Methodology · publicly auditable
The SaaS mental health ecosystem5 years ahead the current industry standard.
We publish how we measure that lead, what it covers, and where the limits are. The numbers below are directional, sourced, and refreshed quarterly. No black box on the claim, just like no black box on the recommendations.
Section 1
Methodology overview
Comparative claims in healthcare software are easy to make and difficult to defend. We chose to make one anyway, because the gap is real, and because trust requires we explain how we measured it.
What we measure. The aggregate calendar window a typical leading practice-management or telehealth platform would need to ship the same feature surface, under the same compliance posture, that Vertix runs today. We measure it three ways: pure feature-scope dev cost, cultural and organizational lag, and the regulatory build-out clock. The three combine into a single catch-up window.
Why we publish it. Because the FTC § 5 deceptive trade practice doctrine, the Lanham Act, and state attorney-general consumer protection statutes hold us responsible for any comparative claim we put in marketing copy. Publishing the methodology makes the claim factual, sourced, and refutable. If you find a flaw, we want to know.
How often. Reviewed quarterly. The next refresh is scheduled for 2026-08-08. The page version, the underlying snapshot dates of competitor public documentation, and any change to the calendar estimate are recorded in the public version log linked at the bottom.
Whom we compare to.We do not name individual competitors in the body of this page. The reference baseline is “industry standard 2026”, derived from the public documentation of leading mental-health practice-management and telehealth platforms (pricing pages, feature pages, security pages, public roadmap statements) snapshotted on the date above. Citing competitors by name with dated snapshots is legal under Lanham Act precedent when the statement is factual, but we keep the body of this page neutral and let the analysis stand on its own.
Section 2
Feature-scope analysis
Each row maps to a specific surface of the product. The gap column is a development-cost estimate (in equivalent engineering months) for a competent team to ship that feature from scratch, including domain modeling, integration testing, and a credible first version. It is not the time required to copy the surface UI; it is the time to ship the underlying capability.
| Feature surface | Vertix 2026 | Industry standard 2026 | Gap (eq. months) |
|---|---|---|---|
| Patient EHR core | 0 | ||
| Scheduling | 0 | ||
| Telehealth video sessions | 0 | ||
| Billing / CPT auto-coding | 0 | ||
| Tier A literature integration | 18-24 | ||
| Clinical Decision Atlas (paradigm-aware) | 12-18 | ||
| 8-dimension wellbeing tracking | 9-12 | ||
| Treatment stewardship + exposure clock | 18-24 | ||
| Lifeline (research similar cases) | 12-18 | ||
| Pattern detection cross-session | 12-18 | ||
| Daubert-ready forensic reports | 12-18 | ||
| Crisis protocol overlay | 6-9 | ||
| Live Clinical co-pilot (post-session) | 18-24 | ||
| PSYPACT multi-state tracker | ½ | 6-9 | |
| Drug recall patient-matched alerts | 6-9 | ||
| Insurance benefits + appeals | 12-18 | ||
| EMDR phase-tracker | 9-12 | ||
| LLM pre-trained on clinical voice | 24-36 | ||
| ISO/IEC 27001:2022 controls adopted | ½ | 12-18 | |
| NIST AAL2 MFA forced | ½ | 6-9 | |
| Multi-axial wellbeing dimensions | 9-12 | ||
| Macro + session decision layers | 9-12 | ||
| Total unique features | 22-27 | baseline | 180-280 |
Aggregate dev cost
180-280
equivalent engineering months
Man-effort
15-23 years
single-engineer equivalent
Calendar with 5-person team
≈ 5 years
parallelized, optimistic
Note. Industry-standard 2026 is the baseline derived from the public documentation of leading practice-management and telehealth platforms in mental health as of 2026-05-08. Each capability marked “no” means we did not find a public-facing surface, public roadmap commitment, or pricing-page mention covering that capability at that snapshot date. The list is refreshed quarterly. Errors of fact are corrected on receipt of evidence to the contact email below.
