The Profession's First Component Standard Was Written for the Machine

"A brief, pleading, motion, or any other paper filed in any court shall not contain any citations that the attorney responsible for submitting the pleading has not personally read and verified, including any citation provided by generative artificial intelligence." That sentence is California Senate Bill 574, amending Section 128.7 of the Code of Civil Procedure. It passed the Senate 39 to 0 in late January and sits in the Assembly as this memo goes out, with an August 31 deadline to act. Track its fate if you like; its fate is the least interesting thing about it. Whether the Assembly enacts it, amends it, or lets it die, a legislature has now drafted that sentence, a chamber has passed it unanimously, and the genre it belongs to · verify every output, in writing, as law · exists either way.

Now ask when the profession last wrote a sentence like that about anything else that produces work inside a law firm. It never did. Not for the paralegal, not for the first-year, not for the senior associate's markup, not for the partner's own first draft. Not once.

The commentary of the last month argues in the other direction, and the best of it deserves its name. CJ Saretto and Daniel van Binsbergen, in "We Were Both Wrong About Legal AI" (Mondaq, June 2026), make the forward case precisely: "The probabilistic tool with human review is a better system than the human review alone, not because the AI is always right, but because the combination catches more than either does separately." They are right, and their argument runs forward · add review to the machine and the system improves. This memo runs the argument backward. The law firm always was that system: ninety percent components under one hundred percent signature, every input presumed fallible and consumed under review. The genuinely new event of 2024 through 2026 is not the architecture. It is that the standard finally got written down · and that it was written for one component only.

1 · The standard nobody ever wrote

Define the genre before claiming anything about it. SB 574 is its purest specimen, but the form is everywhere now. ABA Formal Opinion 512 (July 2024) warns that "uncritical reliance on content created by a GAI tool can result in inaccurate legal advice to clients or misleading representations to courts," and that reliance on output "without an appropriate degree of independent verification or review" can violate the duty of competence under Rule 1.1. Call the genre what it is: the verify-every-output standard, a written rule that addresses the reliability of a component's individual work product before release.

Now the claim, stated in the only form that survives scrutiny, because precision here is the whole memo. The profession has written plenty of rules that touch the people in its production system. What it never wrote, in two centuries, is a per-output verification standard for any component of that system · a spec, addressed to the work product itself, prescribing what must happen to each unit before it counts. For every human component, the written law speaks in norms: Rule 5.3's operative verb is "reasonable efforts," calibrated, discretionary, supervisory. SB 574 speaks in spec: personally read and verified, every citation, every filing, no discretion in the sentence. A norm tells the supervisor to be careful. A spec tells the system what each output must survive. The profession ran on norms for its entire history, and the first time it reached for a spec, the component was a machine.

That distinction · norm versus spec, supervision versus per-output verification · is doing all the work in this memo, so it goes on the record here, in the first section, not in a footnote. Everything that follows defends it.

2 · The adverse witness, taken first

Before claiming a first, survive the witness who says there was never a chain at all. Jordan Furlong, in "The newest lawyer competence: Quality control" (December 2023), states the adverse finding exactly: "There are relatively few standardized quality checks for the work of associates and none at all for partners." He goes further: "Imagine telling a senior lawyer that their work will be evaluated by a fellow partner before it goes out to the client, or that it has to be carried out in a systematic manner based on quality standards and procedural protocols to which all work product must adhere."

Concede everything he documents. There was never a written spec. There was never a standardized check. If this memo claimed the profession ran a formal quality system, Furlong's sentence would kill it in the second section, which is why the claim is different: the chain was real but informal, and the informality is provable on the page.

