United Behavioral Health's Level of Care Guidelines

What this models

This example runs on the Payer-authored-criteria-class behavioral coverage review network: 10 components and 24 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: 1 assumed · 8 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

• baseline

  D48-derived org (Phase 6, clinical-decision-support), re-derived in 2026 at the coarsest granularity that still draws every documented mechanism. REGISTER FIRST, because it governs every value here: there is no machine-learning system in this record and none is implied anywhere in this network. The governed object is a document — annual editions of codified admission, continued-stay and discharge criteria per level of care — applied as binding criteria by the company's own clinicians. The model node is the criteria-application step, not an inference engine, and every number below is a structural reading of an adjudicated record rather than a measured per-determination rate, because the record supplies no such rate.

• baseline

  TOPOLOGY. Ten nodes and twenty-four pathways, all documented, none decorative. An earlier drawing of this same record used eleven nodes and thirty-three pathways; it was re-drawn coarser in 2026, and every fact it carried is still on this page. Two operator classes, because the record separates them by AUTHORITY: care advocates collect the member's clinical information and may approve on clinical grounds, while a clinical denial may issue only from a peer reviewer, a physician or doctoral-level psychologist, at about thirty minutes a case, and the escalation between them is mandatory. ONE development-and-approval body, because the sentence the court adjudicated is about one process, developing and adopting the guidelines: drafting workgroups submitted each year's revisions (findings paragraphs 40 and 168) and the committee approved them, and the finding is that the company placed Finance and Affordability representatives in key roles in that development process throughout the class period instead of insulating its guideline developers from financial pressure. The earlier drawing split the workgroups from the committee and drew the same cost pressure arriving at each of them; both halves carry the same adjudicated finding, so they are drawn once, with the sign-off the body held drawn on it as the approval gate. Two further review authorities stay separate because each acted differently and each has its own adjudicated outcome: the insurance regulators of four states whose statutes required external criteria, and the appeal-and-litigation channel that finally landed. One inputSource because the external professional-society criteria are a real, named, legally required component of this deployment's governance rather than a modelling convenience: four state statutes required them, the vacated 2020 remedy named the catalogue outright, and California SB 855 later named the same associations by statute. What the coarser drawing folded, and where each fact now reads: the front line's adoption of the criteria and the approved edition binding both clinician classes are read on the two edition pathways; the thin trickle of case output to the approving table is read in that body's own description; the cost case reaching the drafting stage and the sign-off reaching back into what was proposed are read on the expense pathway into authorship and on the approval gate; the drafting write and the approval write are one write onto the edition; the clinical non-coverage determination written up with its cited guideline is read on the determination write, which carries that pathway's privacy flag; and the member's appeal, adjudicated futile, is read on the one route out of the company that it begins. Absences are derived too: no enforcement node (the adjudicated defect is upstream in authorship, and the denial is the determination's own written form rather than a separate downstream action system), no worklist (no queue or backlog appears in the record; the documented constraint is a per-case time budget, which is dynamics), no guardrail (no automated output screen exists), no retriever, and no externalBoundary (the record documents no egress of member information beyond the pipeline and its own appeal, regulator and litigation channels).

• baseline

  DEMAND 3 / CAPACITY 2. Demand 3 because the review load in this record is high by design: a 2014 internal presentation named continued use of concurrent review to ensure appropriate utilization as the mitigation strategy for the 2008 Parity Act's removal of day and visit limits, so the same admission is re-reviewed repeatedly against the criteria; benefit-expense forecasts and targets were tracked monthly with action taken where expenses exceeded projections; and the certified classes alone cover roughly 67,000 coverage requests over a six-year class period across a national commercial book. Capacity 2 rather than 1 because the clinical channel is real and credentialed — only a physician or doctoral-level psychologist may issue a clinical non-coverage determination, and the trial record puts about thirty minutes on the peer-to-peer call with the requesting physician plus the write-up, plus further time on the care advocate's file. Capacity 2 rather than 3 because that window plus a file read is the whole clinical check: there is no second clinical tier, and no independent evaluation of the deployment exists anywhere in the record. Neither number moved when the network was re-drawn coarser: both read off the deployment, not off how finely it is diagrammed.

