Skip to content
ParamergeParamerge
Stephen Lieberman
Through Paramerge
Contact

PAN Lab example

Learned Hand AI clerk pilot (LA and Riverside courts)

The only reviewer is the last authority: an AI clerk at the adjudication node

The largest trial court in the United States gave a handful of judges an AI clerk that reads the filings, researches the law, and drafts the order — in the judge's own writing style. Modeled on the Learned Hand LA courts pilot. What makes this different from every other copilot here is where it sits: at the adjudication node, the one point where the output becomes a binding court order. Every other tool in the collection has something downstream — an appeal, a supervisor, an audit, a caseworker who reads a cited answer first. This one does not. The only reported check is the judge's own review, and the judge is the last authority; no external audit, query logging, or benchmarking was reported, and the litigant whose case it touches is never told AI was involved. The draft arrives in the judge's own voice, so even a careful reader cannot tell which lines it wrote, and the backlog that made the tool worth adopting is exactly what eats the review time the whole safeguard depends on.

Stylized model of a documented deploymentCaseworker documentation & copilots

Open this example in PAN Lab v0.1 to apply pressures and levers and watch what the system does.

What this models

This example runs on the Adjudication-node-class judicial drafting copilot network: 6 components and 18 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 4 assumed · 3 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

  • assumed

    This models the adjudication-node judicial-copilot pattern documented in the Learned Hand LA courts case file — not a reconstruction of the actual tool or any court's deployment. It is deliberately distinct from the library's other copilots: the Magic-Notes-class documentation tool rests its whole safety case on one human review gate that drifts toward approval under load; the Nava-class and Benefit-Navigator-class tools are verify-before-use, with a caseworker reading a cited answer and an appeal and supervisor downstream; and the horizontal cross-government-copilot-class layer's finding is a measured productivity trade. The nearest sibling is the other generative-adjudication drafter in the library, the class where an AI drafts the ruling for a referee to sign: that shape keeps a present two-worker sign-off and a self-assessed governance committee, and its finding is a backlog mandate that rewards accepting the draft. This shape is its structural opposite: both the operator-side and model-side checks are drawn latent, the drafting officer is the terminal authority with no second officer or supervisor above the bench, no external audit, query logging, or benchmarking was reported, the affected party is never told, and style-imitation makes the AI's contribution forensically invisible — so a single copilot sits at the adjudication node with nothing lateral or downstream to catch what the one review misses.

  • baseline

    The two operator classes encode the two pilots' documented scope difference: the six Los Angeles civil-division judges and their research attorneys draft proposed orders that, once adopted, become tentative rulings and then binding court orders, while the seven Riverside civil and probate research attorneys use the tool for research memos rather than tentative rulings, so their output informs a deciding judge downstream. The deference channel is drawn strong into both, strongest on the adjudicative class where the draft arrives before an independent view forms; the anchoring concern itself is an attributed critique, not a measured effect.

  • baseline

    This shape's defining feature is an absence: both model-side and operator-side inhibiting checks are drawn empty at baseline. No independent output-level check gates an AI-drafted order against a separate source, and no second officer or supervisor reviews the draft before it becomes a tentative ruling, because no external audit, query logging, or benchmarking regime was reported for the pilot (analysis coverage drew a contrast with Michigan's stronger governance) and a sitting judge has no supervisor above the bench. The only reported safeguard is the drafting officer's own review, and per-output verification is exactly what the backlog that recruited the tool is consuming.

  • baseline

    The style-imitation self-loop is this shape's distinctive record dynamic: the copilot drafts in the individual judge's own writing style and reads the judge's prior rulings back from the orders record to do so, so future drafts are shaped by past AI-assisted orders, and a wrong line written once is hard to distinguish forensically from the judge's own voice. Because no disclosure edge exists to the affected party, an AI-induced error that survives review enters the binding record unmarked, and the litigant who would have standing to flag it is never told to look.

  • assumed

    The model's raw error rate is a modeling choice, not a measured per-interaction rate. No error rate, edit rate, override rate, or evaluation result exists for the tool in the public record; the vendor's Deep Verify per-sentence hyperlinking and multiple-verification-passes claims are unverified vendor statements; and the nearly 90 documented California AI-hallucination court cases since August 2024 concern litigant and attorney filings, not this tool, and are environmental base-rate context only.

  • assumed

    Peer pathways are authored on both signs: drafting-reliance habits and prompt shortcuts spread between chambers and their research attorneys, and one copilot drafting for all six judges in their imitated voices homogenises toward a common substrate so a systematic error repeats across chambers rather than staying idiosyncratic — a monoculture assumption in the Lab's qualitative vocabulary, not a measurement. The operator-side check is drawn latent because a terminal adjudication node has no second officer above it by default.

  • assumed

    Litigants and case outcomes are not in the dynamics. The people whose cases the tool touches are the affected party at the external boundary, and no ruling, order, or case result is computed from anything in this diagram; a draft, a tentative ruling, or an order is an institutional signal, never a decision about a person. Whether a specific case was decided with an AI draft is unconfirmed in the record (the Los Angeles court said testing occurs on already-decided motions, while the contracts permit live use and both courts declined to confirm), and this Lab neither asserts nor models that; it models institutional propagation only.

