Over two days this week the firm researched, wrote, verified, and published thirty long-form articles on Minnesota law. Five drafting agents worked in parallel from written assignment briefs. Independent verification agents — which had written none of what they were checking — then re-derived every citation, every quotation, and every pin cite from the primary source before anything went live.
The most useful thing that run produced was not the articles.
At least eight of the legal premises I handed down in those briefs were wrong. Not debatable, not close questions. Wrong in the way that publishes false law: a one-year bar described as two years, a modification standard attributed to a subdivision the legislature had repealed, a rule of decision stated exactly backwards, a statutory attorney-fee remedy that does not exist. Every one was caught before publication by a subordinate who had been instructed to verify the assignment rather than execute it — and who did.
So here is the thesis, stated plainly enough to argue with: the instructions you hand down are hypotheses, not ground truth, and the system has to be built so the subordinate can prove the boss wrong. In any pipeline that touches primary law, the senior lawyer’s memory is the least reliable component in it.
I should be precise about who “I” am in that sentence. The briefs came off the top of the pipeline — from me and from the coordinating agent that works from my direction and its own accumulated familiarity with this body of law. For the purposes of this essay the two are indistinguishable, and that is itself the point. Both were writing down what they were confident the law said. Neither had opened the statute.
Confidence and currency run in opposite directions
Here is the structural problem, and it is not about anyone’s competence.
The person who assigns the work is the one with the most experience of the subject and the least recent contact with its text. That is what seniority is. You know the shape of chapter 518 because you have litigated inside it for years, and the thing you actually know is a compressed, load-bearing summary built out of a hundred prior encounters — most of them with a version of the statute that has since been amended.
Meanwhile the subordinate has no compressed summary at all. They have to go read it. Which means the least experienced participant is reliably the only one working from the current text.
Run that forward and you get a specific, predictable failure: the confidence attaches to the memory, and the currency attaches to the reading, and they sit in different people. A senior lawyer will state a rule with the fluency earned by fifteen years of practice, and be describing a subdivision repealed in 2024. Nothing about the delivery signals which one it is. That is why this failure mode survives — it does not feel like guessing from the inside.
This is also why the AI framing is a red herring. I have been describing agents, but every sentence above applies unchanged to a partner briefing an associate. The agents just made the failure rate measurable, because for once there was a written record of what the boss asserted and an auditable record of what the subordinate found.
The eight
Concreteness is worth more than argument, so here is the whole list, with what I asserted and what the source actually says. Everything below was re-verified against the primary source for this essay, not carried over from the run.
1. A “two-year moratorium” on custody modification. I described Minn. Stat. § 518.18 as containing a two-year bar. It contains two different clocks and the first one is one year: paragraph (a) bars a modification motion “earlier than one year after the date of the entry of a decree,” and paragraph (b) bars a subsequent motion “within two years after disposition of the prior motion on its merits.” Collapsing those into one number is a malpractice-grade error in a family practice. The brief also referred to the statute’s “subdivisions.” Section 518.18 has none — it runs (a) through (f).
2. Maintenance modification under § 518A.39. I sent the drafter to chapter 518A for the spousal-maintenance modification standard. The Revisor’s page for § 518A.39 now shows, at subdivision 3, the line “MS 2022 [Repealed, 2024 c 101 art 2 s 11].” The 2024 act moved maintenance modification into § 518.552, subd. 5b, a new subdivision — the section’s history line ends “2024 c 101 art 2 s 1-8.” I was citing a subdivision that had been repealed out from under me two years earlier.
3. Promissory estoppel as a consideration substitute. I proposed the thesis that promissory estoppel “substitutes for consideration, not for definiteness.” Minnesota holds the opposite. Constructors Supply Co. v. Bostrom Sheet Metal Works, Inc., 291 Minn. 113, 120, 190 N.W.2d 71 (1971): “Promissory estoppel is not a substitute for acceptance, consideration, or mutuality, but a doctrine based on reliance which courts may use in a proper case to prevent injustice.” That is not a nuance I got slightly off. It is the sentence that says the assigned thesis is wrong, and the drafter declined to write the assigned thesis.
4. Part payment revives a time-barred debt. I handed down the standard collection-defense chestnut — that a small payment restarts the limitations clock — as the article’s thesis. Minnesota abolished it for consumer debt in 2013. Minn. Stat. § 541.053: “After its expiration, the statute of limitations is not revived by the collection of a payment on an account, a discharge in a bankruptcy proceeding, or an oral or written reaffirmation of the debt.” History line: 2013 c 104 s 2. The article that published says the reverse of the article I assigned.
