The Drafting Table

Reasonable Doubts · 18 chapters · ~2 min each

Cast →

Chapter One (~2 minutes)

Terms of Reference

The committee's mandate was one paragraph long, and the argument over what it meant lasted the whole first sitting. Justice (ret.) Eleanor Vance read it aloud, a Model Practice Directive on the Use and Non-Use of AI in Legal Work, and asked the only question that mattered before any drafting could begin: were they writing a rule about a technology, or about care?

Marguerite Loomis wanted the technology. A rule about care was already in the books; what the profession lacked was a bright line for a specific new tool, something a partner could point to and a carrier could underwrite. Peter Okafor wanted the opposite and said so plainly: name the technology and the directive was obsolete before it was printed, because the thing changed every quarter.¹ The duty of care survived model updates; a rule keyed to "generative AI, version whatever" did not.

Vance did not resolve it. She ruled the question into the first bracket instead ("a directive concerning the use of AI, or the exercise of professional judgment in an AI environment") and told the committee they would carry the brackets forward and tally them, because a drafting body shipped its disagreements as text rather than losing them to a vote nobody remembered.

Then the seating fight. The field of exposure² the committee was meant to map ran, Loomis argued, to lawyers: malpractice, unauthorized practice, confidentiality, privilege. So the drafters should be lawyers. Two appointed members were not: Renata Cruz, a senior litigation paralegal, and Dale Whitfield, a self-represented litigant. Nadia Haddad, general counsel, put it as a governance point rather than a snub. Governance built for law firms³ was firm-facing work, and a directive drafted by non-practitioners would not be adopted by practitioners.

Renata answered without heat. She was the person in her firm who actually operated the tools. The uploads, the prompts, the first drafts were hers, not the partners', and a rule addressed to "the lawyer" that never named her would regulate a fiction. Literacy in these systems,⁴ she said, was being drafted as though it lived only in the person who signed the letter, when in her building it lived in the person who ran the software. If the committee wanted the directive obeyed, it needed someone who knew what obedience would cost on a Tuesday.

Dale's claim was simpler. He used these tools heavily, because the alternative to a machine that answered was no answer at all. Whatever the directive said about verifying AI work, he did legal work with AI and no supervisor, no insurer, and no budget. A rule that pretended people like him did not exist would still govern the courtrooms he stood in.

Vance called the seating to a vote. It carried four to three. She recorded the split rather than smoothing it, and noted for the minutes that the narrowness was itself a finding the committee would return to. Okafor, voting yes, observed that advising clients on their own AI compliance⁵ was something lawyers now did for clients who were exactly as unadvised as Dale; excluding him excluded the client they all claimed to serve.

The recital that justified the whole enterprise was agreed without a bracket: the profession had to remain competent in a world the law now governed, and had to anticipate changes in professional services whether or not any given lawyer adopted anything. The proportionality objection, that a small matter might not warrant a heavy rule, Loomis raised and Vance parked; it would attach to particular clauses, not to the committee's existence. They had a mandate, a bracket, and two seats the lawyers had nearly refused. It was enough to begin.

Chapter Two (~2 minutes)

Recitals: Why Any Directive at All

A preamble is where a drafting body says what changed and when, and the committee discovered at its second sitting that it could not. Vance had pinned List A to the wall, two dozen reasons a lawyer must understand AI, and proposed they become the recitals: the whereas-clauses that justified every operative clause to follow. The exercise was meant to take an hour. It took the sitting.

The trouble surfaced when Loomis tried to convert the first reason into a recital and found she could state the conclusion but not the mechanism. "Whereas generative AI can produce confident false law." True, and she could not say why. Okafor could not either. Neither could Haddad, who deployed models at scale. They knew the sanction cases; they did not know what a model was doing when it fabricated.

Varma, the technologist the committee had retained, was asked to supply the missing sentence and declined to supply only one. The definition itself¹ was not a slogan, he said; it was a category. The thing was not a database that occasionally erred but a system that generated the statistically likeliest continuation of a prompt, which is a different object with a different failure mode.² It converted words to numbers, moved them through layered mathematics, and predicted the next token, with enough deliberate randomness that the same question could yield two different answers. "Confident false law" was therefore not a bug in the recital but the recital's whole point.

That reframed the preamble. A directive was needed not because a tool was occasionally wrong but because it was fluently, structurally wrong in a way a busy reader would not catch. Haddad pressed further: the thing did not sit still. The way these systems had evolved³ meant the model her company ran in March was not the model running in September; it had grown from single-turn chat to tool-using systems that reached outside themselves and acted. A directive keyed to today's capability would be describing a fossil.

So the recitals had to justify the directive on the ground of understanding rather than of any particular tool. Literacy in these systems⁴ became the spine of the preamble: not a duty to use, but a duty to understand capabilities and limitations well enough to protect a client who depended on them. It was a duty the abstainer owed as squarely as the enthusiast. Renata sharpened it. The recital should not say lawyers must understand AI; it should say whoever performs the work must, because in her firm the understanding was needed most by the person at the keyboard, who was not a lawyer.

Dale offered the recital the lawyers had left out. The field of exposure⁵ they were mapping (malpractice, confidentiality, privilege) described their own risk. His was different: no insurer to deny a claim and no privilege to waive, only a case he might lose because a machine told him something false with great confidence. If the preamble spoke only of professional liability, it would justify a directive that protected lawyers from consequences and left everyone else facing them.

Vance let both amendments in and noted the meeting's real finding: the committee had agreed to write a rule about a technology none of them, an hour earlier, could describe. Varma was asked to prepare a standing technical annex, the modules the drafters would consult whenever a clause's feasibility was in doubt, and it was understood, without anyone quite saying it, that the annex would be doing a great deal of the committee's thinking. The recitals were agreed. The mechanism, now that they had it, made the rest of the work look harder than it had the day before.

Chapter Three (~2 minutes)

Application: To Whom Is This Addressed?

The draft the subcommittee brought back bound "lawyers," and Renata took it apart before the coffee was cold.

The operative verb throughout was addressed to the admitted practitioner: the lawyer shall verify, the lawyer shall not disclose, the lawyer shall supervise. Renata read three clauses aloud and then described her Tuesday. The document came in; she classified it, uploaded it, prompted the model, produced the first draft. A contract paralegal in another city did the same on overflow. The firm's translation vendor ran client material through its own system. By the time a lawyer touched any of it, the work the directive purported to govern was done, and done by people the directive did not mention.