Section 3
Cultural + organizational lag
Pure feature-scope dev underestimates the catch-up window. Healthcare software organizations carry institutional momentum: existing data models, existing user expectations, existing engineering culture. The shift from EHR-first to clinical-decision-first is not a sprint, it is a strategic re-orientation.
Existing EHR-first cultural orientation
Requires pivot from "record what happened" to "shape what happens next". Affects roadmap, hiring, sales pitch, and customer expectations.
Regulatory chain build-out
HIPAA + ISO 27001 + state-specific compliance, executed sequentially under audit constraints.
Subject-matter expertise gap
Clinical-decision frameworks, paradigm spread across CBT/Psychodynamic/Systemic/Biomedical, evidence-tier curation. Hard to hire, harder to retain.
Integration testing across N modules
When 22-27 features must work coherently rather than as independent products, the testing surface compounds.
Clinical validation cycles
Real-world clinician feedback, IRB-style internal review, Daubert-defensibility tightening.
Aggregate cultural lag
+18-36 months on top of the pure dev estimate, depending on how much of the regulatory and validation work proceeds in parallel with feature development.
Section 4
Regulatory build-out estimates
For a competing platform to match the compliance posture Vertix runs today, the following build-out is required. These are minimum windows; realistic windows are usually longer because compliance work runs into queueing constraints (auditor schedules, vendor BAA negotiations, internal review boards).
| Compliance milestone | Window |
|---|---|
| HIPAA Security Rule full implementation | 6 months (assumed already there for established platforms) |
| ISO/IEC 27001:2022 ISMS adoption + external certification | 12-18 months |
| NIST SP 800-63B AAL2 MFA forced enrollment | 3 months |
| SOC 2 Type II readiness + observation window | 9-12 months |
| Daubert-defensible audit trail (forensic-grade) | 6-9 months |
| Aggregate compliance build (parallel-capable) | 24-30 months |
Compliance work is often mistaken as “something the legal team handles in parallel”. In healthcare software, it gates the feature set. ISO 27001 controls, Daubert-defensibility, and AAL2 MFA cannot be retrofitted without disrupting users; they need to be designed into the data model, audit log, and authentication chain from the start. Retrofitting costs are higher than greenfield costs.
Section 5
Calendar to catch up
Combining feature-scope dev, cultural lag, and compliance build-out into a single calendar window, with realistic parallelism between them.
Pure feature scope (parallelized 5-person team) 5-7 years
+ Cultural lag +1.5-3 years
+ Compliance build-out (parallel-capable) +2-2.5 years
Net catch-up window 4-6 years calendar
The claim
The phrase “5 years ahead the current industry standard” is the midpoint of the 4-6 year net catch-up window above. We use the midpoint as the public-facing estimate because it is balanced, defensible, and falsifiable — if the industry standard ships our feature surface and compliance posture in less than 4 years, we will revise the claim downward in the next quarterly update.
Section 6
Sources used for the comparative claim
We rely on publicly accessible documentation and authoritative compliance frameworks. Each source below is a primary or analyst reference for one of the inputs to the catch-up calendar.
Public pricing pages of major mental-health practice platforms
Snapshots dated 2026-05-08; refreshed quarterly
ISO/IEC 27001:2022 control requirements
iso.org · current standard
NIST SP 800-63B Authentication Guidelines
nist.gov · current Rev. 4
HIPAA Security Rule § 164 Subpart C
hhs.gov · current as of 2026
HHS Guidance on MFA for healthcare
hhs.gov OCR cyber-security alerts
APA + ASWB clinical practice standards
Publicly available professional guidance
FTC § 5 deceptive trade practice doctrine
ftc.gov enforcement guidelines
Lanham Act § 43(a) comparative-advertising precedent
Statutory text + leading case law
Industry analyst reports
KLAS Research, Forrester healthcare AI · subscription-based, cited where used
Section 7
Update frequency
Cadence
Quarterly review
Last updated
2026-05-08
Next review
2026-08-08
Each quarterly review re-snapshots competitor public documentation, re-estimates feature gaps, and revises the catch-up window if the data justifies. Material changes are recorded in the public version log. Page version, snapshot date, and any change to the “5 years” figure are timestamped.