• baseline

  BASELINES, and exactly how far the PAN org carries them. Fifteen of the twenty-four pathways here have a one-to-one counterpart in the PAN org's own list, and every one of them mirrors that counterpart's evidence-derived width on a single stated rung mapping recorded in this file's derivation comments, with no exceptions; no width was averaged, raised or lowered when two pathways were folded into one, because a fold re-homes a sentence and never a number. The other nine are derived directly from the court record cited beside each of them, because PAN's shape does not carry them: it has no edge kind for a check, so the five check pathways are not in that list at all; it records the drafting workgroups as a governance actor rather than as one of its four operator classes, so the prior-edition read into revision is not either; it carries no edge for what an operator puts INTO the determination, which is where the information-collection pathway and the peer-to-peer call live; and the one-document-every-determination self-loop is a property of the deployment rather than a coupling between two of its parts. Two check pathways are drawn deliberately away from where PAN's correction figure for the same class would put them, and each says why on its own line: the approval gate sits at the low rung rather than at zero because the body met, minuted and could decline; and the fiduciary review of the authorship sits below the highest rung because it reaches the criteria forward while the class-wide benefits claim ended in judgment for the company. The full per-pathway table, marked with which of the two bases each value comes from, is in the derivation record. Three contrasts are load-bearing and each is a finding drawn rather than a preference. First, the criteria pathway into the determination runs at the top rung while the member's own clinical information runs at the bottom one — that width asymmetry is the eighth declared deviation from generally accepted standards, mandatory prerequisites in place of a multidimensional assessment of the individual. Second, what feeds the table that writes the criteria — the expense forecasts, and last year's edition — runs wider than the gate that table could have used on itself: the court's finding is not that governance was missing but that it was present, constituted and infected. Third, the fiduciary channel's per-item effect runs a rung above its inbound reach, and the two are never averaged into one number.

• baseline

  THE LOOP THIS NETWORK EXISTS TO DRAW, and why it runs store to store rather than through any model. Cost pressure enters the criteria through the table that writes and approves them: benefit-expense forecasts and targets are what that table reads (paragraph 177), impact on benefit expense was modelled before a change was adopted and proposed liberalizations waited on a green light from finance (paragraphs 182 to 184), the resulting criteria bind every determination, denials suppress the benefit expense the targets measure, and the forecasts then validate the strategy the next annual edition inherits. Nothing in that loop is an inference engine; it is a document, a committee and a spreadsheet, which is exactly why this deployment belongs in a catalogue about governance rather than about accuracy.

• baseline

  THE TWO LATENT PATHWAYS ARE FINDINGS, NOT GAPS. The conformance check from the edition against the professional-society criteria is drawn at zero because the record adjudicates that it was not performed: three ASAM residential levels (3.1, 3.3 and 3.5) were omitted from the editions, the criteria mandates of Connecticut, Illinois, Rhode Island and Texas were each adjudicated violated, and the 2013 and 2015 crosswalks telling Connecticut regulators that all three levels were included are declared in the operative judgment to be statements the company knew to be false when it made them. The independent read of the criteria is drawn at zero because no independent scientific or actuarial evaluation of this deployment exists in the record; the outside measurement that exists was performed by a federal court through a ten-day bench trial, and PAN leaves its own independent-auditor slot deliberately empty for the same reason. Neither was touched when the network was re-drawn coarser: a documented absence is a finding, and a finding is never folded away to make a diagram smaller.

• baseline

  POSTURE DISCIPLINE, binding on every sentence in this network. This is an ERISA bench record: findings of fact and conclusions of law by a judge, no jury, no damages, no verdict. What stands adjudicated is breach of the fiduciary duties of loyalty and due care in DEVELOPING AND ADOPTING the guidelines, declared willful and systematic on 3 February 2026, and violations of four state criteria mandates that the company never appealed. What did NOT survive is carried with equal precision and appears nowhere on this diagram as a value: the class-wide wrongful-denial theory failed, the Ninth Circuit reversed the denial-of-benefits classes and held that neither ERISA nor the plans required the guidelines to be coextensive with generally accepted standards of care, a 2024 mandamus compelled judgment for the company on the benefits claim, and the roughly 67,000-request reprocessing ordered in November 2020 was stayed, held not to be appropriate equitable relief, and vacated in its entirety without ever taking effect. No claim was reprocessed. The ten-year injunction and the special master belong to that vacated order and never operated; the operative remedy is a permanent bar on the adjudicated editions for implementing plan terms about generally accepted standards plus a five-year accuracy mandate with jurisdiction retained through 3 February 2031.

