WALDHORN.AIThe Review

No. 02Technology

The secret isnot the model.

A model is one part of it. The rest is what no model carries in its memory: the law of the place you sign, read from its sources and dated; the judgment of the music business; and a record no one can quietly change.

Abstract

Asked specific, verifiable questions about real court cases, general-purpose models got the law wrong between 58 and 88 per cent of the time, and did not reliably know when they had.1 The research tools built for lawyers on top of such models, tested the same way, still invented answers between a sixth and a third of the time.2 A more capable model does not close that gap. It writes the wrong answer more fluently.

Nor does a larger pile of contracts. A library of past agreements tells a model what other people signed, under other laws; it cannot say what the law requires where you sign, or what your deal needs. WALDHORN.AI is built from those instead: what a deal must answer, what the law of the place it is signed requires, and what a review must catch.

38
kinds of music agreement it drafts
2,577
questions it can put about a deal, up to 118 for a single one
60
places whose contracts get verified rules, and every rule dated
456
known traps in the review's playbook, written by hand

The model reads and writes; what the law requires, the engine computes. This piece shows what it does at each step of a deal, from the first reading to the seal. How it does it is ours.

Computed, not remembered

Before anything is drafted or reviewed, the engine works out what the answer has to be.

A general model answers from what it absorbed in training, fluently, and without accounting for where the law it states came from. Legal competence is not one skill for such a model either: when lawyers and computer scientists set out to measure it, LegalBench took 162 tasks across six kinds of legal reasoning, and the models were uneven across them.3 Contracts are harder still. CUAD, a set of more than 500 commercial contracts annotated by lawyers, found models' contract review “nascent”,4 and on long documents models reliably miss what sits in the middle of the text.5 A recording agreement is a long document, and the clause that follows an artist for a career is rarely on the first page.

So the engine does not ask a model what the law is. Before a word is drafted or a clause is graded, it computes what the answer must satisfy:

  1. which law governs the agreement, and what that law inherits;
  2. which of that law's rules apply to this kind of agreement;
  3. which of those this deal's own terms bring into play;
  4. and what the agreement must therefore contain.

The result is determinate. The same deal under the same law yields the same requirements, every time, and every one of them traces to a rule with a source and a date. The model reads and writes within those requirements; it does not choose them.

That law can be computed is established science. In 1986 a team at Imperial College wrote the British Nationality Act as a logic program and ran it;6 Catala, a programming language designed for the law, turns statutes into code with a formal meaning;7 and Surden set out how contract terms can be written so that a computer can assess compliance with them.8 What this engine brings to it is the music business: its deals, its terms, and the law of the places where they are signed.

Read, graded, and pinned to the words

Upload a contract, as a PDF, a scan or a photo, and it is read as pages, the way counsel reads a hard copy.

Every clause is weighed against how the music business actually works, and graded: red for a term that can follow a career for decades, amber for a material risk worth negotiating, green for a term that is standard or in your favor. Each finding is pinned to the exact words it concerns, with the reason and the fix beside it.

The review leads to an annotated document you can keep and share or, on business and enterprise accounts, to a research memorandum that sets the clause against the law of the jurisdictions in play, every authority named.9

Figure 1.A page of a real deal: each finding marked on its words and graded, then carried to the annotated document or, on business and enterprise accounts, to the research.

Drafted for the place it will be signed

A new agreement starts from a guided intake: the deal in plain words, field by field.

From there it is drafted through the law of the place it will be signed, the rules that govern each kind of clause, and every term of the deal, resolved together rather than one after another. What comes out is written for the world the deal will live in: consent for AI training and voice cloning, streaming-era royalties, and reversion with real dates.

The rings of the dial below are the 15 jurisdictions whose law the ledger has read directly. Every rule a draft draws on is tied to the statute or judgment behind it, with the date a person last checked it, and each of them can be opened and checked: see the register.

Figure 2.Every field, every jurisdiction and every rule, resolved into one draft for New York.

