Practice

Method Cannot Be Downloaded

Ten associates, one same clause. After five years, they converge. What makes them converge is written nowhere.

Ten associates. One same clause to draft. The first year, ten different versions. After five years, three or four, nearly identical. This convergence is an observable fact in any firm, and it is the best entry point for understanding what a legal organization’s method truly is.

For what made these ten practitioners converge is in no manual. There is no manual. There is a tacit method: a way of opening a clause, of prioritizing a risk, of posing a reservation, of concluding. It is learned by immersion, never by reading. No one taught it explicitly; each absorbed it by working alongside the others, by being corrected, by seeing what passed and what did not.

This method has a disconcerting feature: everyone recognizes it, no one can write it. A practitioner of the firm immediately identifies a document that “looks like” the house, and another that does not, without being able to formulate the criterion that separates them. The knowledge is there, whole, operating, but it lives in the gesture and not in speech. It is this gap, between a perfectly real and perfectly inexpressible competence, that makes the method so hard to tool.

Everyone recognizes a document that “looks like” the firm. No one can write it.

The test of the partner who tries to write it

Do the test, it is illuminating. Ask a partner to write the firm’s method. They will begin, hesitate, produce three pages of generalities true for every firm and false for their own. It is not that they refuse, nor that they lack lucidity; it is that it is not formulable. It lives in the gesture, not in speech, and what lives in the gesture does not let itself be transcribed without being lost.

And yet, this inexpressible method is of a formidable precision. The same partner spots in three seconds that a note “is off,” without being able to say why in the moment. They feel the gap before being able to name it. The knowledge is there, whole, operating, able to judge with extreme finesse; it simply was never put into words, and never will be, because it is not of the order of the word.

It is nonetheless the firm’s real asset, more than the precedents, more than the templates. It is what holds the coherence of deliverables when people change, what makes a client find the same signature from one matter to the next, what distinguishes one house from another at equal competence. This asset has a fragility: it transmits only by apprenticeship. A junior absorbs it by being corrected, for years, by seniors who take the time. When that time is lacking, transmission seizes up, and the asset erodes without a sound.

A firm’s asset is not what it has written. It is its way of writing.

Why training the model does not suffice

It is often believed that it would suffice to train a model on the firm’s documents to transmit its method. This confuses style and method. Training makes the output resemble past documents, on the surface. It does not capture the decision logic that produced them, because that logic is not in the final text: it is in the sequence of choices that led to it, and that sequence is written nowhere.

An example shows it clearly. Take two warranty clauses drafted by the firm, identical in appearance. One protects a seller, the other a buyer. What distinguishes them is not in the words; it is in the intention that guided each choice of phrasing. A model trained on both will learn to produce clauses that resemble them. It will not learn when to protect the seller rather than the buyer, because that decision is written in neither clause. It is in the drafter’s head, in the matter’s context, in a trade-off that left no textual trace.

And convincing imitation is exactly the trap. A clause that resembles the firm’s passes the first reading; it fails the second, that of the practitioner who looks not at whether it resembles, but at whether it protects the right party, in the right way, for this precise matter. Style is copied, method is not, because method is a decision logic and not an appearance of result. It is a difference of nature, not of degree: a model, even excellent, remains a generator, when method is a decision process.

Fine-tuning imitates the result. It does not inherit the reasoning.

One can push the example further to see where imitation breaks. A model trained on a thousand of the firm’s clauses will reproduce their turns of phrase, their vocabulary, their formatting; it will produce, on demand, a perfectly idiomatic clause. But confront it with a new situation, the one where you must decide, precisely, which party to protect and how far, and it will have no resource, because that decision was not in the thousand clauses: it preceded them. It will imitate the appearance of a decision it never saw being made. This is where imitation, brilliant on familiar ground, collapses on real ground.

Method resides in the sequence

If method is neither in the documents, nor in a model trained on them, where can it reside? In the sequence of steps. A firm’s method is not a set of texts, it is a way of chaining: how reasoning precedes drafting, how review articulates with drafting, how verification is placed, in what order questions are posed and settled. This chaining logic can be encoded, not in a model, but in the orchestration that governs production.

This is what MAX carries. A layer that does not merely produce, but produces in order, according to the method, step after step. Where a generator returns a plausible text in one shot, orchestration unrolls the sequence of choices as a practitioner of the firm would, applying at each step the logic proper to the house. The method is no longer imitated in the appearance of the result; it is active in the way of producing, which is the only way to make it live in a tool.

There is here a survival stake for the firm’s most precious asset. Transmission by apprenticeship was already fragile before AI, for lack of time; an AI that produces outside the method finishes it off, by short-circuiting learning through correction. An AI that produces within the method, on the contrary, can support transmission instead of replacing it, by making the sequence visible, reusable, transmissible. It is not the same AI, and it is not the same future for the firm.

One then measures what “downloading the method” would have contradictory. To download supposes a finished object, transferable at once; the method is a living process, existing only in its execution. You no more download a method than you download a knack: you can encode its sequence, make it play, make it reproducible, but never freeze it in a file you would just load. This is why the right question is not “how to copy the method into the tool,” but “how to make the tool play the method,” and these two questions do not call for the same architecture.

Producing outside the method is plausible text. Producing within the method is firm work.