• baseline

  EVIDENCE STATUS. Every deployment-specific fact on this diagram comes from a primary court document read in full: the 106-page redacted findings of fact and conclusions of law, the 99-page November 2020 remedies order (cited as history and for the alternative-criteria catalogue it named, never as operative relief), the Ninth Circuit's amended August 2023 opinion, the 57-page August 2025 remand order, and the seven-page operative amended remedies order of February 2026, with a Congressional Research Service analysis and the enacted text of California SB 855 alongside. Two sources are labelled where they are used: a plaintiff-aligned case tracker corroborates the timeline and carries the class-scale characterization of roughly 50,000 people of whom about half were children or adolescents, and a law firm's April 2021 note is a contemporaneous snapshot of the 2020 order and its stay that predates every reversal. Class-size arithmetic is stated the way the record supports it: roughly 67,000 coverage determinations for roughly 50,000 people, because a member can have more than one denial. The 2026 re-drawing read no new source and changed no fact: it is a coarser diagram of the same record.

• assumed

  Served members are not in the dynamics. No coverage decision, level of care, clinical judgement, health outcome or financial outcome for any person is computed from anything drawn here, and no score over any person is authored anywhere in this network. The class-size figures are recorded litigation facts, and the characterization that about half the affected people were children or adolescents is plaintiff-side reporting carried with that attribution. The structural equity fact of the case — behavioral-health coverage governed by criteria the court found more restrictive than generally accepted standards — is a finding about a benefit category, not a measured disparity between subpopulations of served people, and no such disaggregated measurement exists in the record read here.

What this example does not show

• LITIGATION POSTURE, verbatim from the evidence dossier and load-bearing: Wit v. UBH, N.D. Cal. 14-cv-02346-JCS (related 14-cv-05337): FFCL Feb. 28/Mar. 5, 2019; Remedies Order Nov. 3, 2020 (VACATED); 9th Cir.: Wit I memorandum Mar. 22, 2022 (withdrawn), Wit II opinion Jan. 26, 2023 (withdrawn), Wit III amended opinion Aug. 22, 2023, 79 F.4th 1068 (affirm standing + fiduciary classes; reverse benefits classes, GASC-coextensive reading, reprocessing; remand exhaustion; unappealed state-mandate rulings intact); Wit IV mandamus granted Sept. 4, 2024, No. 24-242 (judgment for UBH on benefits claim); dkt. 669 Aug. 5, 2025 (loyalty/care breach claim survives; no exhaustion required, alternatively futile); dkt. 695 Feb. 3, 2026 — OPERATIVE: declaratory judgment + permanent injunction on the tainted Guidelines + five-year GASC-accuracy mandate and retained jurisdiction through Feb. 3, 2031. Fee motion pending per mid-2026 trade reporting.
• The class-wide wrongful-denial theory did not survive, and no sentence in this scenario says otherwise. Judgment on the denial-of-benefits claim runs FOR the company; the Ninth Circuit held that neither ERISA nor the plans required the guidelines to be coextensive with generally accepted standards of care, and that reprocessing was not appropriate equitable relief. The surviving theory is process-based — disloyal and careless authorship of the criteria — which is exactly why this network draws the authorship rather than the determinations.
• REMEDY ARITHMETIC. The ten-year injunction, the special master, the court-specified external criteria and the roughly 67,000-request reprocessing all belong to the November 2020 order, which was stayed in February 2021, reversed in relevant part, and vacated in its entirety on 3 February 2026. It never took effect and no claim was reprocessed under it. The operative order imposes a permanent injunction against using the adjudicated editions to implement plan terms about generally accepted standards, plus a NEW five-year mandate through 3 February 2031 that any such criteria accurately reflect those standards and applicable state law, with jurisdiction retained. There is no special master and no reprocessing in the operative order.
• This is a bench record, not a jury one, and no damages were awarded: every liability statement here is a judge's finding of fact or conclusion of law under ERISA, and the relief is declaratory and injunctive. Attorney-fee litigation was reported ongoing in mid-2026 and appellate review of the 2026 remedies order remained possible as of this bundle's date, so the injunction posture should be re-verified before republication.
• There is no machine-learning system in this record. The criteria are codified decision rules applied by human care advocates and peer reviewers; the model node on this diagram is the criteria-application step. No accuracy figure, error rate or independent evaluation of this deployment exists — the values here are structural readings of an adjudicated record, and every one of them would be different for a deployment whose evidence said something else.
• Served members are not modeled. Coverage decisions, levels of care, clinical judgements and health outcomes are boundary quantities recorded in the case file; nothing on this diagram computes any of them. Class-scale figures are recorded litigation facts — roughly 67,000 coverage determinations for roughly 50,000 people, because a member can have more than one denial — and the characterization that about half were children or adolescents is plaintiff-side reporting, carried with that attribution. The record's structural inequity is between benefit categories, not between subpopulations of served people, and this network does not manufacture the second from the first.