5. An eleven-element fraud claim. I asked the drafter to verify “the eleven-ish fraud elements as Minnesota actually states them,” pointing at Hoyt Properties. Hoyt Props., Inc. v. Production Res. Grp., L.L.C., 736 N.W.2d 313, 318 (Minn. 2007), states five, quoting Specialized Tours, Inc. v. Hagen, 392 N.W.2d 520, 532 (Minn. 1986). The eleven-factor formulation is real but older and from a different line — the Court of Appeals collects all three formulations in TCI Business Capital, Inc. v. Five Star American Die Casting, LLC, 890 N.W.2d 423, 432 (Minn. App. 2017), which cites Davis v. Re-Trac Mfg. Corp. for the “similar eleven-factor test.” The published article reports all three rather than adopting my premise, which is the better article.
6. An attorney-fee provision in the civil theft statute. I asked the drafter to “verify the punitive multiplier, who can recover, attorney-fee language” in Minn. Stat. § 604.14. There is no attorney-fee language. The section has six subdivisions — liability, notice, parent-or-guardian liability, criminal action, recovery of property, right to demand payment — and none of them mentions fees. The remedy is value plus “punitive damages of either $50 or up to 100 percent of its value when stolen, whichever is greater.” That negative became a load-bearing point in the published piece, because a plaintiff who assumes fee-shifting here has mispriced the case.
7. A total-loss rule in the wrong rule chapter. I pointed at “Minn. R. ch. 2767 unfair claims settlement” for auto total-loss valuation. Chapter 2767 is titled INSURANCE PLAN ADMINISTRATORS. The automobile chapter is 2770, AUTOMOBILE INSURANCE, and the standard the drafter was actually looking for is a statute, not a rule: Minn. Stat. § 72A.201, subd. 6, “Standards for automobile insurance claims handling, settlement offers, and agreements.”
8. An eviction-expungement recodification that never happened. I flagged a “possible recodification into ch. 504B.” It did not happen. Minn. Stat. § 484.014 is live, titled “HOUSING RECORDS; EXPUNGEMENT OF EVICTION INFORMATION,” with a history line running through “2024 c 118 s 1.” The drafter’s job here was to prove a negative, and the article now says expressly that the law was not moved.
Read as a set, those are not typos. Two of them would have inverted an article’s central holding. Four of them would have sent a reader to a provision that does not say what the brief claimed — one of which no longer exists. All eight were stated in the assignment with no hedge at all, because I believed them.
The deference failure mode
Now consider how this ordinarily goes in a law office.
A partner tells an associate the rule. The associate, reading the statute an hour later, finds something that does not match. What happens next is decided almost entirely by economics that have nothing to do with the law.
Correcting upward is expensive. It costs social capital, it costs time the associate is being judged on, and it carries a live risk of being wrong in front of the person who writes the review. Deferring is free. So the associate does what the incentives say: assumes the partner is working from something they have not read yet, writes around the discrepancy, and softens the sentence until it is true of both versions. The brief goes out with a hedge in it where a holding should be, and nobody ever learns that the partner’s recollection was two amendments stale.
That is the deference failure mode, and it is not a character flaw in associates. It is a designed-in property of any hierarchy where the cost of raising a correction exceeds the cost of absorbing an error. If you want the correction, you have to change the price.
There is a second-order cost too, which is that the senior never gets the update. A lawyer whose subordinates route around their errors instead of reporting them gets more confident over time, not less, because the errors keep not surfacing. The feedback loop that would fix the memory is the exact loop the deference norm severs.
Say “lead” in the brief
The first and cheapest intervention is a sentence. The 2026-08-18 brief opened with this, above everything else:
Nothing below is verified. Every statute, subdivision, case, and holding named in an assignment is a lead, not a fact. Verify against the primary source before writing it; several section numbers handed to drafters on this project have turned out not to exist.
The next day’s brief escalated it, because by then there was evidence:
…section numbers handed to drafters on this project have turned out not to exist, and three assignment theses in the last batch were wrong and had to be reversed by the drafter.
Note what those sentences do. They are not a disclaimer and they are not humility as a posture. They relocate the burden. “Verify against the primary source before writing it” converts the assignment from an instruction into a research question, and it makes writing an unverified premise — even mine — a violation of the brief rather than compliance with it.
The second sentence does more work than the first. Citing the prior failures is what makes the permission credible. Standing permission to contradict the boss reads as a formality until the boss documents having been contradicted and having liked it. Three reversals last batch is not modesty. It is a base rate, and a base rate tells the subordinate what number of corrections is expected of them.
Permission is worthless without retrieval
A grant of authority to verify is empty if verifying is hard. Most of the reason associates defer is not courage; it is that checking the partner’s rule properly costs forty minutes and the deadline is tonight.
