SignedOff

1. One-liner

Builds the dated record of who checked every citation, before a judge asks you to prove it.

2. Trend signal — why now?

Three things moved, and they moved in a specific order that matters.

The sanctions curve is steepening, not flattening. Damien Charlotin’s AI Hallucination Cases database — the reference tracker, maintained by a research fellow at HEC Paris — went from roughly 200 cases in mid-2025 to 719 by January 2026, 1,227 by early April 2026, and 1,598 by 9 June 2026. Between 22 May and 9 June it added 140 cases, just under 8 per day, up from 5–6 per day in April. Courts imposed roughly $145K in AI-citation sanctions in Q1 2026 alone, with penalties climbing from four-figure fines to $15,000 per attorney at the federal appellate level, plus the first bar suspensions tied to AI filings.

Courts flipped from punishing to pre-certifying. This is the part most people missed. Through 2025 the enforcement posture was reactive — sanction the lawyer after the fake case surfaces. In 2026 it became proactive. Florida’s Supreme Court amended Rule 2.515(d)(2), effective 15 June 2026, so that every signer of a filing represents that “the legal authorities identified exist and are accurately cited.” New York’s 22 NYCRR Part 161 took effect 1 June 2026. Ohio’s Garfield Heights Municipal Court Rule 14 landed 23 July 2026. Note what Florida and New York did not do: they did not require you to disclose AI use. They made you certify the output is real, regardless of how you produced it. The duty attached to the signature, not the tool.

And the layer underneath is a mess. The binding obligations mostly live judge-by-judge, in standing orders and individual practice rules. Ropes & Gray counted 145 tracked orders requiring disclosure and/or verification as of June 2026; another tracker counts 113 active orders binding attorney filings. They are not uniform — some require disclosure of any AI use, some require certification that citations were verified, some require both, some merely caution. At least 25 federal district courts have adopted standing orders or local rules on AI certification.

The consequence, and the reason this is a product: courts have begun asking attorneys to describe the verification, not just assert it. Judge Cronan requires parties to certify they have “personally reviewed the filing for accuracy” and to describe “in detail the steps taken to verify” AI-generated material. Practitioner guidance on surviving a show-cause order is now explicit that “a dated attestation in the matter file documenting who owns verification and providing sign-off is a strong mitigator.”

The fourth signal, which turns a rare fear into a recurring purchase: malpractice carriers writing lawyers professional liability policies began issuing AI-specific underwriting questionnaires in the 2026 cycle. They ask which AI tools the firm uses, whether there is a written AI governance policy, whether attorneys get AI-specific training, and — the operative one — how AI-generated work product is reviewed and verified. AI-specific endorsements run roughly 5–15% above base premium, and firms that cannot demonstrate governance face higher premiums or restrictive endorsements.

Provenance:

• Signal 1 (Demand): AI hallucination sanctions database grew 200 → 1,598 cases (mid-2025 → 9 June 2026), ~8 new cases/day and accelerating; $145K in Q1 2026 sanctions, $15K/attorney at appellate level, first bar suspensions — https://www.damiencharlotin.com/hallucinations/ and https://complexdiscovery.com/the-ai-sanction-wave-145k-in-q1-penalties-signals-courts-have-lost-patience-with-genai-filing-failures/ — observed 2026-09-02
• Signal 2 (Feasibility/Regulatory): Courts shifted from reactive sanctions to mandatory pre-filing certification — FL Rule 2.515(d)(2) eff. 2026-06-15, NY 22 NYCRR Part 161 eff. 2026-06-01, OH Garfield Heights Rule 14 eff. 2026-07-23; 145 tracked judge/court orders requiring disclosure or verification as of June 2026 — https://www.workplaceprivacyreport.com/2026/05/articles/artificial-intelligence/ai-hallucinations-in-court-filings-continue-florida-supreme-court-responds-with-a-new-certification-requirement/ and https://www.druganddevicelawblog.com/2026/04/courts-get-proactive-on-ai-disclosure-certification-and-consequences.html — observed 2026-09-02
• Signal 3 (Economic): Malpractice carriers now issue AI-specific underwriting questionnaires asking how AI work product is “reviewed and verified”; AI endorsements price at 5–15% above base premium, and firms without demonstrable governance face higher premiums or restrictive endorsements — https://legalaigovernance.com/resources/ai-liability-insurance/ and https://www.aivortex.io/legal/ai-governance/ai-insurance-liability-law-firms/ — observed 2026-09-02 Category: Regulatory arbitrage