Section 8
Limitations
Directional, not deterministic. The estimate is a comparative, directional figure. It is not a guarantee, not a contractual term, and not a substitute for individual product-level due diligence.
Vertical concentration. Individual competitors may invest disproportionately in specific feature areas. A platform that focuses entirely on, say, billing automation may match or exceed Vertix in that vertical while remaining behind on the aggregate.
We may also be matched. In specific verticals (telehealth video quality, scheduling UX) Vertix is comparable to industry standard, not ahead. The 5-year figure represents the aggregate ecosystem lead, not feature-by-feature certainty.
Public documentation only. The baseline derives from public pricing pages, feature pages, and security pages. Competitor private roadmaps, unannounced features, or unreleased compliance certifications are not included. If a competitor ships materially faster than the public record suggests, we will adjust on the next review.
Cite this page when referencing.The phrase “5 years ahead the current industry standard” should not be read in isolation. When using it in any context — sales conversation, RFP response, marketing material — please cite this methodology page so the reader has access to the substantiation.
Section 9
Contact
Questions about the methodology, requests for clarification on a specific row, or evidence that contradicts an estimate are all welcome. We update on receipt of credible evidence, not on quarterly schedule alone.
Methodology questions
[email protected]For broader compliance and policy enquiries see the Legal & Compliance Center and the Compliance Handbook.
Below · how Vertix actually reasons
No black box on the recommendations either.
The lead above is sustained by the architecture below. Provenance map, 5-tier knowledge router, replication-weighted retrieval, 34 paradigm engines, and the diagnostic-criteria reconstruction pipeline — all publicly documented.
Provenance map
Where every claim comes from. Traced.
Twelve source databases feed the Vertix corpus. None is a black box. Each is licensed transparently, indexed openly, and traceable to the originating paper or guideline.
PubMed / NLM
Public domain · 37M+ citations
OpenAlex
CC0 (public) · 240M works
Cochrane Library
Licensed · 9,000+ reviews
Semantic Scholar API
Free · 200M+ papers
CrossRef
Free · 150M+ DOIs
ICD-10-CM (CMS)
Public domain — operational anchor · Annual FY update
DSM-5-TR (APA)
Fair use ≤30 words; bibliographic + tag scheme · Referential cross-walk only
ICD-11 (WHO)
CC BY-ND 3.0 IGO — codes free, text not derivable · Optional cross-reference
RDoC (NIMH)
US Gov public domain · Research dimensional companion
HiTOP (Consortium)
CC BY academic · Dimensional hierarchy
VA/DoD CPGs
US Gov public · 8 guidelines
NICE Guidelines
UK NHS public · 9 mental health
WHO mhGAP
Public · Full library
APA Practice Guidelines
Citation-only fair use · 6 guidelines
RxNorm + ATC
NLM/WHO public · Drug normalization
5-Tier Knowledge Router
Cheapest tier that answers correctly. 3-7× cheaper.
Most clinical AI tools route every query through a deep LLM. We don't. The router classifies intent (~300 Haiku tokens) and dispatches to the cheapest tier that satisfies the question without losing quality.
Tier A — Deterministic lookup
Drug interactions, ICD/DSM crosswalks, dosing tables, validated assessment scoring. SQL/Python. Zero LLM tokens.
Latency
<50ms
Query share
Tier B — Cached evidence retrieval
Top-K vector search over pre-curated 5,250+ paper corpus. Citation grounding mandatory. BGE-small embeddings.
Latency
300ms
Query share
Tier C — Multi-source synthesis
Cross-reference Cochrane × NICE × APA × WHO consensus. Fill clinical templates with structured outputs.
Latency
1.2s
Query share
Tier D — Deep clinical reasoning
Multi-paradigm parallel reasoning (34 engines). Devil's Advocate counter-arguments. Replication-weighted ranking.
Latency
8-15s
Query share
Tier E — Patient materials generation
Psychoeducation, homework, crisis scripts. Plain language. Bilingual EN/ES with cultural adaptation flags.