Sources and evidence

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

• After a ten-day ERISA bench trial in October 2017, the United States District Court for the Northern District of California issued 106 pages of Findings of Fact and Conclusions of Law (28 February 2019; public redacted version 5 March 2019) holding that the 2011-2017 editions of United Behavioral Health's Level of Care Guidelines and Coverage Determination Guidelines were significantly and pervasively more restrictive than generally accepted standards of care. The operative 3 February 2026 judgment declares eight specific deviations, among them excessive emphasis on acute crisis stabilization, no effective treatment of co-occurring conditions, failure to err toward a higher level of care where the indicated level is ambiguous, no coverage to maintain function, motivation-based exclusions in the 2014-2017 editions, no child-and-adolescent-specific criteria, an overbroad custodial-care exclusion paired with a narrow active-treatment requirement, and mandatory prerequisites in place of a multidimensional assessment; the editions also omitted the ASAM residential levels 3.1, 3.3 and 3.5. The court found the criteria operated as binding rules rather than guidance: only a physician or doctoral-level psychologist could issue a clinical non-coverage determination, such a reviewer typically spent about thirty minutes talking to the requesting physician and writing up conclusions, every denial letter had to cite the specific guideline relied on, and the company's testimony that reviewers could deviate from the guidelines on clinical judgment was found not credible. No machine-learning system is involved; the guidelines are codified decision criteria applied by human reviewers.

  empirical
  • Government Wit v. United Behavioral Health, Findings of Fact and Conclusions of Law (redacted), N.D. Cal. No. 14-cv-02346-JCS, dkt. 418 (filed 5 March 2019; order dated 28 February 2019) https://storage.courtlistener.com/recap/gov.uscourts.cand.277588/gov.uscourts.cand.277588.418.0.pdf
  • Government Wit v. United Behavioral Health, Amended Remedies Order, N.D. Cal. dkt. 695 (3 February 2026) https://psych-appeal.com/wp-content/uploads/2026/07/2026-2-3-Amended-Remedies-Order.pdf

• The court's central finding is about who wrote the criteria and under what pressure. In the words of the Findings of Fact at paragraph 180: 'The Court finds that the financial incentives discussed above have, in fact, infected the Guideline development process. In particular, instead of insulating its Guideline developers from these financial pressures, UBH has placed representatives of its Finance and Affordability Departments in key roles in the Guidelines development process throughout the class period.' Named Finance and Affordability representatives sat as members of the approving committees — the Behavioral Policy and Analytics Committee from 2011 to 2016 and the successor Utilization Management Committee from 2016 — which reviewed and reissued the guidelines at least annually; proposed changes were modelled for their benefit-expense impact before adoption and proposed liberalizations are recorded awaiting a 'green light' from finance. The court also found that the company prepared detailed benefit-expense forecasts and targets, tracked monthly trends, and took action to address benefit expenses exceeding its projections, and that a 2014 internal presentation named continued use of concurrent review to ensure appropriate utilization as the 'Mitigation Strateg[y]' for the 2008 Parity Act's removal of day and visit limits. Because the guidelines were kept uniform across fully-insured plans, where the company bears the benefit-expense risk, and self-funded plans, where it does not, the court held the conflict tainted its decision-making as to both categories.

  empirical
  • Government Wit v. United Behavioral Health, Findings of Fact and Conclusions of Law (redacted), N.D. Cal. No. 14-cv-02346-JCS, dkt. 418 (filed 5 March 2019; order dated 28 February 2019) https://storage.courtlistener.com/recap/gov.uscourts.cand.277588/gov.uscourts.cand.277588.418.0.pdf
  • Government Wit v. United Behavioral Health, Amended Remedies Order, N.D. Cal. dkt. 695 (3 February 2026) https://psych-appeal.com/wp-content/uploads/2026/07/2026-2-3-Amended-Remedies-Order.pdf