So the same briefs that granted the permission also handed over the tools and named the traps, in the same breath:
Verify statutes with
python3 drafts/mnstat.py <cite>(run from repo root), court rules withdrafts/mnrule.py, cases withdrafts/findcase.py/ static.case.law / MN Judicial Branch slip-opinion archive. NOT WebFetch on revisor.mn.gov.
That last prohibition exists because a summarizing fetch layer silently truncates long statutory text and has produced wrong quotes on this project — the reason I wrote a scraper instead. With the extractor, checking whether § 604.14 contains an attorney-fee provision costs about four seconds. At four seconds, nobody rationalizes. At forty minutes, everybody does.
The design rule generalizes past agents: verification has to be cheaper than deference, or deference wins every time. If you want associates who check you, the investment is not a speech about intellectual courage at the retreat. It is whatever makes the check take a minute.
Make “the assignment was wrong” a reportable success
Here is the part I did not design, and should have.
The brief told drafters to write a manifest for each article listing every authority, the proposition it supports, and the URL it was verified at. It said nothing about reporting errors in the assignment. So the drafters invented the channel themselves. Three separate agents, working independently, added sections to their manifests that I never asked for:
- “Correction to the assignment lead — flagged for the coordinator”
- “Deliberate corrections to the assignment’s leads — flag for the gate”
- “Assignment leads that did not check out — reported to the coordinator”
That is what it looks like when the permission is real. Given a brief that called its own contents leads, three subordinates independently concluded that contradicting the boss was a deliverable and built a place to put it.
But the fact that they had to build the place is a design failure on my part, and it is the kind that silently reduces the yield. Some fraction of corrections never gets written down when there is no field for it — the drafter fixes the article, moves on, and the senior’s wrong belief survives to be handed to the next person. If the correction has nowhere to go in the output format, you are relying on initiative to carry your quality control. Every assignment brief I write from here forward has an explicit section for it, empty by default and expected to be filled.
The same logic applies to associates, and it is the cheaper half of the reform. Put “assignment premises that did not check out” on the research memo template. Not as a courtesy. As a line item, like the question presented.
Never let the author of a claim be its verifier
The last piece is structural and it is the one I would not give up.
The drafters wrote manifests asserting that every citation had been verified. Those manifests are claims to be tested, not proof — a point the citation gate exists to enforce. So the verification agents were given a different brief and told, in terms:
Assume at least one error exists in every article and go find it. An article that comes back clean on the first pass has usually not been read adversarially.
And verification runs in both directions. The same run that caught eight of my errors caught this one of a drafter’s. An article quoted a Minnesota Supreme Court opinion as describing conduct amounting to “nothing other than routine accounting services.” What Witzman v. Lehrman, Lehrman & Flom, 601 N.W.2d 179, 189 (Minn. 1999), actually says is that the plaintiff failed to allege that the firm “provided anything other than routine accounting services” — a negated clause. One inverted word turned a court’s statement about a pleading failure into an affirmative judicial characterization of the defendant’s conduct. The phrase as quoted appears nowhere in the opinion.
The reason that got caught is the reason the gate is a separate agent: the drafter’s own manifest did not flag it. It could not. The manifest was written by the same process that made the error, and it recorded the page the material came from without re-reading the sentence. An author checking their own quotations re-reads what they meant to write. The whole value of the verifier is that it has no idea what you meant.
Which yields the rule in one line: the author of a claim is disqualified as its verifier — and that disqualification runs up the hierarchy, not just down. I do not get to certify my own briefs either.
What would prove this wrong
Two things, and I would want to know about both.
If the corrections were mostly false alarms. An over-eager verification culture generates its own error class: subordinates who “correct” a senior who was right, and a pipeline that burns hours litigating settled points. I checked against that here by re-deriving all eight for this essay from the Revisor and from archival opinion text, and all eight held. But eight-for-eight in one run is not a stable claim about the world, and if a later run produced ten corrections of which six were the subordinate’s error, the design would need a different balance.
If a more careful senior simply would not make these mistakes. Maybe the honest reading is that I should write better briefs rather than build a system that tolerates bad ones. I do not believe it, and here is the falsifiable version: any experienced lawyer who writes down thirty confident statements of law from memory will be wrong on something like a quarter of them. The test is free and you can run it this afternoon. Write down the ten rules you are surest of in your own practice area — subdivision numbers, deadlines, elements, the operative sentence. Then open the statute. My prediction is two or three, and my prediction is that the two or three will be the ones you have said out loud to a client most recently, because frequent use is what stops you from re-reading.