3. The opportunity

The citation-checking market is finished. It is solved, crowded, and cheap. CiteCheck AI runs $25/mo for 100 reports and is free for the first five. Clearbrief is $300/user/mo and wired into LexisNexis. Vincent AI (vLex, now part of Clio) sits around $69/user/mo self-serve. Lexis+ with Protege has Shepard’s. Sentinel scores citations with a three-model jury. Anyone building citation verification in September 2026 is arriving five years late to a knife fight.

Every one of those tools answers the same question: is this citation real? They emit a verdict — valid, invalid, misattributed — and that verdict is the deliverable.

The question the court actually asks is different: prove what you did before you signed. Judge Cronan wants the steps described in detail. Florida wants the signer’s representation that the authorities exist. A show-cause respondent wants a dated attestation showing who owned verification and when they signed off. That is not a verdict. That is a chain of custody over an act of professional diligence — who ran what check, against which source, on which draft, at what timestamp, and which human accepted the result.

No checker produces this. They are stateless. You paste a brief, you get a report, you close the tab. If the report is wrong — and open-source checkers cross-referencing CourtListener miss things, because CourtListener is not complete — you have no record that you checked at all, let alone that you checked reasonably. The tool that told you the citation was fine does not remember telling you.

The same asymmetry appears on the other side of the file. The trackers that list which judge requires what are free and permanent: Ropes & Gray, Law360’s AI tracker, RAILS, Bloomberg Law, legalaigovernance.com. They are law-firm and vendor marketing funnels. Nobody will ever charge for that list, and I would not try. But nobody connects the list to the filing — no tool looks at the caption on your brief, resolves the assigned judge, and tells you this judge’s standing order requires a certificate of compliance while the one down the hall does not.

So the gap is a seam between three things that all exist separately: a commoditized checker, a free tracker, and a signature that now carries a documented duty. The product is the record that ties them together and survives being handed to a judge eighteen months later.

The 10× is not better verification. It is that today a solo litigator’s verification record is nothing — a memory, maybe a printed Westlaw page in a folder, most likely an assumption that the associate did it. When the show-cause order arrives, the difference between a discharged order and $15,000 plus a bar referral is whether you can produce a contemporaneous, dated account of your process. Manufacturing that record after the fact is both impossible and, if attempted, a much worse problem than the original citation.

4. Target market

Primary customer: Owner of a US litigation practice with 1–10 attorneys — solo practitioners and small firms doing civil litigation, family law, personal injury, immigration, and criminal defense. Filing at least a few briefs or motions a month in state or federal court. Using ChatGPT, Claude, Copilot, or a legal AI tool for drafting, often without telling anyone. No knowledge-management staff, no compliance officer, no professional-responsibility partner. The buyer is the named partner, because there is nobody else.

Why they buy: Two triggers, one annual and one existential.

The annual one is the malpractice renewal questionnaire, which now asks in writing how AI work product is reviewed and verified, and prices the answer into the premium. A solo pays roughly $2,300/yr in premium on average; small firms of 2–5 attorneys run $5,000–$25,000. A 5–15% AI endorsement is real money, and “we have a documented verification protocol with per-filing attestations” is a materially better answer than a shrug.

The existential one is the show-cause order. It is low-probability per filing and catastrophic per occurrence — public, reportable, insurance-relevant, and increasingly a bar-discipline matter rather than a fine. Note that courts have responded more leniently to attorneys who accepted responsibility and could account for their process, and more harshly to those who denied AI use when evidence showed otherwise.

I want to flag an honesty problem in this section rather than paper over it. I found extensive trade-press and law-firm commentary written at solo litigators about this, and a well-documented sanctions record, but I did not find a body of verbatim solo-practitioner complaints — no Reddit thread of lawyers saying “I wish something logged my cite checks.” The demand evidence here is regulatory, actuarial, and judicial rather than vocal. That cuts both ways: it is why the seam is unclaimed, and it is the single biggest risk in this proposal. See §13.