• Four states required these determinations to use external professional-society criteria rather than payer-authored ones, and the court adjudicated violations of all four: Connecticut, where ASAM criteria have been required since 1 October 2013, violated throughout; Illinois, required from 18 August 2011, violated until 1 January 2016; Rhode Island, required from 10 July 2015, violated through the class period; and Texas, under the criteria of its Department of Insurance, violated throughout. The operative judgment also declares that the company's 2013 and 2015 crosswalks told Connecticut regulators that all three ASAM residential levels were included in its admission criteria and that, at the time these statements were made to Connecticut regulators, the company knew them to be false. United Behavioral Health did not appeal this portion of the judgment, and the Ninth Circuit recorded that it therefore remains intact — so the state-mandate rulings survived every appellate cycle unchanged. California later answered legislatively: SB 855 (Stats. 2020 ch. 151, effective 1 January 2021) requires commercial plans to make mental-health and substance-use medical-necessity determinations using the current criteria of the relevant nonprofit clinical specialty association, naming ASAM, LOCUS/CALOCUS, and CASII/ECSII, and forbids applying different, additional, conflicting, or more restrictive utilization review criteria.

  empirical
  • Government Wit v. United Behavioral Health, Amended Remedies Order, N.D. Cal. dkt. 695 (3 February 2026) https://psych-appeal.com/wp-content/uploads/2026/07/2026-2-3-Amended-Remedies-Order.pdf
  • Government Wit v. United Behavioral Health, 79 F.4th 1068 (9th Cir., amended opinion filed 22 August 2023) (Wit III), Nos. 20-17363/20-17364, 21-15193/21-15194 https://cdn.ca9.uscourts.gov/datastore/opinions/2023/08/22/20-17363.pdf
  • Government California SB 855 (Wiener), health coverage: mental health or substance use disorders (Stats. 2020, ch. 151; approved 25 September 2020; effective 1 January 2021) https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=201920200SB855

• The class-wide wrongful-denial theory did not survive. In its amended opinion of 22 August 2023 (Wit III, 79 F.4th 1068) the Ninth Circuit affirmed Article III standing and affirmed certification of the three classes for the fiduciary-duty claim, but reversed certification of the denial-of-benefits classes under the Rules Enabling Act; held that the district court erred to the extent it determined that the ERISA plans required the guidelines to be coextensive with generally accepted standards of care; held that reprocessing was not appropriate equitable relief under 29 U.S.C. section 1132(a)(3); and remanded the exhaustion question. When the district court's scope-of-remand order tried to preserve more than the mandate allowed, the Ninth Circuit granted mandamus on 4 September 2024 (No. 24-242) and directed entry of judgment for United Behavioral Health on the denial-of-benefits claim, writing that in its thorough analysis of the spirit of the mandate, the district court lost the letter. On remand the district court held on 5 August 2025 that the fiduciary claim survives Wit III insofar as it rests on the duties of loyalty and due care — entering judgment for the company on the duty-to-follow-plan-terms theory — and that the surviving statutory claim requires no administrative exhaustion, alternatively excused as futile on the trial findings. The roughly 67,000-request reprocessing ordered in November 2020 was stayed on 12 February 2021, held unavailable on appeal, and vacated; no coverage request was ever reprocessed under it. This is a bench record: findings of fact by a judge, no jury, and no damages, with relief declaratory and injunctive.

  empirical
  • Government Wit v. United Behavioral Health, 79 F.4th 1068 (9th Cir., amended opinion filed 22 August 2023) (Wit III), Nos. 20-17363/20-17364, 21-15193/21-15194 https://cdn.ca9.uscourts.gov/datastore/opinions/2023/08/22/20-17363.pdf
  • Government Wit v. United Behavioral Health, Order re Breach of Fiduciary Duty Claim, N.D. Cal. dkt. 669 (5 August 2025) https://storage.courtlistener.com/recap/gov.uscourts.cand.277588/gov.uscourts.cand.277588.669.0.pdf
  • Government Wit v. United Behavioral Health, Amended Remedies Order, N.D. Cal. dkt. 695 (3 February 2026) https://psych-appeal.com/wp-content/uploads/2026/07/2026-2-3-Amended-Remedies-Order.pdf
  • Government Wit v. United Behavioral Health, Remedies Order, N.D. Cal. dkt. 491 (3 November 2020) (VACATED in its entirety, 3 February 2026) https://storage.courtlistener.com/recap/gov.uscourts.cand.277588/gov.uscourts.cand.277588.491.0.pdf
  • Reference Manatt, Phelps & Phillips, LLP (2021, April 6). Updates from the Wit v. United Behavioral Health ERISA Class Action (Health Highlights) https://www.manatt.com/insights/newsletters/health-highlights/updates-from-the-wit-v-united-behavioral-health-er
  • Government Congressional Research Service, Rogers, H.-A. (2022, December 21). Behavioral Health Benefit Coverage and Wit v. United Behavioral Health (Legal Sidebar LSB10881) https://www.everycrsreport.com/reports/LSB10881.html