Negotiated point by point, on the record

Proposed changes travel to the other side as a redline, and the answers come back against it, point by point: accepted, countered, or declined. The other side responds from the invitation it receives.

A docket keeps where every point stands until the last one closes, so nothing agreed along the way is lost, and nothing is agreed that was not seen.

Figure 3.Proposed, answered, countered and agreed, with every point on the record.

Amended against the original, and signed

When the terms change, the amendment is written against the signed original, clause by clause, and bound to it as one instrument, so the agreement and its amendment are always read together.

Both parties sign electronically, and what they sign is sealed at the moment the last signature lands.

Figure 4.Written against the signed original, signed by both sides, and sealed.

Sealed, and impossible to alter unseen

An executed agreement is sealed with a fingerprint taken from every character of it. Anyone holding a copy can check that copy against the record, and the contract itself is never published to do it.

Change one character, a single figure in a royalty rate, and the fingerprint is a different number. The copy no longer matches the record, and the check says so.

Document verification record

✓ Authentic & fully executed

Gilt-Halo-6AF3 Confirm this seal matches the one printed on your document.

SHA-256 6af3af0f477a…f308Sealed

Check a copy Change the figure. Only one matches.

Clause 7.2: a royalty of fifteen per cent (%) of Net Receipts

SHA-256 bc2a09b4135d…df25No matching record

Figure 5.One character changed, and the fingerprint is a different number: the pattern woven from it no longer matches the one on record.

Where a human stays in the loop

A tool used around consequential agreements has to earn trust. Four commitments are built into the workspace rather than bolted on:

  • Conservative under uncertainty. When the answer turns on facts or law it cannot see, it says so and points you to counsel, rather than inventing certainty.
  • No invented law. It does not cite a statute or a case that does not exist. Where it is unsure of an authority, it names none.
  • Your documents are yours. They are processed only to perform the analysis you request, never sold, and not used to train models unless you explicitly opt in.
  • A tool, not a lawyer. WALDHORN.AI is not a law firm and forms no attorney-client relationship. For the decisions that matter, it prepares you for counsel rather than replacing them.

Those are not disclaimers in a footer. They are why a lawyer, and an artist signing a first deal, can rely on what they read here.10

Notes and references

  1. Dahl, Magesh, Suzgun & Ho, 'Large Legal Fictions', Journal of Legal Analysis 16(1) (2024).
  2. Magesh, Surani, Dahl, Suzgun, Manning & Ho, 'Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools', Journal of Empirical Legal Studies (2025).
  3. Guha, Nyarko, Ho, Ré, Chilton et al., 'LegalBench: A Collaboratively Built Benchmark for Measuring Legal Reasoning in Large Language Models', Advances in Neural Information Processing Systems 36, Datasets and Benchmarks Track (2023).
  4. Hendrycks, Burns, Chen & Ball, 'CUAD: An Expert-Annotated NLP Dataset for Legal Contract Review', NeurIPS Datasets and Benchmarks Track (2021).
  5. Liu, Lin, Hewitt, Paranjape, Bevilacqua, Petroni & Liang, 'Lost in the Middle: How Language Models Use Long Contexts', Transactions of the Association for Computational Linguistics 12 (2024), 157–173.
  6. Sergot, Sadri, Kowalski, Kriwaczek, Hammond & Cory, 'The British Nationality Act as a Logic Program', Communications of the ACM 29(5) (1986), 370–386.
  7. Merigoux, Chataing & Protzenko, 'Catala: A Programming Language for the Law', Proceedings of the ACM on Programming Languages 5 (ICFP) (2021).
  8. Surden, 'Computable Contracts', 46 UC Davis Law Review 629 (2012).
  9. Research memoranda come with business and enterprise accounts; see pricing.
  10. See the Privacy Policy and Terms of Service for how documents are handled and the limits of the service.

One engine under every step. See it read yours.

Free to start. WALDHORN.AI is a tool, not a law firm, and does not provide legal advice.