Rough TAM reasoning: Roughly 49% of US private-practice lawyers work solo and about 40% of US firms are one lawyer; small firms of 2–10 add another ~20% of lawyers. That is a very large denominator, but the addressable slice is narrower: firms that litigate (rather than transact), file regularly, and use AI. Call it 60,000–90,000 US firms with a genuine, recurring filing-certification exposure. At $49/mo blended that is a theoretical ceiling well north of $40M/yr, which is larger than I need and larger than I believe — realistically this is a $2–4M ARR business, which is exactly the target.

Why now for them: Florida’s rule took effect 15 June 2026 and New York’s 1 June 2026. Their next malpractice renewal will ask the AI question in writing. And the sanctions database is adding eight cases a day, several of which are solo practitioners in exactly their practice areas.

5. Product sketch (MVP)

• Drop a draft, get a verification pass. Upload the brief; every citation is extracted and checked against CourtListener and the free case-law corpus, flagged as verified, unverified, or not-found — with the unverified ones ranked by how load-bearing they are in the argument.
• Court-aware rule lookup. Enter the caption or select the court and judge; the product tells you what that forum currently requires — bare Rule 11, a disclosure statement, a certificate of compliance, or a describe-your-steps certification — with a link to the primary-source order.
• The attestation. One dated PDF per filing: which citations were checked, against what source, on which draft version, with what result, which human reviewed each flagged item, and who signed. Hash-stamped and immutable once signed.
• Certificate and disclosure drafting. Generates the specific certificate language the assigned judge’s order calls for, in the format that court expects — not a generic template.
• Matter file. Every attestation filed under its matter, retained and searchable, so an order to show cause eighteen months later is a two-minute retrieval, not an archaeology project.
• Insurance packet. One-click export summarizing firm-wide verification practice — policy, training log, per-filing attestation counts — sized for the carrier’s AI questionnaire at renewal.
• Written AI-use policy generator. The firm’s governance document, which the carrier questionnaire asks about directly and which almost no small firm has.

6. AI angle — what’s load-bearing

Honest answer: partially load-bearing, and I would rather say that than oversell it.

AI does real work in three places. It extracts and normalizes citations from messy filed documents — varied Bluebook practice, short-form cites, OCR’d scans, inline string cites — which is genuinely hard with regex and genuinely easy with a modern model. It assesses whether a retrieved case actually supports the proposition it is cited for, which is the failure mode beyond pure fabrication and the one checkers handle worst. And it drafts the court-specific certificate language by reading the assigned judge’s standing order.

But the durable asset is not the model. It is the record — dated, hashed, attributed, retained. If you removed the AI, a worse version of this product would still exist and would still be worth money at renewal time, because the artifact is what the carrier and the court want. That is a weakness in the “AI-first” framing and a strength in the business: the moat such as it is comes from the record and the retention, not from a model anyone can rent.

7. Localization angle (if any)

N/A — this is a US-only play, and deliberately so. The product’s value is entirely a function of US court rules, US judicial standing orders, and US lawyers professional liability underwriting. The same pressure exists in England & Wales, Canada, and Australia (the Charlotin database is global), and the shape would transfer, but the rule corpus would have to be rebuilt from scratch per jurisdiction. Do the US first; it is large enough on its own.

8. Business model — path to $1M–$5M ARR

• Pricing: $39/mo solo, $99/mo for firms up to 5 attorneys, $199/mo up to 10. Annual billing discounted 2 months. Priced deliberately below the checkers, because this is not a research tool and must not be evaluated against one.
• ACV: ~$700 blended, assuming a mix skewed toward solo and 2–5.
• Rough math to $1M ARR: ~1,430 firms at $700 ACV. Against 60,000–90,000 addressable litigating small firms, that is roughly 2% penetration.
• Rough math to $5M ARR: ~7,000 firms, or ~8–10% penetration — which I do not think happens on self-serve alone. It happens if a malpractice carrier or a state bar endorses it, which is the real expansion lever (see risk flags: it is also the concentration risk).
• Expansion path: Seats as firms grow; per-matter retention tiers for firms with long-tail litigation; a paid “audit response” tier that assembles the full show-cause packet when an order actually lands. That last one is the highest-willingness-to-pay moment in the whole business and the worst possible thing to make the core of a subscription, so it stays an add-on.