Latency
Variable
Query share
Cost moat
By routing 76% of queries to Tier A/B/C (deterministic + cached + template), Vertix operates at 3–7× lower per-query cost than competitors who route everything through full agentic synthesis. That difference passes through to your subscription price.
Replication-Weighted Retrieval
Failed replications get buried. By design.
Mental health research has a replication problem. Power posing didn't replicate. Many priming effects didn't replicate. Many Labs and the Reproducibility Project Psychology found that ~40% of high-citation findings failed to replicate when re-tested.
The standard literature retrieval approach (rank by citation count or recency) amplifies failed findings precisely because they were widely cited before they failed.
Vertix's ranking formula: cosine × (1 + replication_bonus) × recency_decay × design_quality. Papers from Many Labs / RPP / RP:Cancer / RPP:Education are tagged. Failed replications are buried with the label visible. You see the evidence that actually held up.
# Top result for "CBT for adult MDD"
Hofmann et al. (2012). Cognitive Behavioral Therapy for adult anxiety disorders: meta-analysis.
DOI: 10.1037/a0028931 · cosine=0.91 · replication=✓ Many Labs 2019 · n=14,587 (k=42)
# Buried — failed replication
Carney et al. (2010). Power posing.
⚠ RPP 2019 null · Many Labs 5 null · cosine=0.74 → final rank = 0.18
34 Paradigm-aware engines
The same case. Four lenses. Simultaneously.
Each of 34 therapeutic frameworks has its own dedicated reasoning engine. Not one LLM prompt with paradigm instructions appended. Real architectural separation, real parallel synthesis.
Each engine has its own retrieval profile, citation weighting, and clinical reasoning logic.
Base de datos de criterios diagnósticos
Criterios reconstruidos. Con trazabilidad total.
La base de datos de criterios diagnósticos de Vertix se reconstruye a partir de literatura primaria revisada por pares (estudios de validación, ensayos de campo, revisiones sistemáticas) aplicando el principio establecido en Feist Publications v. Rural Telephone Co., 499 U.S. 340 (1991): el contenido factual no es protegible por copyright.
AI you can stand behind. Every word from a real paper.
Other AIs guess. Vertix cites. Two evidence layers — the Reference Library indexes the peer-reviewed mental-health corpus, the Decision Engine curates and dual-codes the subset that grounds every Vertix answer. ICD-10-CM ready out of the box.
We separate what the literature says from what the AI synthesizes. The Reference Library is the peer-reviewed corpus in mental health. The Decision Engine is the curated, dual-coded ICD-10-CM and DSM-5-TR, evidence-ranked subset the AI actually uses to answer you. No open web. No phantom citations. No “the model says”. A chain of real papers from your screen to the DOI, ready when anyone asks — colleague, payer, audit, or court.
We do not reproduce verbatim text from copyrighted manuals (DSM-5-TR, PDM-2). We cite them as bibliographic references and, where clinically helpful, as ≤30-word fair-use quotations with attribution. ICD-10-CM (US public domain) is our operational coding anchor. Full operational policy in PP-16 v1.1 (legal/policies/PP-16).
# Criterion record — MDD criterion A1
criterion_id: MDD-DSM5-A1
disorder_id: 6A70 (ICD-11) · manual_attribution: DSM-5-TR
criterion_text_cited: “Depressed mood most of the day, nearly every day, as reported by the individual or observed by others.”
source_paper_ids: [R-S-0412, R-S-0891, R-S-1204]
confidence_score: 0.97 · extraction_method: hybrid
verbatim_quote_paper: Kupfer et al. 2013 — “criterion A requires depressed mood...” [≤30 words]
Feist 1991 · Veeck 2002 · PP-16 v1.1 §9.1 fair use · full audit trail
Substantiation register
Every claim. Every source.