Loomis said the answer was supervision: the lawyer remained responsible for all of it.¹ Renata agreed that responsibility stayed with the lawyer and said that was the problem. Responsibility was not performance. A rule that assigned the duty to the person who did not do the work, and stayed silent on the person who did, told her nothing about how to behave at nine in the morning. The annex's account of handing a task off² was blunt: agents and assistants failed on vague delegation and succeeded on precise scope, which meant the directive had to name the task and its handoff, not just the person who would answer for it.

The committee saw the fork. It could write a rule about persons, binding named categories of human, or a rule about tasks, binding whoever performed a defined operation at whatever seniority. Haddad, who ran AI across a department³ as her daily reality, said the person-rule was already broken there, because the fluent operator was often the most junior and the partner who signed had never opened the tool. Oversight of the kind the draft assumed⁴ presupposed a supervisor who could recognize the failure being supervised; hers frequently could not.

Okafor worried about the opposite failure. A task-rule reaching every contractor, translator, and agency risked demanding the sort of firm-wide policy⁵ no solo could write. He could not issue attestation forms and tiered approvals to a translation vendor he used twice a year. The rule had to bind the task without requiring a compliance department to discharge it.

What defeated the original draft, in the end, was Renata's single amendment. Strike "the lawyer" as the universal subject, she proposed, and make the directive apply to any person who performs a covered task in legal work, with the responsible lawyer named separately as the person who must ensure it. Two obligations, not one: the performer's duty to do the task to standard, and the lawyer's duty to see that they did. The shadow-AI objection, that staff would quietly wire in unapproved tools the firm could not inventory, became an argument for her version rather than against it, because a rule addressed only to lawyers could not even reach the keystroke where the disclosure happened.

Loomis resisted, then conceded the drafting point while reserving the principle. She still wanted the lawyer at the center; she accepted that the clause had to describe the work as it was done rather than as the rules imagined it. Vance recorded the application clause as narrowed, from persons to tasks-and-responsible-persons, and noted that a paralegal had just rewritten the sentence the lawyers had considered settled, which was, she said dryly, precisely why the seat had been worth the fight in Chapter One. The directive now knew who it was talking to. It had cost the committee its comfortable fiction to find out.

Chapter Four (~2 minutes)

Definitions

Seven words did the load-bearing work of the directive, and the committee spent a sitting discovering that it agreed on none of them: use, verify, reasonable, supervise, rely, disclose, understand. A directive that left them to intuition would be litigated into meaninglessness the first time a lawyer wanted an escape hatch.

"Use" went first, and split immediately. Loomis wanted it broad: any AI touching the matter. Renata wanted it precise, because operating a model on invented facts was not the same act as feeding it a client's file, and a definition that collapsed the two would forbid the safe thing to prevent the dangerous one. The committee provisionally separated operation from disclosure from reliance, and moved on before the separation hardened.

"Verify" and "rely" turned out to be the same fight from two directions, and the annex decided it. The way these systems fail¹ established that the output was not occasionally wrong but fluently wrong, the product of a pattern-matching engine rather than a truth engine, confident precisely where it was least grounded. So "rely" could not mean "take on faith"; it had to mean "adopt after verification proportional to the stakes." Trust calibrated to consequence² gave "reasonable" a spine: trust was not a switch but a graduated judgment, and a reasonable lawyer was one whose checking scaled with what a failure would cost. A framework for checking³ gave "verify" a testable floor: retrieve the source directly, confirm it says what the machine claims. The word now named an act someone could later prove they had performed.

"Disclose" the committee could define against the emerging rules without much argument;⁴ the disagreement there belonged to Chapters Eight and Eleven, and Vance sent it forward.

Then Dale objected to the whole enterprise, and the objection landed. The definitions were being written in the register of people who had never had one applied to them. Literacy as the annex described it⁵ assumed a reader with a firm, a policy, and a colleague to ask. He had none. If "understand" meant what the annex said competent lawyers meant by it, the directive was defining a standard he could not read, applied to work he was nonetheless doing. A definition of "reasonable verification" pegged to a lawyer's resources was, for him, a definition of an obligation he could never satisfy, which made it either a dead letter or a trap.

Vance did not wave it off. She asked Dale to say what he could satisfy, and the answer reshaped "verify": a floor that did not assume a database subscription, so that the test became whether the case exists and says what is claimed, by whatever authoritative means the reader can actually reach. Loomis disliked the softening and Okafor defended it, noting that a solo's version of verification was closer to Dale's than to a large firm's, and a definition only the insured could meet would exclude most of the profession, not just the unrepresented.

"Reasonable" stayed partly bracketed at the end. Reasonable "in the circumstances of the person performing the task" was agreed in principle and left open in degree, because Loomis wanted a floor no circumstances could lower and Dale wanted a ceiling no circumstances could raise. Vance carried the bracket forward. The definitions were operable now, which was all a definitions clause had to be. They had also, she noted, been improved by the one member the drafting style had been quietly written to exclude.

Chapter Five (~2 minutes)

The Competence Clause

For the first time the committee reached genuine consensus, and then split on the very next question, which was, Vance observed, the most honest thing it had done yet.

The consensus was that the duty to understand attached to every lawyer, not only to adopters. Loomis and Okafor, who agreed on almost nothing procedurally, agreed here. The standard of care as it now stood¹ had moved the profession from "optional competence" to "mandatory competence," and nothing in the authorities made the duty conditional on use. A lawyer who advised on a contract a model had drafted, or cross-examined on an exhibit a model had generated, practiced in the world these systems made whether or not she had opened one. The competence clause would bind the abstainer and the enthusiast alike.

Literacy in these systems² gave the clause its content: competence meant understanding capabilities and limitations well enough to know which tasks were safe, which needed verification, and how thoroughly. It did not mean fluency with any particular product. That framing let the non-user in. Understanding, the annex insisted, could be built by someone who would never rely on the thing. The boundaries of what a model can be asked to do³ described a constrained instrument governed by hidden instructions and a chain of command, whose limits could be mapped by anyone willing to probe them, which meant a lawyer could know the shape of the cage without living in it.

The split came over demonstration. Everyone agreed the duty was continuing rather than discharged by a single seminar,⁴ roughly the fifteen-to-twenty annual hours the annex costed out, renewed as the tools changed. The fight was over what evidence of it the directive should require.