9. Go-to-market wedge — first 100 customers

• Mine the sanctions database as a list. The Charlotin database is public and names cases, courts, and frequently the attorneys and firms involved. Every entry is a firm that just lived through this, plus — more usefully — a court whose bar now knows a colleague got sanctioned. Do not cold-pitch the sanctioned lawyer, that is tasteless and they are lawyering up. Pitch the local bar: “three attorneys in this district were sanctioned this year, here is a free 45-minute CLE on the new certification rules.” CLE is how you reach solo litigators, and they must earn credits anyway.
• Give away the judge-lookup, charge for the record. Build the free judge-by-judge certification lookup — the trackers exist but are PDFs and blog tables, not tools. A searchable “what does my judge require” utility is the SEO and the funnel, and it is the query a lawyer types the night before filing. The paywall sits at the attestation, not the lookup.
• Go through the malpractice brokers, not the carriers. Carriers move slowly; the independent LPL brokers who place solo and small-firm policies are motivated to hand clients something that improves the answer to the AI questionnaire before renewal. Fifty broker relationships is a tractable target for one person, and each one sits on hundreds of small firms with a known annual renewal date.
• State bar practice-management advisors. Most state bars run free practice-management assistance programs whose job is exactly this — telling solos how to not get disciplined. They publish checklists and newsletters and are actively looking for material on AI certification right now.
• Florida and New York first. Two states with dated statewide rules that just took effect, which means a defined population, a defined trigger date, and a natural pitch. Do not launch nationally; launch where the rule has a number.

10. Build complexity — justification

Low. Citation extraction, retrieval against CourtListener and free case-law APIs, PDF generation, hashing, and document storage are all off-the-shelf. The model work is prompt-level, not training-level. A competent pair ships a credible v1 in 8–10 weeks.

The genuinely non-trivial part is not engineering, it is corpus maintenance: keeping 145+ judge and court orders current as judges are appointed, reassigned, and amend their standing orders. That is a content operation, not a build, and it is ongoing forever. Budget for a part-time person reading court websites, or accept the corpus decaying into a liability — a tool that tells a lawyer the wrong thing about their judge’s requirements is worse than no tool.

11. Gating checklist

Gate	Pass?	Note
Legal in target market	✅	Documentation tooling. Explicitly not legal advice; ships with UPL-safe framing and attorney-signs-everything design.
Ethical — no harm / dark patterns	✅	Encourages verification and candor, both of which courts reward. No incentive to misrepresent.
Market exists (evidence above)	✅	1,598 sanction cases, 145 binding orders, carrier questionnaires with priced consequences.
1–5 person team can build this	✅	Off-the-shelf stack; the hard part is content maintenance, not engineering.
Launchable with <$50K / ₹40L	✅	Realistically $15–25K including initial corpus build and CLE production.

All five pass.

12. Feasibility score

Axis	Weight	Score	Notes
Problem intensity	20	15/20	Severe when it hits, but low frequency per firm. The annual carrier questionnaire is what makes it recurring rather than lottery-shaped. Not hair-on-fire weekly.
Demand evidence	15	13/15	Very strong regulatory, judicial, and actuarial evidence with dates and numbers. Docked for the absence of verbatim practitioner complaints — the demand is inferred from institutions, not heard from customers.
Build feasibility	15	13/15	Off-the-shelf v1 in 8–10 weeks. Docked because the rule corpus is a permanent content liability, not a one-time build.
Distribution clarity	15	12/15	CLE, LPL brokers, bar practice-management programs, and a free judge-lookup funnel are all concrete and named. Docked because solo attorneys are notoriously slow, skeptical buyers.
Revenue mechanics	15	11/15	Pricing is defensible and deliberately under the checkers. $1M needs ~2% penetration, which is fine; $5M needs a channel endorsement, which is not in my control.
Time to first revenue	10	8/10	Sellable at first CLE. Realistically 6–10 weeks to first paying firm.
Defensibility	10	2/10	Nearly none. Any checker vendor can bolt on an attestation log in a quarter. The only real moat is retention lock-in and corpus freshness, and both take a year to matter. This is the weakest axis by far and I am not going to dress it up.
Total	100	74/100

13. Qualitative modifiers

Founder-fit tags

technical-heavy · content-heavy

Technical for the extraction and retrieval pipeline; content-heavy because the rule corpus and the CLE distribution motion are both editorial products. A founder who will not write is the wrong founder here. Domain expertise helps but is not required — the primary sources are all public.