The evidence set. The supervision rules predate the machine by four decades: the ABA Model Rules, including Rules 5.1 and 5.3, were adopted on August 2, 1983, and in 2012 the Ethics 20/20 amendments retitled Rule 5.3 from "Nonlawyer Assistants" to "Nonlawyer Assistance" · a one-word retitle that commentators have read as making clear the rule "encompassed any type of assistance, whether human or not," a reading they credit to Pierce and Goutos, "Why Lawyers Must Responsibly Embrace Generative AI" (2023). The clients wrote their half in outside counsel guidelines: GE's provide that "GE will generally not agree to the assignment of first year associates or summer associates to work on GE matters unless special permission is obtained"; Georgia Tech's.pdf) provide for advance agreement on "the number, names, and billing rates of the partners, associates, and paralegals who will be assigned" and prohibit billing for work "performed by over-qualified or under-qualified professionals." Note what these clauses are: staffing gates and review running to the client's own representative, never a spec for the firm's internal chain · the internal chain rests on Rule 5.1 plus those gates, which is precisely the point. And the carriers priced it: malpractice applications ask about docket control and conflict checks · Protexure calls a docketing process in which "at least two people are aware of what needs to be done and when" the single most effective risk control for deadline claims · and Illinois bar risk-management literature has urged peer-review-style checks and balances since at least 2001. Rules, client gates, carrier questionnaires, norms. A real chain. Not one written standard for any link in it.

Three counterexamples will have occurred to any careful reader, and each deserves its floor before the argument proceeds. Licensure first: the bar exam, character and fitness, continuing legal education · a written, profession-wide standard applied to every human in the chain. True, and look at where it sits: at the entrance. Licensure qualifies the producer once, at the door, and says nothing about any particular piece of work the producer ever releases. It is a gate-at-entry standard, the opposite of a per-output one. The ABA's Model Guidelines for the Utilization of Paralegal Services, first adopted in 1991 for legal assistant services and retitled for paralegals in 2004: a written standard for a specific component, but read it and it is Rule 5.3 at paragraph length · reasonable supervision, appropriate delegation, responsibility retained by the lawyer. Norms about the supervisor, never a spec for the output. Technology-assisted review is the serious one: since Da Silva Moore v. Publicis Groupe (S.D.N.Y. 2012), courts have blessed written validation protocols for predictive coding in discovery · sampled, measured, quality-controlled, on paper. A written, per-output-sampled standard for a machine's work, twelve years before SB 574. But mark its three limits: matter-specific, negotiated per case; court-supervised, not profession-wide; and confined to one task in one phase of litigation. TAR does not falsify the claim. It rehearses it: the one prior time the profession wrote down a sampled quality standard for anything, the component was also a machine. The pattern is now two for two. The profession writes specs only when the component cannot be a colleague.

3 · The backward read

So read the firm's history in the direction the evidence points. The paralegal's chronology arrived presumed incomplete. The first-year's memo arrived presumed wrong in ways the assigning partner budgeted time to find. The senior associate's markup arrived presumed to need a final pass. The partner's own first draft arrived presumed to be a draft · that is what the word means. Every unit of work in the building was created at component grade, consumed under review, and released by exactly one instrument: the signature. No component, human or otherwise, ever had a solo-release path. And no component was ever asked to be perfect, because the architecture never required it; it required only that the chain converge on one accountable hundred at the line where the firm's name attached.

One sentence on terms, because this series has used the word before: "component" here means a trust grade · a unit of work presumed to need review before release · not the unit-of-cost sense it carries in "The Matter Never Repeats. The Component Always Does." and "A Legal Transaction Has a Bill of Materials. Here Is One." And one sentence on the pyramid: read here, the leverage pyramid is reliability architecture · layered fallibility converging on one signature · while its reading as a financing instrument belongs to "Stranded Tuition: The Pyramid Was a Financing Instrument, and the Financier Pulled the Line First" and is not re-run here.

Name the structure, because it needs a name to be forwardable. The only place one hundred percent ever lived in a law firm was the signature line. Everything beneath it ran at ninety. Call it the signature's hundred. The verify-every-output genre takes the signature's hundred and bills it to a component · a category error about where the firm's certainty has ever lived. Which raises the question the rest of this memo answers: if the building tolerated component-grade work from every human who ever drafted in it, why did the machine, alone, get a statute?