Loomis wanted an attestation: a signed annual statement, auditable by a carrier, that the lawyer had completed defined training. Bright line, safe harbor, insurable. Haddad wanted the same for a different reason. Bringing a whole department along⁵ told her that competence in an organization was a human problem, not a technical one, and only a documented cycle with named accountability actually moved a department. Renata objected on behalf of the people the attestation would paper over: a signature proved attendance, not understanding, and in her firm the person who would sign was rarely the person who operated the tool. Requiring the wrong person to attest would produce a file that looked like compliance and described nothing.

Dale objected from the other end. An attestation regime assumed an employer, a CLE budget, and a compliance officer. Hold him to it and the clause became a credential he could not obtain, gatekeeping competence behind resources rather than testing it. He would rather be examined than certified: asked whether he understood the thing, not asked to prove he had paid to be told about it.

Vance found the consensus she could and bracketed the rest. The duty: agreed, universal, continuing. The content: agreed, understanding over fluency. The demonstration: bracketed (by attestation of completed training, by demonstrated understanding, or by both) and carried forward with a note that the choice was really a choice about who could afford to comply. Okafor said the quiet part: a competence clause that could only be satisfied with a training budget was a competence clause about money. The clause shipped with its hardest word in brackets, which was, Vance said, better than shipping a false agreement about it.

Chapter Six (~2 minutes)

The Non-Use Clause

The clause nobody had planned to draft became, by the sixth sitting, the one that told the committee what kind of instrument it was writing.

Okafor proposed it. The directive so far assumed adoption and regulated it; nothing protected the lawyer who looked at a matter and declined. In his practice non-use was often the competent choice, and he wanted it recorded as a professional decision rather than treated, by omission, as a confession of falling behind. Vance took List B off the wall, the defensible grounds for declining, and told the committee this was where it earned its keep, because the two lists were not opposing teams: List A said why the profession had to understand these systems, and List B said why understanding sometimes counseled against using a particular one. A clause could serve both at once.

The drafting was disciplined by insisting every ground attach to something specific. Classifying risk by task and consequence¹ gave the structure: not "AI is risky" but this task, at this consequence level, on this tool, is one where the risk cannot be adequately controlled. The way these systems fail² justified declining to rely on an unverified generation in a filing that left the office. The limits of what one can be asked to do³ justified declining a tool whose instruction layers and limits the lawyer could not map. The apparatus a firm needs to hold the risk down⁴ justified declining where the firm lacked the four pillars (policy, training, quality control, technical controls) to use the thing safely at all. Each was a reason to decline a particular use, never a verdict on the technology.

Then Renata nearly sank the clause, and improved it instead. As drafted, non-use was a shield: state that you declined and you were protected. She said that in her firm the sentence "we don't use AI for this" was frequently false. Someone did, quietly, on an unapproved tool, and the shield would protect the partner who had waved the flag while the actual disclosure happened three doors down. A bare right to decline, unaccompanied by any duty, protected the appearance of non-use, not the fact of it.

Trust treated as a calibration⁵ supplied the fix. If declining was a calibrated judgment rather than a mood, the decliner should be able to show the calibration. The committee added a companion obligation: a lawyer invoking the non-use protection must be able to state the reason for declining and the alternative relied on instead. Not a filing, not a form. An answerable account. The shield attached only to a decision someone had actually made.

Dale, who used AI precisely because he had no alternative, tested the other edge. His non-use was never a choice; his use was compelled by the absence of counsel. He asked whether the clause protected him at all, and the honest answer was that it protected a decision to decline, which presupposed the resources to do the work another way. Vance recorded that the non-use protection was, for the unrepresented, mostly theoretical, a finding she flagged forward to Chapter Fifteen rather than solving here.

Loomis, who had wanted bright lines, found she liked this one: a documented reason and a named alternative was exactly the auditable artifact a carrier could respect. The clause survived intact and stronger, a right to decline married to a duty to explain the declining. Non-use, the committee had decided, was not the absence of a professional act. It was one, and the directive would treat it as such.

Chapter Seven (~2 minutes)

Confidentiality and Information Boundaries

This was the clause the annex all but wrote, and the one where Renata's ordinary Tuesday defeated two drafts before lunch.

The first draft was a prohibition: no client information into any AI system. Clean, and useless. Haddad pointed out it would forbid the enterprise tools her company had licensed precisely to keep data in-bounds, while doing nothing about the intern's phone. Governing where client data actually goes¹ made the real question not whether information entered a system but what the system did with it (retention, sub-processors, training use, employee access), and those varied wildly between a consumer tier and a contracted one. A flat ban regulated the word "AI" instead of the flow of the data.

The second draft fixed the tier and broke elsewhere. It permitted "approved enterprise tools." Renata read it and described her Tuesday. The approved research tool called out to three services she had never been shown; the drafting assistant synced to the document system through a connector nobody in the room could name; the translation vendor's tool was its own black box. How a model is reached and where it runs² was the distinction the draft had missed. The same model reached through a public portal, a no-retention enterprise account, an on-premise install, or a sovereign cloud was four different confidentiality postures, and "approved tool" named none of them. Whether a privileged memo left the building depended on the channel, not the brand.

Varma made it worse before he made it better. The connections a tool makes outward³ meant the tool was rarely the whole story: it reached through integrations, each with its own credentials and terms, each another party with a hand near the data. The standard wiring between them⁴ let a connector be built once and reused anywhere, which was efficient and meant the client's file could travel further, through more hands, than any engagement letter described. The confidentiality boundary was not the tool's edge. It was the edge of a stack of agreements incorporated by reference.

So the clause had to be written to the data flow. Evaluating a tool before trusting it⁵ supplied the pre-conditions: encryption, access controls, audit logging, data residency, and a stated policy on retention and training, all of them things a tool must satisfy before client material touched it. Evaluating the whole stack rather than one product⁶ added the questions buyers forgot: where does the data go, who touches it, and what does it cost to leave. The third draft forbade nothing by name and required that, for any tool touching client information, the responsible lawyer be able to answer those questions about retention, sub-processors, training use, residency, and the connector chain, or not use it for that matter.

Okafor raised the cost. He could not audit a service chain; he barely had time to read the terms. The clause seemed to demand a security review no solo could perform. Renata, who had just defeated two drafts, offered the practical floor: the questions did not require an auditor, only an answer, and a vendor who would not answer them in writing had answered the most important one. Dale noted that his confidentiality risk was the same as any lawyer's and his ability to investigate it far smaller, a point Vance again flagged toward Fifteen.