• The operative remedy is the Amended Remedies Order of 3 February 2026 (dkt. 695), which vacated the 3 November 2020 Remedies Order in its entirety and superseded it. It declares that the company's misconduct in developing and adopting the guidelines was willful and systematic, that the adjudicated editions are irreparably tainted by the company's disloyalty and lack of care, and that the company breached the ERISA duties of loyalty and care under 29 U.S.C. sections 1104(a)(1)(A) and (B) and violated the Connecticut, Illinois, Rhode Island, and Texas criteria mandates. It permanently enjoins use of those editions to implement plan terms about generally accepted standards of care, and orders that for five years — through 3 February 2031, with jurisdiction retained — any criteria the company adopts for that purpose shall accurately reflect those standards as established in the court's Findings of Fact and the requirements of any applicable state law. The vacated 2020 order's features are not part of it: there is no special master, no supervised retraining programme, no court-specified external criteria catalogue, and no reprocessing. The ten-year injunction belonged to that vacated order and never took effect; the 2031 endpoint belongs to the new five-year mandate. Attorney-fee litigation was reported ongoing in mid-2026 and appellate review of the 2026 order remained possible as of August 2026, so the injunction posture should be re-verified before republication. Class-scale figures are recorded litigation facts — roughly 67,000 coverage determinations for roughly 50,000 people, because a member can have more than one denial — and the characterization that about half of those people were children or adolescents is plaintiff-side reporting.

  empirical
  • Government Wit v. United Behavioral Health, Amended Remedies Order, N.D. Cal. dkt. 695 (3 February 2026) https://psych-appeal.com/wp-content/uploads/2026/07/2026-2-3-Amended-Remedies-Order.pdf
  • Government Wit v. United Behavioral Health, Remedies Order, N.D. Cal. dkt. 491 (3 November 2020) (VACATED in its entirety, 3 February 2026) https://storage.courtlistener.com/recap/gov.uscourts.cand.277588/gov.uscourts.cand.277588.491.0.pdf
  • Advocacy The Kennedy Forum. Wit v. United Behavioral Health case tracker (read August 2026) https://www.thekennedyforum.org/wit/

Where this connects

Institutional pressures in this domain

• Workload surge — Demand outruns staffing; per-case attention shrinks and review becomes triage.
• Reviewer bottleneck — One fixed-capacity checking stage sits between AI output and consequence; everything queues behind it.
• Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
• Data & policy drift — The world, the intake process, and the rules change under a system trained on how things used to be — two mechanisms with different remedies: the statistical properties of what the system processes move (concept drift), or the mixture of inputs arriving in deployment differs from the mixture it was trained on (covariate shift).
• Deadline pressure — Statutory or managerial timeliness rules reward fast approval of machine output over slow disagreement.

All of them in context on the Clinical decision support & deterioration alerting domain page.

Levers available here and the patterns behind them

• Upgrade model — Improve the model
• Keep prompts neutral — Framing and mirroring reduction
• Mark AI-written records — Provenance labeling
• Gate record entries — Human-in-the-loop write gating
• Pause AI on alarms — Deployment circuit-breaker
• Escalate checks — State-feedback vigilance
• Understand the system — Understand the system
• Assign a challenger — Structured dissent
• Check with a second model — Cross-model verification
• Check copied records — Reconcile copied records
• Review on schedule — Oversight cadence & retrospectives

Documented case histories

• UBH Level of Care Guidelines (Wit v. UBH)
• TREWS sepsis early-warning system
• Advance Alert Monitor (AAM) deterioration model
• Sepsis Watch deep-learning detection system
• Proprietary EHR sepsis model (external validation)
• nH Predict Utilization Review
• Cost-Proxy Care Stratification
• CA-CDS Child Abuse Alerting
• IDx-DR Autonomous Screening
• Viz.ai LVO Stroke Triage
• IBM Watson for Oncology
• OPTN eGFR Waiting-Time Correction
• Practice Fusion Pain CDS
• EviCore by Evernorth: the review threshold
• Cigna PxDx