4 · The failure that arrives in the costume of competence

Because the machine broke the chain's error detector, and the honest version of that point belongs to Jack Shepherd, who published it first. In "Generative AI in the legal industry: is accuracy everything?" (November 2024), Shepherd observes that "well-presented work free of typos often instils confidence from the start," and then lands the asymmetry: humans "do sometimes produce bad but well-presented work. LLMs always do." And: "That makes it harder to spot the errors."

That is the conceded limit of the backward argument, and it deserves full credit. The informal chain worked because human failure is legible to humans: the junior's bad memo usually looks bad, and a partner's heuristics were trained on that distribution for a career. Machine failure inverts it. D.C. Bar Ethics Opinion 388 (April 2024) states the mechanism plainly: GAI products sometimes "hallucinate," meaning they make up things that do not exist. Judge Castel, sanctioning the Mata v. Avianca lawyers (S.D.N.Y. 2023), described the fabricated opinion their tool produced: it "shows stylistic and reasoning flaws that do not generally appear in decisions issued by United States Courts of Appeals. Its legal analysis is gibberish" · yet it carried a caption, a docket number, and the full costume of authority. The Stanford benchmarking team (Magesh et al.) found that even in purpose-built legal research tools, hallucinations "remain substantial, wide-ranging, and potentially insidious."

So the standard got written for the machine because the machine is the first component whose failures defeat the informal chain's detection habits. The chain ran on the assumption that error announces itself. This component's errors are fluent. That is why the spec exists · and it is also why the obvious fix is wrong.

5 · The analogy that fails in both directions

The cliché in every partner meeting is "supervise it like a first-year." Refuse the analogy explicitly, because it fails twice.

It understates how systemic the old tolerance was. Component-grade work was the entire architecture, up to and including the partner's own first draft; treating AI "like a junior" implies the building once held everything else to a higher grade. Section 3 says it never did.

And it overstates how transferable the supervision is. The partner's review habits are calibrated to human failure distributions, and the automation literature has spent forty years documenting what happens when those habits meet machine output. Parasuraman and Manzey (Human Factors, 2010) found that automation complacency "is found in both naive and expert participants and cannot be overcome with simple practice," and that it bites hardest under exactly the multitasking load a supervising partner works under. Goddard, Roudsari, and Wyatt (JAMIA, 2012) documented the same bias in clinicians reviewing decision-support output · the professional-services bridge, and the very authority the Stanford team imports into law. Magesh et al., writing of the subtler errors their legal benchmark surfaced, state the danger directly: "These errors are potentially more dangerous than fabricating a case outright, because they are subtler and more difficult to spot." Recycled human-review habit, pointed at a component that fails fluently, does not transfer the old chain's reliability. It amplifies the new component's errors while feeling like diligence.

The analogy is not a plan. It is nostalgia for an error distribution that no longer describes the input.

6 · The practice pushes back

The strongest objection comes from the 30-year practitioner, and it deserves printing at full strength:

"Your history is backwards in the wrong way. First, every human in my chain carried a written standard · the bar exam, character and fitness, CLE, the whole licensure apparatus. Your machine is the first component in the building with no license, which is exactly why it got a statute; that is continuity, not novelty. Second, the informal chain worked because every layer was a trained human whose failure modes I learned over a career and whose own career depended on not failing me. The associate had skin in the game. Your machine has none. Third, your 'first component standard' is a press release: SB 574 amends Section 128.7, restating a duty to check my citations that Rule 11 and basic competence put on me decades ago. You are dressing the oldest rule in practice as a historic novelty in order to sell an operations layer."

Concede what must be conceded, and concede it in full. The substance of the duty is old: SB 574 amends an existing sanctions statute, and Opinion 512 expressly runs through Rule 1.1; no lawyer was ever free to file unread citations. Concede the incentive point entirely: the associate's career was a genuine reliability mechanism, a decade of skin in the game that no machine can replicate, and the informal chain was load-bearing on it. And concede the licensure record: every human in the chain did carry a written standard.