The clause carried, grounded entirely in the annex. Confidentiality, the committee had learned, was not lost by "using AI." It was lost or kept at a specific point of entry, through a specific channel, into a specific chain, and the directive now made the lawyer say which.

Chapter Eight (~2 minutes)

Client Consent and Disclosure to Clients

The question looked procedural and turned out to be about who the directive was for: what must a client be told about AI in their matter, in what words, and could an engagement letter ever carry it?

Haddad opened for less. As general counsel she received these disclosures as well as sent them, and a paragraph in every retainer saying "we may use AI tools" told her client nothing she could act on; it satisfied a rule while informing no one.¹ What she wanted was assurance, not announcement: fewer words to the client and better controls behind them. A sophisticated client, she argued, was protected by the firm's verification and confidentiality regime, not by a boilerplate sentence she would never read.

Loomis was sympathetic and cautious. Transparency as the thing that earns trust² cut the other way for her: the emerging rules, and the malpractice exposure, rewarded disclosure at the most protective standard across jurisdictions, and a carrier looked more kindly on the firm that over-disclosed than the one that reasoned its way to silence. She did not want the directive blessing less disclosure because a given client happened to be sophisticated.

Then Dale spoke as the client the clause forgot. Haddad's "less disclosure, better assurance" assumed a client who could evaluate assurance, one who had a general counsel of her own to read the controls. He could not. For him, the only protection was being told, plainly, that a machine had been used and what that meant for his matter, because he had no capacity to audit the firm's verification and no leverage to demand it. A rule calibrated to the sophisticated client would strip the disclosure precisely where it was the only safeguard left.

Governing where client data goes³ anchored the substance both wanted addressed: informed consent to AI use was meaningless unless the client was told what information would be exposed, where it would travel, and what residual risk remained. The honest version was hard to write, because the answer was often "further than we can fully trace." Boilerplate could not carry it. Haddad conceded the point about traceability and held her ground on volume.

The compromise was Dale's, and it won. The directive would scale the disclosure to the client, not the convenience of the firm: a sophisticated client who so agreed could receive a defined, standing disclosure and rely on assurance; but a baseline of plain, matter-specific disclosure was owed to any client who had not knowingly waived it, and could not be buried in a retainer as a condition of engagement. Advising a client on AI compliance⁴ framed it as the counsel lawyers already gave others about their AI governance, now owed to their own clients: material use, material risk, in words the recipient could actually use.

Haddad accepted it because it gave her the efficient path with her own clients while she kept the protective floor for everyone else's. Loomis accepted it because the floor was the most-protective standard she had wanted. Renata noted the operational catch: the person who knew whether a matter's AI use was "material" was usually her, not the partner drafting the letter, so the clause quietly created a duty to ask the operator before certifying the disclosure. Vance recorded the consent clause as agreed and narrowed, a two-tier disclosure with a floor that could not be waived by boilerplate, and observed that the institutional client and the unrepresented litigant had, between them, drafted a better clause than either would have alone.

Chapter Nine (~2 minutes)

Supervision and Delegation

The supervision clause read well and assumed a person who did not exist in most of the firms it would govern: a supervisor who understood what he was supervising.

The draft restated the settled rule: delegation does not transfer responsibility; the lawyer supervises and answers for the work.¹ Nobody disputed it. Renata disputed what it silently assumed. Supervision, on the annex, was not a signature; it was an activity. The oversight it described² ran to checkpoints, output validation, and catching drift before it cascaded, and every part of it required the supervisor to recognize the failure in front of him. In her firm the partner who signed frequently could not. The most fluent operator was the most junior person, and the most consequential review was performed by whoever understood the tool least.

Loomis said the answer was that the partner must therefore become competent, which was Chapter Five's duty, now with teeth. Renata agreed and said it did not describe Tuesday. The model for working alongside an agent³ set out real oversight in pre-execution review, output verification and escalation triggers, and warned against its counterfeit: the rubber stamp, approval without actual review, supervision as theater. A clause that required supervision without requiring the supervisor to understand the work would license exactly the theater the annex named.

Delegating a task precisely⁴ gave her the constructive half. The reason vague delegation failed was that the delegator had not specified the task, the constraints, the exit conditions; good delegation was precise about all three. Applied to the directive, that meant the supervising lawyer could not discharge the duty by generic oversight. He had to define the covered task, its verification standard, and the point at which the operator must escalate to him rather than proceed. Supervision became a set of specified handoffs, not a posture.

Haddad brought what she knew about spreading this across a department⁵ to bear and confirmed the diagnosis at scale: the fluent minority became a bottleneck and the untrained majority either waited or improvised, and the only thing that held quality was a defined standard the supervisor actually enforced. She wanted the clause to require that whoever supervised a covered task be competent in that task, not in AI generally but in the specific thing being checked.

That was the amendment that defeated the draft. Renata proposed, and the committee accepted, that the supervision duty be tied to task-competence: a lawyer may not supervise a covered task he is not himself competent to verify, and must either acquire that competence, assign a supervisor who has it, or decline the delegation. Loomis flinched. It meant a senior partner could be, on a given tool, less qualified to supervise than his paralegal, and the clause now said so out loud. She accepted it because the alternative was a rule that blessed the rubber stamp.

Okafor added the resource caveat: a solo often was the whole chain, supervisor and operator both, and the clause had to let one competent person occupy both roles without pretending they were two. Vance recorded the supervision clause as narrowed, so that supervision requires task-competence and unsupervisable delegation must be refused, and noted, for the second time, that Renata had rewritten a clause the lawyers thought finished. The directive no longer assumed the supervisor understood the work. It required it, and admitted what that meant about who, in a modern firm, was actually qualified to sign.

Chapter Ten (~2 minutes)

Verification and Independent Judgment

Everyone had assumed the verification clause would be quick. It consumed two sittings, because "verify" would not hold still until the committee decided what it was measured against.

The first draft required that AI output be "verified before use." Vasquez, the adjudicator, said it was unenforceable from the bench: he could not tell, looking at a filing, whether "verified" meant a glance or a re-research, and a standard he could not apply was a standard that would be argued about rather than met. The professional standard for checking authority¹ set the ceiling for the work that mattered: every citation checked for existence, holding, currency, and relevance, each check documented. The sanction cases behind it were not hypothetical. But applying that ceiling to everything was impossible, and Loomis knew it.