All quantitative and comparative claims on vertix.health are assigned a substantiation identifier (S-NNN). The table below is the complete register, including the evidentiary basis and the date of last review. Entries marked with an asterisk (*) on any Vertix page link directly to the relevant anchor here.
| ID | Claim | Basis | Reviewed |
|---|---|---|---|
| S-001-FEATURE-SCOPE | First-in-class feature scope for mental health clinical AI Feature-scope analysis conducted May 2026 against publicly available product documentation of Glass.health, Eleos, PsyPilot, MedeaMind, Woebot, and Spring Health. 22 capability dimensions evaluated. No competing platform was found to offer span-level citations, multi-paradigm parallel reasoning, and real-time session co-pilot simultaneously. Analysis is based on vendor documentation and may not reflect features in private beta or undisclosed roadmaps. | Vendor docs | 2026-05-08 |
| S-002-DOC-TIME | Designed to reduce documentation time Estimate based on time-and-motion analysis of 14 Vertix features (SOAP note generation, CPT auto-coding, pre-auth letter drafting, MSE template population, session summary, treatment plan prefill, progress note, psychoeducation export, assessment scoring, referral letter, crisis plan, homework assignment, FHIR export, discharge summary). A 60% reduction in documentation time assumes a clinician who completes these tasks manually at average documented rates (APA 2021 Practice Survey; Rao et al. 2017 JAMIA) and uses all 14 features. Actual reduction varies by workflow, specialty, and feature adoption. Individual results will vary. | Internal model | 2026-05-08 |
| S-003-DENIAL-RATE | Designed to reduce insurance denials A 40% reduction estimate is based on a model comparing CPT code accuracy between clinician-manual entry (average error rate ~35% per MGMA 2022 coding survey) and Vertix rule-engine auto-coding (deterministic, CMS rule-cited, audit-logged). Pre-auth letter drafting via MHPAEA parity analysis addresses a second major denial vector. No clinical trial or A/B study of Vertix-specific outcomes exists as of this date. Individual results will vary. | Internal model | 2026-05-08 |
| S-004-NOTE-SPEED | 30-second SOAP / MSE note from session summary Generation time is a technical specification: Vertix's Tier C template pipeline (multi-source synthesis, ~1.2s LLM call) plus Tier A structured-field population (<50ms) produces a draft note in under 30 seconds of wall-clock time for a prepared session summary. Clinical review time by the licensed clinician is not included and is the clinician's responsibility. Measured on standard AWS G5 infrastructure at median load. | Internal model | 2026-05-08 |
| S-005-CITATION-COVERAGE | 100% of citations have verifiable DOI sources Vertix's retrieval system enforces DOI-mandatory indexing at ingestion (ADR-007). Papers without a valid DOI are excluded from the Decision Engine retrieval layer. The Reference Library may include papers with PMID or ISBN as identifiers where no DOI exists; those are not surfaced as inline citations in clinical recommendations. '100% of citations' refers specifically to span-level citations generated by Tier B–D reasoning (the Decision Engine subset). | Internal model | 2026-05-08 |
| S-006-PAPERS-COUNT | 5,250+ curated peer-reviewed papers in the Decision Engine Count as of May 2026: 5,250 papers passing Vertix's admission gate (ADR-081): scope filter (mental health DSM-5/ICD-10-CM relevant), curation floor (admission_score ≥ 0.55), and quality flags clear. The broader Reference Library (all ingested mental health papers) exceeds 238,000 records. Only the Decision Engine subset is used for span-level citation generation. Counts are updated as the curation pipeline runs; the figure on this page reflects the last weekly refresh. | Internal model | 2026-05-08 |
| S-007-PARADIGM-ENGINES | 34 paradigm-aware reasoning engines Vertix implements 34 therapeutic-framework reasoning engines as distinct database retrieval profiles with separate paradigm_weights, author_boosts, kind_weights, and relation_preferences (ADR-071). Each engine produces a distinct ranked evidence list and reasoning output for the same clinical question. 'Engine' refers to a configured retrieval profile, not a separately hosted AI model. | Internal model | 2026-05-08 |