If that number comes back zero for you consistently, I am wrong about the mechanism and you should write it up.
What this costs, and who it is for
None of this is expensive. It is a sentence in the brief that calls the assignment a lead, a base rate that proves the permission is real, tooling that makes checking cost seconds, a field in the output where the correction goes, and a verifier who did not write the thing.
What it buys is a firm where being wrong is discovered internally, early, by someone who reports it — rather than externally, late, by opposing counsel in a responsive brief or by a client who relied on it. That is not a technology outcome. It is the ordinary supervisory obligation, discharged by design instead of by hoping.
And it matters most at the small end. A large firm can absorb a stale-subdivision error because there are three more lawyers between the assumption and the filing. A solo has none of that, and until very recently had no realistic way to buy it. The reason I am writing this in a section that gives everything away is that the correction layer — the thing that used to be a senior associate you could not afford — is now a brief, a scraper, and a standing instruction. Any lawyer serving people who could not otherwise afford one can have it by Friday.
The first sentence of every assignment I write now is the same one. Nothing below is verified. Go find out.
Sources
- Minn. Stat. § 518.18 — paragraph (a) one-year bar on custody modification motions; paragraph (b) two-year bar on a subsequent motion after disposition on the merits; section is organized in paragraphs (a)–(f), not subdivisions
- Minn. Stat. § 518A.39 — subd. 3 shown as “MS 2022 [Repealed, 2024 c 101 art 2 s 11]”
- Minn. Stat. § 518.552, subd. 5b — “Modification”; history line ending “2024 c 101 art 2 s 1-8”
- Minn. Stat. § 541.053 — consumer-debt limitations period; revival by payment, bankruptcy discharge, or reaffirmation abolished; enacted 2013 c 104 s 2
- Minn. Stat. § 604.14 — civil liability for theft, subds. 1–6; punitive damages of “$50 or up to 100 percent of its value when stolen, whichever is greater”; no attorney-fee provision in any subdivision
- Minn. Stat. § 72A.201, subd. 6 — “Standards for automobile insurance claims handling, settlement offers, and agreements”
- Minn. Stat. § 484.014 — “HOUSING RECORDS; EXPUNGEMENT OF EVICTION INFORMATION,” history through 2024 c 118 s 1
- Minn. R. ch. 2767 — “CHAPTER 2767, INSURANCE PLAN ADMINISTRATORS”; Minn. R. ch. 2770 — “CHAPTER 2770, AUTOMOBILE INSURANCE”
- Constructors Supply Co. v. Bostrom Sheet Metal Works, Inc., 291 Minn. 113, 120, 190 N.W.2d 71 (1971) (No. 42562) — quoted sentence read at star page 120 of the Caselaw Access Project archive, static.case.law/minn/291/html/0113-01.html
- Hoyt Props., Inc. v. Production Res. Grp., L.L.C., 736 N.W.2d 313, 318 (Minn. 2007) (No. A05-1293) — five-element fraudulent-misrepresentation statement quoting Specialized Tours, Inc. v. Hagen, 392 N.W.2d 520, 532 (Minn. 1986); read at static.case.law/nw2d/736/html/0313-01.html
- TCI Business Capital, Inc. v. Five Star American Die Casting, LLC, 890 N.W.2d 423, 432 (Minn. App. 2017) (No. A16-0741) — collecting the Hoyt five-element, Martens seven-factor, and Davis v. Re-Trac Mfg. Corp. eleven-factor formulations; read at static.case.law/nw2d/890/html/0423-01.html
- Witzman v. Lehrman, Lehrman & Flom, 601 N.W.2d 179, 189 (Minn. 1999) (No. C6-98-555) — “provided anything other than routine accounting services”; the phrase “nothing other than routine” appears nowhere in the opinion; read at static.case.law/nw2d/601/html/0179-01.html
- Reporter citations above are taken from the Caselaw Access Project’s structured case metadata rather than from opinion body text. The Caselaw Access Project has no citator; nothing here should be treated as a subsequent-treatment check.
- The assignment briefs, per-article manifests, and independent gate reports quoted in this article are this firm’s own working files from the 2026-08-18 and 2026-08-19 publishing runs, maintained in the site repository. Thirty articles carry those two publication dates.
Commentary on practice management and supervision — the opinions and the predictions are the author’s. Not legal advice, not ethics advice, and not a statement of the law for any reader’s situation: the statutes and cases above are cited only as examples of a delegation failure and were current as of the date of this article. No client information appears in this piece; every example comes from the firm’s own published work on its own website. Everything in this section is free, and the only ask is that you use it to serve someone who could not otherwise afford a lawyer. Questions about anything here: Send us a message or 612-470-6529.