She wanted a bright line: a fixed rule keyed to the tool. AI-drafted filings get full verification; everything else, a spot-check. Clean, auditable, and wrong, said Varma, because the tool was a poor proxy for the risk. Quality control across the whole of the work² tied check-depth to the cost of error rather than the instrument, since a marketing email and a motion could come from the same model and warranted opposite scrutiny. Criteria for what counts as good enough³ made the same point from the quality side: "good enough" was defined by use, not by whether a machine was involved.

So the committee tried to tier by consequence, and Loomis fought it for two sittings. A consequence-based standard, she argued, gave every lawyer a discretion to characterize his own risk downward, and a bright line at least could not be argued away. Trust treated as a calibration⁴ was the annex's answer to her, and it was honest about the trade: trust was not a switch but a graduated judgment, and the alternative to graduation was not safety but a rule that over-verified the trivial and under-verified the novel, because the tool told you nothing about the stakes. A framework for checking⁵ gave the graduation a testable spine: retrieve the source, confirm it says what is claimed, check currency, look for contrary authority, record what you checked. "Proportional" therefore did not mean "discretionary" but "scaled by a stated method to a stated consequence."

Vasquez broke the deadlock from the docket's side. What he needed from the bench was not a fixed depth but a legible record: tell him the consequence level the lawyer assigned and the checks that level required, and he could test whether the assignment was reasonable. A bright line gave him a number that lied; a tiered standard with a documented rationale gave him something he could actually review. That was the argument that moved Okafor, who had feared the tier would demand a quality department, and now saw it demanded a defensible sentence instead.

The clause tiered verification by consequence, with the legal ceiling fixed for anything filed, relied on, or delivered outside the office, and a documented rationale required for the level assigned. Loomis lost the bright line and said so for the record, believing still that a rule you could argue about was a rule that would be, and signed the tier because the annex had convinced her the bright line protected the appearance of rigor rather than the fact of it. Independent judgment, the clause said in its last line, could be assisted but not delegated: the verification was the lawyer's act, whoever ran the tool. Vance recorded it agreed over a noted dissent, and observed that the committee had spent two sittings turning one verb into something a tribunal could enforce.

Chapter Eleven (~2 minutes)

Certification and Candour to Tribunals

The draft certification was ambitious, and the adjudicator at the table shrank it to something a court could actually use.

The proposal was a filed certificate: the lawyer certifies whether AI was involved in preparing the document, identifies the tool, and certifies that the output was verified.¹ Loomis liked its completeness. The decided cases and the sanctions in them² stood behind it. The wave of sanctions since Mata v. Avianca had been for unverified fabricated citations, and a certification regime, she argued, would have caught them. Transparency as the thing that earns trust³ pointed the same way: disclose at the most protective standard, and certification was the most protective form of disclosure.

Vasquez let them finish and then reported what certifications actually did to his docket. He saw hundreds of filings. A universal AI-involvement certificate would appear on every one of them within a month, an unread line of boilerplate identical across the bar, and it would tell him nothing about the only thing he cared about: whether the citations were real. Worse, it would give him a false comfort and a false target, because lawyers would litigate whether a spell-checker was "AI" instead of whether the brief was true. A certificate everyone signed and no one read was not candor; it was a new place to hide.

What he wanted was narrower and harder. Not "AI was used" but "the authorities in this filing have been verified and are accurate." The professional standard for checking authority⁴ already required the underlying work; the certificate should certify the thing that mattered, the verification itself, and make the lawyer personally answerable for it, rather than certify the tool, which deceived no one who mattered and misdirected everyone else. The deceived party in a fabricated-citation case was the court, and the court was not deceived by the presence of AI; it was deceived by an unchecked cite.

That reframed who was deceived by the difference. Certifying involvement protected the lawyer who could say "I disclosed the tool" while filing false law. Certifying verification protected the tribunal. Haddad, who disliked disclosure for its own sake, preferred the narrow version for the same reason Vasquez did: it demanded assurance rather than announcement. Renata flagged the operational truth. The person who could truthfully certify verification was whoever performed it, so the certificate had to run to the responsible lawyer who had actually confirmed the work, not the one whose name headed the caption by default.

The committee kept a residual disclosure hook for the cases that needed it. Where a court's own standing order or a contract required identification of AI use, the directive deferred to it; the local rules were multiplying and inconsistent, and the directive would not override a tribunal's specific demand. But the directive's own default shrank to the verification certificate: personal, answerable, about truth rather than tooling.

Whether the certificate must also name the tool where no local rule required it stayed bracketed, Loomis still wanting it in and Vasquez still thinking it noise, and Vance carried the optional rider forward, "and identifying any AI tool used," rather than resolve it. The certification clause left the sitting smaller than it arrived and, everyone agreed, better aimed: a lawyer would certify not that a machine had helped but that the authorities were real and checked, which was the promise the court had always actually needed and the one the sanctions had always actually been about.

Chapter Twelve (~2 minutes)

Vendor Diligence and Procurement

The diligence clause was drafted by people who had procurement departments, and the solo at the table pointed out that it could only be obeyed by people who had procurement departments.

The obligations were sound and, as written, heavy. Before adopting a tool, the responsible lawyer must interrogate the vendor's claims, map the service chain, and refuse what could not be explained. Evaluating a tool before trusting it¹ supplied the checklist: security, data residency, retention, certifications, performance on real tasks. Evaluating the whole stack rather than one product² added lock-in and exit cost, the questions buyers regretted skipping. Haddad, who ran real procurement, thought it about right.

Varma pushed the substance harder, because most of the claims a vendor made were meant not to be understood. The foundation models and the companies behind them³ meant the first question was which model sat underneath and who else was therefore in the chain. A "legal AI platform" was usually someone else's foundation model with a wrapper, and its knowledge cutoff, context window and failure modes were inherited from a provider the brochure never named. How a model is reached and where it runs⁴ determined where the data lived, whether on the web, through an API, on-premise or in a sovereign cloud, which mattered more than any warranty. The surrounding ecosystem and its integrations⁵ made the product a lattice of connectors, each a maturity risk and each another party, so "our AI understands your documents" collapsed on the question of whether it was trained on them or merely retrieving from them at query time. A diligent buyer had to know which layer could fail and how he would see it.

Then Okafor read the clause as it would land on him. Every question in it was right, and he could satisfy none of them at a vendor table. He had no security team to run a questionnaire, no leverage to demand a data-processing agreement, no time to trace a service chain across four sub-processors. The clause as drafted did not regulate AI adoption; it regulated firm size, permitting adoption only by those wealthy enough to interrogate a vendor and forbidding it, in practice, to the solo who most needed the leverage of a written answer.