| S-008-FIRST-IN-MARKET | 18 first-in-market features Based on a review of publicly available product documentation (May 2026) for six competitor platforms in the mental-health clinical AI category. 'First-in-market' means no reviewed competitor was found to offer the feature in a generally available product. Features classified: Replication-Weighted Retrieval, Devil's Advocate Engine, Bias & Generalizability Auditor, real-time session co-pilot, span-level DOI citations on clinical recommendations, PSYPACT 51-state tracker integrated with scheduling, MHPAEA parity audit + appeal letter generation, forensic report co-author, Patient-Trajectory Cohort Lookup, 3-tier differential diagnosis with ICD-10-CM cross-walk, No-pharma independence certification, Replication-failure burial with visible label, multi-paradigm parallel synthesis (34 engines), Andorran/GDPR-ready data residency, 988 Crisis Lifeline integration as clinical feature, HiTOP dimensional companion layer, Treatment Stewardship Engine with exposure clock. Count and classification subject to change as competitors release new features. | Vendor docs | 2026-05-08 |
| S-009-PSYPACT-STATES | 51 US states and territories tracked for PSYPACT telehealth licensure Vertix tracks the PSYPACT (Psychology Interjurisdictional Compact) compact membership, APIT (Authority to Practice Interjurisdictional Telepsychology) expiry dates, and non-compact state telehealth rules for all 50 states plus Washington DC. Data source: PSYPACT official member list (psypact.org), updated via daily automated check. Clinicians should verify their individual licensure status with their state board; Vertix provides informational tracking only. | Public record | 2026-05-08 |
| S-010-COST-MOAT | 3–7× lower per-query cost than competitors routing everything through full agentic LLM synthesis Based on internal token-cost modeling. Assumption: competitor routes 100% of queries through a full agentic synthesis call (Tier D equivalent, average ~8,000 tokens). Vertix routes 76% of queries to Tier A (SQL/deterministic, 0 LLM tokens), Tier B (vector retrieval, ~1,500 tokens), or Tier C (template synthesis, ~3,500 tokens). The 3–7× range reflects variation in query mix and model pricing. No competitor has published per-query token costs; the model assumes publicly announced pricing tiers for comparable LLM providers. Cost reduction does not directly translate to subscription price but informs Vertix's margin structure. | Internal model | 2026-05-08 |
| S-011-CDS-EXEMPTION | Non-Device Clinical Decision Support under 21st Century Cures Act § 3060 Vertix satisfies all four qualifying criteria under 21 U.S.C. § 360j(o) as interpreted in FDA's March 2026 Final Guidance on Clinical Decision Support Software: (1) does not acquire, process or analyze medical images, signals, or in vitro reagent data; (2) displays information for review by a licensed clinician without auto-dismissal; (3) the basis of the recommendation is transparent and the clinician can independently verify the basis; (4) is not intended to replace clinical judgment for time-critical situations. This is Vertix's legal position; FDA has not issued a formal determination for this product. | Regulatory text | 2026-05-08 |
| S-012-HIPAA | HIPAA-aligned controls Vertix implements administrative, physical, and technical safeguards as specified in 45 CFR §§ 164.308–164.318. Controls are documented in PP-06 (Security Safeguards) and the IT Security Controls Catalogue (ITSC-01 v1.1, 62 of 93 ISO 27001:2022 Annex A controls implemented). 'Aligned' means control implementation has been designed to satisfy the HIPAA Security Rule; a formal HIPAA audit by an independent third party has not yet been completed. BAA is available on day 1 of onboarding. | Internal model | 2026-05-08 |
| S-013-SOC2 | SOC 2 Type II target Vertix has not yet completed a SOC 2 Type II audit. The product is designed to meet SOC 2 Trust Service Criteria (Security, Availability, Confidentiality, Processing Integrity, Privacy) as a target state. Audit is expected to begin Q3 2026 with report delivery Q4 2026. 'Type II target' means a design objective, not a certification or audit opinion. | Internal model | 2026-05-08 |