Now look at where that written standard sat, because the licensure point is not the objection's strength · it is the memo's. The license is a gate at the entrance: it qualifies the producer once, at the door, and then the profession trusted the chain, in norms, for the rest of that career. Not one written word ever addressed the producer's individual outputs. The machine cannot sit for the bar, cannot be character-and-fitnessed, cannot fear disbarment · so for the first time, the profession moved the written standard from the entrance to the output. That is the whole event. Not a new duty; a new form. For humans, a written gate at entry, then norms. For the machine, a spec at the output · personally read and verified, each citation, in statute. The practitioner's own framing proves the novelty he denies: the statute exists because no license could.

And his skin-in-the-game point, fully conceded, is the door to the terminus. If the informal chain's reliability ran on incentives the machine cannot carry, then the machine's review layer cannot be assumed the way the associate's diligence was. It has to be engineered.

7 · The standard is an ops product

Here is what the genre's own texts already half-admit. Opinion 512, the same document that warns against uncritical reliance, calibrates the duty rather than absolutizing it: "The appropriate amount of independent verification or review required to satisfy Rule 1.1 will necessarily depend on the GAI tool and the specific task," with subset validation sufficing for some uses and lighter review for idea generation. That is not the language of an ethics homily. That is the language of a system specification: sampling rates, task classes, tool-specific calibration. SB 574 sits at one extreme of the genre · read everything yourself · and Opinion 512 at the other, already gesturing at sampled, logged, evidence-based verification. Between them is a design space, and a design space implies a designer.

A review layer built for AI's actual failure distribution · sampled where validation supports it, logged so the sampling is defensible, calibrated to confident fabrication rather than visible sloppiness, instrumented against the complacency the automation literature guarantees · is an operations product. It has an owner, a budget, a roadmap, and an error rate. It is not an ethics memo that everyone initials and nobody operates. Who pays for verification, and which entity is economically positioned to bend its cost curve, is settled ground in "The Partner Without a Computer Was Reading His Balance Sheet Correctly" and handed off here in one line. That the operations layer beneath the practice is the one parcel of a legal business an operating company can own and rebuild is the published conclusion of "The Only Ownable Acre in a Law Firm", referenced rather than re-derived.

The backward image closes the loop. For two centuries the profession ran components under a signature, governed the humans with a license at the door and norms in between, and never wrote down what any output had to survive · it never needed to, because human failure announced itself and human careers underwrote the rest. Then a component arrived that could do neither, and the profession did something it had never done for any paralegal, associate, or partner's own first draft: it wrote the standard down. The buildings always practiced it. The letter was simply never sent · until it was addressed to the machine.

What the Building Always Tolerated vs. What the Machine Is Asked to Be

This table is the authors' analytical summary; sources are as cited in the text. Two definitions keep it honest. "Solo-release path: Never" means no first draft, including the partner's own, was ever the signed deliverable · release always ran through the signature. It does not mean partners' work was checked by anyone else; Furlong's finding in Section 2 that there were no standardized quality checks for associates "and none at all for partners" stands in full, and is the informality the memo is built on. The last column's first row of "Yes" is the first time the standard was ever written down. The release column never changed.

Jopese operates a legal MSO, and the operations layer this memo ends on is the ground it occupies.

This memo is published by Jopese, a legal management services organization operated by HIRO PARTNERS LLC, a Texas limited liability company. It is offered for educational and analytical purposes only. It is not legal, tax, or investment advice, and it is not an offer to sell or a solicitation of an offer to buy any security or service. Jopese is not a law firm and does not provide legal advice or legal services; legal services are delivered by an independent law firm under a separate engagement in which Jopese does not participate. The table above is the authors' analytical summary and does not describe any actual firm, agreement, or transaction, including any to which Jopese is a party. References to legislation, ethics opinions, bar guidance, judicial decisions, scholarship, publications, and companies are drawn from public sources and are provided as market commentary, not as an endorsement, a recommendation, a prediction of any legislative outcome, or a representation of any relationship. "The signature's hundred," "component standard," and "the first written component standard" are the authors' analytical labels, not industry terms of art. Legislative postures described are as of mid-June 2026 and subject to change. Questions: rene@jopese.com.