That was the objection that reshaped the clause. The committee had confused the diligence with the department. Varma observed that the questions themselves did not require an auditor; they required answers, in writing, from the vendor. A solo could not run a certification review, but he could refuse to adopt a tool whose vendor would not state its model, its data residency, its retention, and its exit terms on paper. The obligation was reframed from "conduct diligence" to "obtain and retain the vendor's answers to defined questions, and decline what the vendor will not answer," a duty a solo could discharge with an email and a filing cabinet, and one that shifted the cost onto the vendor who wanted the sale.

Loomis worried the lighter version let small firms adopt on thinner review than large ones. Okafor agreed and called it honest: the alternative was a rule the large firms met and the small firms quietly ignored, which protected no client. Renata added the operational hook. The person who would actually notice a connector reaching somewhere new was the operator, so the diligence had to be refreshed by whoever ran the tool, not filed once at purchase and forgotten.

Vance recorded the procurement clause as narrowed, diligence becoming documented vendor answers scaled to what any practitioner could actually obtain, and noted that the solo had, like the paralegal and the litigant before him, turned a clause written for the well-resourced into one the rest of the profession could keep.

Chapter Thirteen (~2 minutes)

Records, Retention and Discoverability

Halfway through the retention clause the committee realized it was drafting a discovery problem for the entire profession, and had to decide whether to admit it on the record.

The clause began as housekeeping: what AI-related material must be kept, what destroyed, and for how long. It stopped being housekeeping when Varma described what the material actually was. The question of what persists between sessions¹ established that a model kept nothing on its own. Close the tab and the matter was gone from it, so the record did not live in the machine; it lived in whatever the firm captured around it. And the firm captured a great deal: prompts, outputs, the continuity documents lawyers uploaded to carry context forward, the reusable² libraries of saved prompts and precedent that made the work efficient. Every one of those was a document.

Haddad saw the implication first, because she had lived it. Firm-wide policy³ on the annex treated prompts and outputs as governable records precisely because they were discoverable ones: informal analysis, abandoned drafts, an operator's aside to the model that read, in hindsight, like an admission. The firm that logged everything for the sake of managing its risk,⁴ because the quality-control pillar wanted an audit trail, was simultaneously building a warehouse of material an opponent could one day demand. The two duties pulled against each other: log for verification, and you create the discoverable record; log nothing, and you cannot prove you verified. And the warehouse was full of confidences. Governing where client data goes⁵ meant those retained prompts and outputs carried the same client information Chapter Seven had guarded at entry, so a schedule that held them indefinitely was itself a confidentiality exposure, not a neutral archive.

The professional standard for checking authority⁶ sharpened the bind, because its audit trail of tool, prompt, checks performed and corrections made was exactly the documentation a court would want if verification were questioned, and exactly what an opponent would want if it were not flattering. The committee could not require the trail for candor in Chapter Eleven and pretend here that it would not be reachable by a litigation hold.

Loomis wanted the clause to solve it cleanly: define these materials as privileged work product and be done. Vasquez, from the bench, said she could not. Privilege was contested for exactly this material, and a directive could not manufacture a protection the courts were actively declining to grant. The most it could do was tell lawyers the truth about their exposure and require them to manage it deliberately.

So the real question surfaced: should the directive say, in writing, that AI interactions generate discoverable records ordinary practice was not treating as such? Haddad hesitated. A professional standard that announced a new discovery frontier might be read as an invitation to go mining. Dale argued the opposite and carried it. Silence protected the informed, who already knew; disclosure protected people like him, who generated prompts and outputs with no idea they were creating a record an opponent could subpoena. A directive that stayed quiet to spare the profession embarrassment would be keeping a hazard secret from the people least able to discover it themselves.

The committee decided to say so. The clause required a retention schedule that classified AI prompts, outputs, logs, and drafts as records subject to the same retention, litigation-hold, and destruction rules as any other, no better and no worse, set in advance rather than improvised under a hold. It added a plain-language acknowledgment, over Haddad's noted reservation, that such material was potentially discoverable. Vance recorded the retention clause as agreed, with the discoverability acknowledgment bracketed only as to wording, and noted that the committee had done the harder of the two available things: it had written down the problem it discovered it was creating, rather than shipping a clause that pretended not to see it.

Chapter Fourteen (~2 minutes)

Fees, Billing and the Allocation of Cost

The billing clause asked a question the committee could not answer honestly and comfortably at the same time: did the directive make representation cheaper or more expensive?

The easy part was the abuse. Automating a legal workflow¹ meant a task that took an associate three hours might take a supervised pipeline three minutes, counting intake, extraction and a first draft, and billing the client three hours for three minutes of machine time was the plain wrong the annex and the rules both condemned. Loomis wanted a clean prohibition: no charging for time not actually spent, and transparency about efficiency gains. Agreed without much fight.

The hard part was the cost the directive itself created. Every clause the committee had passed added work: verification proportional to consequence, vendor answers obtained and retained, disclosures scaled to the client, retention schedules maintained. The patterns automation follows² cut both ways, saving time on the drafting and adding it back at the human review gates the same annex insisted on, so the net was not obviously a saving. Someone paid for the verification. The question was who.

Haddad said the market would sort it. Building a position clients could see³ treated demonstrable, well-governed AI competence as a differentiator they would pay for, and a firm that could show faster, cheaper, verified work would win. Okafor said that described her clients, not his. His clients chose him because he was affordable, and a directive that loaded verification and diligence costs onto every matter would either come out of his margin or go onto their bill, and he could not absorb it indefinitely. Planning for the long term⁴ on the annex was candid that the compliance apparatus of documentation, training and controls was an ongoing cost rather than a one-time setup, and it fell hardest on the smallest practices.

Then Dale asked the question directly, as the person who would pay a bill or go without one. Would the directive make a lawyer cheaper to hire? He was not asking rhetorically; he was deciding whether the profession's new rulebook made counsel more or less reachable for someone like him. The committee could not give him a confident yes. Automation lowered the cost of production; the directive's own safeguards raised the cost of doing it responsibly; and where those two lines crossed depended on the firm, the matter, and the tool.

