How RevU analyzes NIL contracts

RevU is a contract-analysis tool. It reads a NIL contract, classifies each clause against a published legal-contract taxonomy, extracts every dollar amount with a semantic role label, and surfaces clauses that look one-sided, unusually aggressive, or out of line with comparable deals. It is fast — most contracts process in under 60 seconds — and it is consistent, because every analysis runs through the same pipeline.

What it is not: legal advice. RevU does not tell you whether to sign a contract, whether a clause is enforceable in your state, or what to do if a dispute arises. For any decision that matters — and most NIL decisions matter — an attorney should review the output alongside the contract itself. RevU is built to make that attorney review faster and more focused, not to replace it.

The rest of this page describes the specific primitives the engine relies on, why they were chosen, and what claims we deliberately do not make. Everything below sits inside the scope set in this section — when you read "RevU classifies" or "RevU extracts," read it as "the analysis tool produces a structured output that an attorney or knowledgeable user can verify against the contract text."

Clause classification is the backbone of every contract-review tool, and most vendors keep their taxonomy proprietary. RevU does not. Clauses are classified against the CUAD (Contract Understanding Atticus Dataset) taxonomy — a 41-class, expert-annotated rubric published in a peer-reviewed academic paper. Using CUAD means the categories themselves are public, citable, and stable; if RevU flags a clause as "Anti-Assignment" or "Most Favored Nation," there is an external definition behind that label, not a marketing term we invented.

Hendrycks, D., Burns, C., Chen, A., & Ball, S. (2021). CUAD: An Expert-Annotated NLP Dataset for Legal Contract Review. NeurIPS Datasets and Benchmarks Track. DOI: 10.48550/arXiv.2103.06268.

The 41 CUAD classes most relevant to NIL contracts include Effective Date, Termination for Convenience, Anti-Assignment, Audit Rights, Most Favored Nation, Non-Compete, Liquidated Damages, Cap on Liability, Choice of Law / Governing Law, and Renewal Term. NIL-specific overlays — exclusivity, approval rights, social-media usage rights, appearance and deliverable obligations, and morality clauses — sit on top of the CUAD base, so the analysis can speak in both academic and NIL-native language.

Why this matters: when RevU disagrees with another tool — or with a reviewer's gut — the disagreement can be located precisely in a public rubric. That's how the engine stays auditable.

On top of these 41 base categories, RevU adds a set of NIL-native overlays — clause types CUAD does not name but that NIL deals turn on:

• Morality / conduct clause (graded M0–M5)
• Clawback of paid compensation
• Social-media deliverables & per-post rates
• Content approval rights
• FTC / state / school disclosure obligations
• NIL Go ($600+) reportability & fair-market-value posture

NIL contracts hide money in unusual places. A $500,000 headline number can be undercut by a $400,000 buyout penalty triggered if the athlete transfers schools, or padded by a "$50 per Instagram post" figure that buries the deliverable count three paragraphs later. Regex-based extractors miss this regularly. So RevU does not use one.

Instead, every $ in the contract is scanned and labeled with a semantic role: stated total compensation, individual installment, buyout penalty (athlete pays brand or school), buyout remedy (brand pays athlete), liquidated damages, IRS reporting threshold, performance bonus, expense cap, or tax obligation. Each labeled amount carries its source clause, page, character range, surrounding context, a confidence score, and the method used to assign the role.

Totals are then derived from the inventory: total_guaranteed sums the guaranteed-role amounts; total_potential sums the same set plus best-case bonuses. The inventory is the single source of truth; no other code path may write compensation values. We deleted the older regex extractor and the GPT-vs-regex reconciliation pass for that reason.

The inventory fails closed. If the scan finds no amounts, or if the labeling cannot meet a confidence threshold, RevU does not invent a number — it returns null compensation values with a quality-issue flag. Invariants enforce the basics: total_guaranteed ≤ total_potential, cash totals must reconcile, no buyout amount can be greater than the stated contract total without an explanation. When the data disagrees with itself, the analysis surfaces the conflict instead of hiding it.

NIL contracts have a recurring asymmetry problem. A clause reads, "Either party may terminate this agreement for material breach upon 30 days' written notice and an opportunity to cure," and it looks mutual. But sub-clauses bury the catch: the brand's notice period is 30 days; the athlete's is 5. The brand gets a 30-day cure period; the athlete gets none. The brand can terminate for "any material breach"; the athlete is limited to non-payment. On the surface, two equally-sized paragraphs. In practice, two completely different rights.

RevU's reciprocity schema runs each paired obligation through a structured symmetry check: notice period, cure rights, monetary cap, scope of triggering events, who bears the burden of proof, who carries the cost of any required process, whether either side has a unilateral approval right, and whether cure or remedy provisions are reciprocal. Any dimension that fails the symmetry test flags the clause as fake-symmetric — and the analysis surfaces both the comparison and the specific dimensions that don't match.

The point is not to demand perfect symmetry — most commercial contracts asymmetrically favor the party with the bigger lawyer. The point is to make the asymmetry visible so the athlete (and their attorney) can decide whether it's negotiable.

Morality clauses are not binary. A clause that lets a brand terminate the deal if the athlete is "convicted of a felony" is very different from one that lets them terminate "in their sole and absolute discretion" if the athlete's conduct, in the brand's opinion, harms its reputation. Both are morality clauses. They are not the same risk.

RevU grades morality clauses on a six-tier ladder:

• M0 — no morality clause present, or unspecified.
• M1 — objective triggers only (e.g., final criminal conviction), with cure or remedy rights and reciprocity with the brand.
• M2 — primarily objective triggers, with cure period, but limited reciprocity.
• M3 — mixed objective and subjective triggers, partial cure rights, one-sided.
• M4 — primarily subjective triggers ("brings the brand into disrepute") without meaningful cure rights.
• M5 — fully subjective and unilateral; brand may terminate in its sole discretion with no cure right.

Alongside the tier, RevU records the structural features that drove the rating: whether pre-deal conduct is covered, whether family members are in scope, whether political speech has a carveout, whether the burden of proof sits with the brand or the athlete, and whether termination is reciprocal. An athlete looking at an M4 or M5 should understand they're carrying real exposure, and an attorney can use the tier as a quick anchor for negotiation priorities.

Scores are derived, not vibes. For each clause, the engine assembles a structured set of inputs and produces a structured set of outputs, each output traceable back to the clause text it came from. Nothing is a single opaque number — the inputs are inspectable and the outputs carry their evidence.

At the contract level, the per-clause outputs roll up into the figures a reviewer actually acts on:

The 1–10 scores are intentionally coarse anchors for a conversation, not precise measurements — the value is in the grounded evidence and the relative comparison across clauses, not in treating an "athlete friendliness 4" as a scientific constant.

The honest part. These are deliberate limitations, not features we haven't built yet:

We list these because credibility comes from saying what you won't do, not just what you will. The CUAD taxonomy, dollar inventory, reciprocity schema, and morality ladder make the analysis defensible. The boundaries above make the analysis honest.

Where the engine is weak, it should be discounted — not trusted by default. These are the constraints we know about, what each one means in practice, and what to do instead.