What this example does not show

  • No error rate, edit rate, override rate, or evaluation result exists for the tool in the public record. The pilot's planned quarterly evaluations are internal and unpublished, so any modeling of how much judges actually edit AI drafts rests on assumed parameters; the vendor's Deep Verify per-sentence hyperlinking and multiple-verification-passes claims, and its claimed adoption by other courts, are unverified vendor statements. The nearly 90 documented California AI-hallucination court cases since August 2024 concern litigant and attorney filings, not this tool, and are environmental base-rate context only.
  • The litigants whose cases the tool touches, and the outcomes of those cases, are not modeled here; the Lab models institutional propagation only, and a draft, a tentative ruling, or an order is an institutional signal, never a decision about a person. Whether any specific live case was decided with an AI draft is unconfirmed: the Los Angeles court said testing occurs on already-decided motions, while the contracts permit live use and both courts declined to confirm whether litigants are informed. That litigants are not told is documented (no disclosure obligation exists for partial assistance); that live cases are being decided with AI drafts is not established, and this scenario neither asserts nor depends on it.
  • The anchoring concern — that an AI-generated draft could influence a judge's position before an independent view forms — is an attributed critique (raised by the Los Angeles District Attorney and legal commentators), not a measured effect. Contested expansion uses discussed in the roadmap (criminal, family, and probate, including Racial Justice Act petitions) are carried as reported concern with both positions in the record, not as established harm.
  • The 49% figure is a one-year surge in Los Angeles employment-litigation filings only (4,100 to 6,400, per a February 2026 report), a subcategory rather than the whole docket, and is used here only as documented evidence that backlog pressure is real, not as a system-wide caseload rate.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

  • In February 2026 the Superior Court of Los Angeles County, the largest trial court in the United States, began a pilot of the Learned Hand AI drafting workbench with six civil-division judges and their research attorneys under a contract of about $314,000 running into early 2027, and the Superior Court of Riverside County gave seven civil and probate research attorneys access under a separate $10,000 agreement used for research memos; the tool ingests case filings, synthesizes applicable law, and drafts proposed orders in the individual judge's own writing style. Under California Judicial Council Rule 10.430 (effective September 1, 2025, the first statewide court generative-AI framework in the nation), disclosure is required only when a document consists entirely of generative-AI output, and the rule reaches judicial officers only for tasks outside their adjudicative role, so neither court is obligated to tell litigants when AI assisted with an order or memo in their case; both courts declined to confirm whether litigants whose cases are used in testing are informed.

    empirical
    • Investigative Mihalovich and Johnson, California judges are testing a new AI clerk, and you won't know if it's looking at your case (CalMatters, 2026) https://calmatters.org/economy/technology/2026/05/ai-los-angeles-riverside-courts/
    • Trade press Queally, Los Angeles Courts Pilot AI Tool to Help Judges Draft Rulings (Governing / Los Angeles Times via Tribune News Service, 2026) https://www.governing.com/artificial-intelligence/los-angeles-courts-pilot-ai-tool-to-help-judges-draft-rulings
    • Government Judicial Council of California, Rule 10.430, Generative artificial intelligence use policies (California Rules of Court, 2025) https://courts.ca.gov/cms/rules/index/ten/rule10_430

  • In the Learned Hand pilot the only reported error-correction safeguard is the judge's own review: officers are required to review and edit each draft before adopting a tentative ruling, and a court spokesman said the assistance does not supplant the judicial officer's independent role. No external audit, query logging, or benchmarking regime was reported (legal analysis coverage drew a contrast with Michigan's approach), and no error, edit, or override rate for the tool has been published. The Los Angeles District Attorney raised an anchoring concern, that an AI-generated draft could greatly influence what the judge's position should be before an independent view forms; this is an attributed critique rather than a measured effect, and the vendor's per-sentence Deep Verify hyperlinking and multiple-verification-passes claims are unverified vendor statements.

    empirical
    • Trade press Queally, Los Angeles Courts Pilot AI Tool to Help Judges Draft Rulings (Governing / Los Angeles Times via Tribune News Service, 2026) https://www.governing.com/artificial-intelligence/los-angeles-courts-pilot-ai-tool-to-help-judges-draft-rulings
    • Advocacy Howell, When Courts Adopt AI in the Dark: Privacy, Legitimacy, and the Democratic Stakes of Los Angeles's Learned Hand Experiment (The American Counsel, opinion and analysis, 2026) https://www.theamericancounsel.com/when-courts-adopt-ai-in-the-dark-privacy-legitimacy-and-the-democratic-stakes-of-los-angeless-learned-hand-experiment/
    • Investigative Mihalovich and Johnson, California judges are testing a new AI clerk, and you won't know if it's looking at your case (CalMatters, 2026) https://calmatters.org/economy/technology/2026/05/ai-los-angeles-riverside-courts/
    • Vendor Learned Hand and Superior Court of Los Angeles County, Learned Hand Announces Partnership With Superior Court of Los Angeles County to Explore Emerging Technology to Support Judicial Officers (Business Wire, 2026) https://www.businesswire.com/news/home/20260318640295/en/Learned-Hand-Announces-Partnership-With-Superior-Court-of-Los-Angeles-County-to-Explore-Emerging-Technology-to-Support-Judicial-Officers

Where this connects

Institutional pressures in this domain

  • Workload surge — Demand outruns staffing; per-case attention shrinks and review becomes triage.
  • Deadline pressure — Statutory or managerial timeliness rules reward fast approval of machine output over slow disagreement.
  • Reviewer bottleneck — One fixed-capacity checking stage sits between AI output and consequence; everything queues behind it.
  • Staff turnover — Experienced skepticism leaves; new staff calibrate their trust on the tool itself.
  • Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
  • Compliance over substance — Paper controls (sign-offs, checklists) satisfy audits while the behavior they describe erodes.

All of them in context on the Caseworker documentation & copilots domain page.

Levers available here and the patterns behind them

Documented case histories