What the committee could do was refuse to let the directive be used to inflate a bill. The clause forbade charging for time not spent; required that technology cost passed to a client be disclosed and reasonable, not marked up as if it were labor; and, in Renata's addition, required that the efficiency of AI-assisted work not be hidden behind a fiction of hours, because the operator knew exactly how long the machine had taken even when the invoice pretended otherwise. It stopped short of promising savings, because the committee would not write a promise it could not keep.

Loomis wanted a safe harbor for value billing, charging for the result rather than the time, and the committee half-gave it, permitting a reasonable fee reflecting value and responsibility while forbidding a fee that merely disguised unspent hours. The line between the two stayed soft, and Vance let it, noting that fee reasonableness was old law the directive need not rewrite.

She recorded the billing clause as agreed on the prohibition and honest on the uncertainty. Dale's question went into the minutes unanswered on purpose: the directive would make responsible AI use auditable, but whether it made representation cheaper for the people who most needed it cheaper was, Vance wrote, a question the committee could pose and not resolve. That was itself worth putting on the record.

Chapter Fifteen (~2 minutes)

Unrepresented Parties and Access

The directive regulated lawyers, and Dale rose to say that most of the AI-assisted legal work now moving through the system was not being done by lawyers at all.

He put it plainly. In the housing list, the small-claims list, the family list, the person on the other side increasingly arrived with a bundle a machine had helped assemble, because the alternative was arriving with nothing. A first interaction that builds confidence¹ described exactly what he had done the first time: asked a plain question, got something usable back, crossed a threshold he had been denied for years. The power to work through a document² described the rest, which was uploading his own file, asking it targeted questions, and building a private understanding of his own matter that no lawyer had ever been paid to give him. The tools the directive was busy fencing were, for him, the first legal help he had ever been able to afford.

The committee's instinct was that this was outside its remit. A professional directive bound the profession; it could not regulate a litigant, and Loomis said so, since the document had no authority over Dale and should not pretend to. Haddad agreed on jurisdiction and worried about scope creep. Vance let the objection stand and asked the harder question: if the directive said nothing about the unrepresented, what did its silence do?

Dale answered that. The applications in legal practice³ on the annex were written for lawyers, covering contract review, research and filing preparation, all under a duty of verification, and every safeguard in them presumed a professional to perform it. Applied to him, the safeguards became obstacles: a verification standard he could not meet, a disclosure regime with no one to disclose to, a competence duty with no CLE to satisfy it. The directive's protections were, for the unrepresented, a wall. And the flag raised in Chapters Six, Seven, and Eight, that verification was a resource clause in disguise, landed here with full weight. Every clause that said "verify" said, to Dale, "have the resources to verify, or fall short of a standard no one will enforce against you but everyone will judge you by."

So he proposed the amendment the lawyers had not written. The directive could not bind him, but it could bind them in their dealings with him. Where a lawyer faced an unrepresented party using AI, the lawyer should owe defined duties: not to exploit an AI-generated error the litigant could not catch, not to bury a self-represented opponent under a verification standard the lawyer knew they could not meet, and to bring a known AI-generated fabrication to the tribunal's attention rather than win quietly on it. Building a position clients could see⁴ gave the professional case for it: a profession that used its AI advantage to crush the unrepresented would forfeit exactly the public confidence the whole directive was meant to preserve.

Loomis resisted hard. It imported a duty toward opponents the adversarial system did not recognize, and she would not invent one in a practice directive. Okafor was torn, sympathetic to Dale and wary of a clause that might expose a solo who simply beat a pro se litigant fairly. Vasquez, from the bench, said the courts were already drowning in exactly this problem and would welcome a professional norm to point to.

The vote was close: four to three, the same margin that had seated Dale in Chapter One, and Vance noted the symmetry aloud. The clause carried in a narrowed form, with no affirmative duty to assist, but a duty not to exploit a known AI error against an unrepresented party and a duty of candor where an opponent's AI fabrication reached the court. Dale had not been protected by the directive. He had, Vance recorded, rewritten it, turning a rule that forgot him into one that at least required the profession not to use its advantage against him. It was the second time the litigant had moved a clause the lawyers thought settled, and the last before the brackets came due.

Chapter Sixteen (~2 minutes)

Incidents, Insurance and Reporting

The incident clause was going nowhere until Loomis read a letter into the record, and the letter did more to move the committee than any argument had.

The draft asked the ordinary questions: what counts as an incident, who must be told, and when. Recovering from an error and building for resilience¹ gave the taxonomy. AI failures fell into patterns, whether the confident false fact, the incomplete execution or the reasoning that broke, and the annex's counsel was to diagnose before redoing and to design workflows that absorbed an error rather than passing it downstream. Classifying risk by task and consequence² tiered the response: a high-consequence failure demanded disclosure, audit, and remediation; a trivial one, a note. The committee could draft that.

What it could not settle was the reporting duty, to whom it ran and how it interacted with the standard the committee was writing, until Loomis produced a renewal letter from a professional liability carrier. She read the exclusions aloud. The gaps opening in coverage³ were no longer theoretical: the letter excluded claims arising from undisclosed AI use, from reliance on unverified output, and from confidentiality breaches through cloud AI tools, and it conditioned coverage on documented verification and training. The room went quiet in a way no clause had produced, because the letter was not the committee's aspiration; it was a bill someone would actually pay.

Professional liability cover as it now stood⁴ explained why it bit. Coverage for AI error was not automatic; it was conditional, and the conditions were exactly the practices the directive had spent fifteen chapters drafting: disclose, verify, document, train. A lawyer who breached the directive was not merely exposed to discipline; he might be uninsured for the very claim the breach produced. What it takes to keep that cover in place⁵ made the converse the selling point: coverage maintenance was an action problem, not a knowledge one, and the documentation the directive required, the verification records and the vendor answers and the incident log, was precisely what a carrier wanted to see.

That reframed the incident clause from a burden into an alignment. Haddad, who had resisted disclosure for its own sake, now wanted the reporting duty tight, because an undisclosed incident was the thing that voided coverage. Okafor, who had fought every added cost, went quietest of all: the letter meant his choice was not between cost and no cost, but between the directive's cost and being personally uninsured for a six-figure claim. Nobody had persuaded him. The carrier had.

The clause required an internal incident record for any AI error caught in verification, the near-miss the annex valued, and external reporting scaled to consequence: to the client where the matter was affected, to the tribunal where a filing was, and to the insurer where the policy required it. Dale, characteristically, marked the asymmetry. The reputational downside of an incident could exceed its legal harm, and he had no carrier to report to and no reputation a firm would rebuild for him. An incident a lawyer would insure against, he would simply suffer. Vance flagged it forward to the adoption debate.