Key assumptions to validate (3–5)

1. Assumption: Solo litigators will pay for a record of diligence they currently do not keep at all, absent an active show-cause order. How to test: Run one CLE for 40 attorneys, present the attestation concept, and measure paid conversion within 14 days — not survey interest, actual cards.
2. Assumption: The malpractice questionnaire is a real purchase trigger, not just underwriting theater. How to test: Interview 15 LPL brokers. Ask directly whether a documented verification protocol has changed a quoted premium or endorsement for any client this cycle. If none can name a case, the annual trigger is fiction and this reverts to lottery-shaped demand.
3. Assumption: Existing checkers will not simply add an attestation log and close the gap. How to test: Read CiteCheck, Clearbrief, and Vincent release notes and roadmaps for 60 days. Also just ask them — vendors answer this question honestly more often than you would expect.
4. Assumption: The judge-by-judge corpus is maintainable by one part-time person. How to test: Build it for one federal district and one state system, then measure hours to keep it current over 6 weeks.

Risk flags

1. Defensibility (severe): Scored 2/10 and it is the real problem. The attestation log is a feature, not a company, until retention and corpus depth accumulate. A funded checker deciding to own this seam wins on distribution. The only credible answer is speed plus a channel — get to the LPL brokers before the checkers think of it.
2. Demand is institutional, not vocal: Every signal here comes from courts, carriers, and commentators. I could not find lawyers asking for this. Products bought because an institution requires them convert far better than products bought because someone theorized a need — but I have not proven the requirement is felt yet.
3. Regulatory whiplash: These rules are new and inconsistent, and some may be walked back or standardized. Notably, Florida and New York both declined to require AI-use disclosure and instead attached the duty to the signature. If a uniform federal rule lands and simply restates Rule 11 with no documentation requirement, the certificate-drafting half of the product loses its reason to exist. The attestation half survives, because the carriers are independent of the courts.
4. Corpus accuracy is a liability, not just a cost: Telling an attorney their judge requires nothing when the judge requires a certificate is a malpractice-adjacent failure. This needs conservative defaults — when unsure, say unsure and link the primary source.
5. Channel concentration: The path to $5M runs through a carrier or bar endorsement. That is one signature away from being someone else’s business decision.

14. Structured verdict

Score:                  74/100
Verdict:                GO
Confidence:             Medium
Best-fit builder:       Technical founder who will also write — legal-adjacent
                        background helpful, not required. Solo or pair.
Time to revenue:        6–10 weeks
Capital to launch:      $15–25K
Top 3 assumptions to validate first:
  1. Solos pay absent an active show-cause order — one CLE, measure paid
     conversion in 14 days, not survey interest
  2. LPL questionnaire moves premiums in practice — interview 15 brokers,
     require a named case where documentation changed a quote
  3. Incumbent checkers won't close the seam — read roadmaps for 60 days,
     and ask them directly
Kill criteria:
  - Abandon if <3% of 100 CLE attendees convert to paid within 30 days
  - Abandon if no LPL broker can name a client whose AI-verification
    documentation affected a quoted premium or endorsement
  - Abandon if CiteCheck, Clearbrief, or Vincent ships a per-filing dated
    attestation log before v1 launches

15. Next step — 1-week validation sprint

• Day 1–2: Build the free judge-lookup for one federal district and Florida state court — 20–30 orders, hand-curated from primary sources. This is the funnel asset and it doubles as the corpus-maintenance time trial from assumption 4.
• Day 3–4: Call 15 LPL brokers who place solo and small-firm policies. One question that matters: “Has a client’s documented AI verification protocol changed a premium or endorsement you quoted this cycle?” Record the count of brokers who can name a specific instance.
• Day 5: Pitch the attestation concept — with a mocked-up sample PDF, not a demo — to a Florida or New York solo-practitioner bar section, and offer a paid pre-order at $39/mo.

Falsifiable outcome: Go if ≥4 of 15 brokers name a real premium or endorsement consequence, and ≥3 attorneys put down a card at the pitch. If brokers cannot name instances and nobody pre-orders, the trigger is theoretical, the demand is commentary rather than pain, and this is a VALIDATE at best — shelve it and re-check when the first wave of post-rule renewals lands in 2027.