option was to invent our own taxonomy. We rejected it for two reasons. First, inventing a new schema would let us optimize for marketing claims — defining the categories to be those we are good at and ignoring the ones we are not. That's exactly the criticism leveled at proprietary contract-analysis products. Second, NIL is changing fast enough that any taxonomy built for it would be obsolete within a year; building on an industry-agnostic foundation lets us add NIL-specific categories incrementally without re-architecting.
What CUAD does not cover (and what we added)
CUAD is industry-agnostic, which is exactly the point and also a limitation. NIL contracts contain clause types that don't fit cleanly into any CUAD category. We have built additional clause categories on top of CUAD specifically for NIL work, and we treat the additions as transparent extensions to the base taxonomy.
Examples of NIL-specific clause categories we have added:
Morality clause structure. CUAD has a generic "termination for cause" category but does not separate out morality language. Our engine classifies morality language separately and applies the M0–M5 ladder (see our morality clause post for the full ladder).
Disclosure-and-compliance clause. CUAD does not have a dedicated category for clauses that govern the athlete's obligations to disclose the contract to school compliance, NIL Go, the FTC, etc. We added this category and built the cross-checks that fire when a contract's disclosure language conflicts with the controlling state regime.
NCAA compliance representations. School-direct contracts and NIL Go-routed contracts now commonly include representations and warranties about compliance with NCAA rules, the House settlement, and applicable state NIL law. CUAD has a generic "representations and warranties" category; we sub-classified the NIL-specific variants.
Dollar-amount inventory. Beyond clause classification, RevU runs a dollar-amount inventory across every contract — every monetary figure is labeled by role (base, bonus, signing, royalty, etc.). This is independent of CUAD and is described in detail on our methodology page.
What this means for athletes reading the output
When RevU's report says your contract contains, for example, a "License Grant — IP ownership assignment" clause, the label corresponds to a publicly-defined CUAD category. You don't have to take our word for what the category means. You can look it up, read the definition, find example clauses from the CUAD dataset, and form your own view about how the language in your specific contract fits.
This matters because contract analysis, ultimately, is decision support. The engine surfaces categories and patterns; the human reading the output decides what to do about them. The more the categories are defined transparently, the easier it is to reason about whether the engine's classification is correct in any given case. A proprietary "AI risk score" is much harder to validate, override, or push back on than a structured clause-level report grounded in published categories.
How CUAD interacts with our scoring
CUAD is a classification taxonomy — it tells you what a clause is, not whether it is good or bad. The good/bad question is what our athlete-friendliness scoring layer handles.
For each clause classified into a CUAD category, RevU's engine evaluates the substantive terms of that clause along multiple dimensions (broad vs. narrow, mutual vs. one-sided, capped vs. uncapped, with or without carve-outs, with or without notice/cure periods, etc.) and produces an athlete-friendliness score on a 1–10 scale and a brand-friendliness score on the same scale. The scoring rubric is documented on our methodology page and is treated as a separately auditable layer on top of CUAD classification.
The separation matters: if you disagree with our score, you can point to the specific CUAD-classified clause and the specific scoring dimension and have a meaningful conversation. With a single proprietary risk score there is nothing to dispute except the score itself.
On the question of "accuracy"
We have publicly committed to not advertising aggregate "accuracy" numbers (e.g., "95% accurate") because the metric is misleading without specification of the test set, the F1 versus exact-match threshold, the clause-category breakdown, and the contract-type breakdown. CUAD provides the public test set against which we can evaluate the engine's classification performance honestly; the methodology page describes the per-category performance the engine has reached on CUAD and on our internal NIL-specific evaluation corpus.
For the curious: the academic side
If you want the underlying research, the foundational CUAD paper is Hendrycks et al., "CUAD: An Expert-Annotated NLP Dataset for Legal Contract Review," published on arXiv in 2021. The dataset itself is available on GitHub under a permissive license. The Atticus Project, which led the annotation effort, maintains a public website with documentation. Multiple follow-up papers in the legal-NLP literature have improved on the CUAD baseline; the field is active and the techniques are still advancing.
The transparency commitment
Building on CUAD is part of a broader commitment we have made about how RevU operates. The methodology page describes the engine's architecture — CUAD classification, dollar-amount inventory, reciprocity schema, M0–M5 morality ladder — at a level of detail that should let a sports-law attorney audit our reasoning. The 10-point checklist and the 8-pattern red-flag posts are the human-readable counterparts to what the engine does mechanically. The goal is for our output to be defensible at attorney-level scrutiny, not just compelling at marketing-page scrutiny.
If you ever see RevU classify a clause in a way that doesn't match how a contract attorney would describe the same clause, we want to hear about it. Email info+support@revu.deals with the contract excerpt and our output, and we'll log it as a benchmark case. CUAD is a public taxonomy; our application of it is meant to be defensible at the same level of scrutiny.