Loomis got the reporting duty she wanted, for a reason she had not expected to give: not candor as virtue but coverage as survival. Vance recorded the incident clause as agreed and noted the sitting's real lesson. The committee had argued about principle for fifteen chapters and been moved, in the end, by an actuary. It was not the noblest way to reach a good clause. It was, she wrote, one of the more durable, because a duty an insurer enforced would outlast a duty only a regulator announced.

Chapter Seventeen (~2 minutes)

Review, Sunset and Amendment

The committee's most honest clause was its admission that the directive would be wrong within eighteen months, and its only real agreement of the sitting was about how it would learn.

The problem was structural and everyone could see it. The way these systems had evolved¹ meant the thing being regulated did not sit still. Single-turn chat had become tool-using agents in the span of the committee's own work, so a directive fixed to today's capability described a receding object. The standard of care as it now stood² compounded it: the standard was moving under them, from optional to mandatory competence in a couple of years, and a rule that did not move with it would ossify into a floor lower than the law's.

Loomis wanted stability; a standard that changed constantly was one no lawyer could rely on and no carrier could underwrite. Okafor wanted responsiveness; a rule that could not absorb next year's tool would be quietly ignored the moment it lagged. They could not agree on how much substance to fix, and Vance made the disagreement itself into the clause.

If the committee could not agree on the content, it could agree on the cadence. Planning for the long term³ on the annex was built for exactly this, with practices designed to stay valid as the ground shifted, resting on transparency, proportionality, and documentation rather than on any specific tool, and it modeled a review cycle that adapted without starting over. The committee adopted a mandatory review at a fixed interval, a standing duty to reassess the bracketed items, and a sunset on any clause keyed to a named capability unless affirmatively renewed.

Bringing a whole team along⁴ supplied the reason the review had to be more than a calendar entry. The annex's finding, that adoption failed on human factors rather than technology and that competence lapsed without a deliberate cycle, applied to the directive itself. A rule reviewed by no one was a rule that decayed; the clause therefore named who would review it, on what evidence, and how amendments would propagate, rather than leaving "periodic review" as the pious nothing it usually was.

Haddad pushed for the review to be evidence-driven: not a committee reconvening to share impressions, but a body that ingested the incident records, the coverage changes, the new sanctions, and the shifting bar guidance, and amended against data. Vasquez agreed from the docket; the courts would generate the evidence fastest, and the review should have a channel to receive it. Renata added the operational amendment that had become her signature. The people who would first see a clause fail were the operators, and the review had to hear from them, not only from the partners who signed.

Dale asked whether the review would reconsider the parts that failed him, the resource clauses and the access question, or only the parts that failed the firms. Vance made him a small promise she could keep: the bracketed items and the minuted unresolved questions, his among them, would be the review's standing agenda, carried forward by rule rather than by anyone's memory.

The clause passed easily, which Vance distrusted and named. Unanimity here was not agreement about the directive; it was agreement to disagree on a schedule, and that was the honest outcome, a body that could not settle the substance settling instead on how it would keep learning. She recorded the review clause as agreed and observed that a directive which admitted its own coming obsolescence was more trustworthy than one that pretended to permanence. One sitting remained, and the brackets were now due.

Chapter Eighteen (~2 minutes)

Adoption: The Brackets and the Minority Note

The last sitting existed to resolve the brackets, and it resolved them the way honest bodies do: some closed, one left permanently open, and the disagreement published rather than hidden.

Vance took the brackets in order. The first, from Chapter One, whether this was a rule about the technology or about care, she closed by making it care: the directive would govern the exercise of professional judgment in an AI environment, with the technology named only where a clause required it. Governance built for law firms¹ had proven the point across eighteen chapters, since every clause that survived was really about judgment, confidentiality, and supervision rather than about any particular model, and even Loomis, who had wanted the technology named, conceded the drafting had vindicated the other choice.

The competence-demonstration bracket from Chapter Five was harder and split the room for the last time. Loomis and Haddad wanted attestation: auditable, insurable, the thing a carrier and a firm-wide² regime could enforce. Renata and Dale wanted demonstrated understanding, because attestation certified a budget and not a mind, and excluded the operator and the unrepresented alike. The committee could not bridge it, and Vance refused to fake a bridge. She closed the bracket for the majority in favor of demonstrated understanding, with attestation as one accepted means but not the only one, and let the dissent stand.

That produced the minority note. Loomis and Haddad declined to sign the majority text and filed a note the committee voted to publish unedited. Their reasoning was not obstruction, and Vance made sure the minutes said so: they believed a standard resting on "demonstrated understanding" was too soft to enforce, too easy to argue about, and that managing the risk in order to be protected against it,³ the four pillars operationalized and documented, required the bright, attestable lines the majority had repeatedly declined. They thought the directive, for all its care, had written a rule that would be litigated rather than obeyed.

The striking thing, which Vance recorded as the committee's real achievement, was that all four principals could state the others' reasoning without caricature. Renata could articulate why Loomis wanted attestation, naming the carrier and the audit and the survival of the firm, and Loomis could articulate why Renata refused it, naming the operator who did the work and would never sign the form. Dale could state Haddad's case for assurance over announcement, and Haddad could state Dale's case that assurance meant nothing to a client who could not evaluate it. Building a position clients could see⁴ had promised that demonstrable, well-governed competence would be a professional advantage; the committee had produced something adjacent to it, which was demonstrable, well-governed disagreement.

Then the signing, which returned to the vote of Chapter One. The paralegal and the self-represented litigant signed the majority text, the two seats the lawyers had nearly refused now anchoring the document. Two lawyers did not. The four-to-three that had seated Renata and Dale had become a durable four-to-three about how competence should be proved, and Vance noted that the narrowness she had flagged eighteen chapters earlier had held all the way through, which was not a failure of the process but its point.

She adopted the directive with its one permanent bracket, its published minority note, and its standing agenda of unresolved questions, Dale's access problem and his unanswered billing question among them, carried into the review by rule. The committee had not reached agreement about AI. It had reached a shared and demanding standard beneath honestly different intentions: alignment on what the profession must understand and how each clause must be tested, and open, respected divergence on what any given lawyer would do with a tool. That, Vance said as she closed the record, was the only kind of consensus a room of serious people ever actually shipped: not sameness of conclusion, but text that survived every objection made to it, including the ones it could not resolve.

End · Reasonable Doubts