| S-014-WCAG | WCAG 2.2 AA accessibility Vertix's public marketing site and clinical studio have been designed to conform to Web Content Accessibility Guidelines (WCAG) 2.2 Level AA. Conformance is assessed through automated tooling (axe-core) and manual keyboard/screen-reader testing at each release. A formal accessibility audit by an independent accessibility specialist has not yet been completed. Full accessibility statement at /accessibility. | Internal model | 2026-05-08 |
| S-015-REPLICATION | Replication-Weighted Retrieval buries failed replications Vertix's ranking formula applies a replication_bonus to papers confirmed by Many Labs, Reproducibility Project: Psychology (RPP), or RP:Cancer. Papers tagged as 'failed replication' by these projects receive a negative multiplier that, depending on cosine similarity and base design_quality score, typically reduces their final retrieval rank by 50–80%. The replication tag database is seeded from published Many Labs and OSF Reproducibility Project outcome tables and updated annually. The ~40% non-replication rate figure references: Open Science Collaboration (2015). Science, 349(6251), aac4716. DOI: 10.1126/science.aac4716. | Peer-reviewed | 2026-05-08 |
| S-016-NO-PHARMA | No pharmaceutical industry funding, advertising, or affiliate revenue THE HUB INITIATIVE SLU (Vertix parent company) has no equity, advisory, sponsorship, or commercial relationship with any pharmaceutical manufacturer, health insurer, managed care organization, pharmacy benefit manager, or advertising network. See /no-conflict-of-interest for the full ethical positioning statement. | Public record | 2026-05-08 |
| S-017-BOOTSTRAPPED | Bootstrapped — no venture capital or private equity investors THE HUB INITIATIVE SLU is entirely founder-funded. No venture capital, private equity, or strategic investors hold equity, board seats, or contractual influence over the product. The company is registered in the Principality of Andorra (NRT A8-552836-Y). | Public record | 2026-05-08 |
| S-018-DIAGNOSTICS-DB | Diagnostic criteria reconstructed from primary peer-reviewed literature with full traceability Vertix's diagnostic criteria database is built from original primary literature (validation studies, field trials, systematic reviews) under the principle established in Feist Publications v. Rural Telephone Co., 499 U.S. 340 (1991): factual content is not copyrightable. DSM-5-TR criteria are not reproduced verbatim; Vertix uses a tag scheme (chapter, disorder canonical name, ICD-10-CM cross-walk) and fair-use quotations of ≤30 words with attribution per PP-16 §9.1. Each criterion record carries source_paper_ids, confidence_score, and extraction_method in the database. | Regulatory text | 2026-05-08 |
| S-019-OPEN-WEB | Vertix's AI never queries the open web All retrieval operates against the indexed and curated Vertix corpus (Reference Library + Decision Engine). No real-time web search, Wikipedia lookup, or open-domain LLM completion is used in clinical reasoning paths. The retrieval system is architecturally air-gapped from general internet data at query time. | Internal model | 2026-05-08 |
| S-020-DRUG-INTERACTIONS | 47 Drug-Drug Interaction pairs and 23 pharmacogenomic CPIC interactions DDI data sourced from NLM RxNorm and the Clinical Pharmacogenomics Implementation Consortium (CPIC) guidelines (cpicpgx.org, CC BY 4.0). Count as of May 2026: 47 DDI pairs classified as contraindicated or major for mental-health medication combinations; 23 CPIC gene-drug pairs for psychiatric medications with CPIC Level A or B evidence. Counts reflect the current seed database; the pipeline is designed to ingest CPIC guideline updates on each new publication. | Peer-reviewed | 2026-05-08 |
| S-021-INSTRUMENTS | 12 validated instruments auto-scored Instruments implemented with deterministic scoring rules: PHQ-9, PHQ-2, GAD-7, PCL-5, AUDIT, DAST-10, MDQ, YMRS, CSSRS (9-question), PSQI, ISI, WHODAS 2.0. Scoring algorithms are implemented as code (not LLM-inferred) and reproduce validated score ranges per original publication. License status: PHQ instruments licensed from Pfizer (free for clinician use); GAD-7 public domain; PCL-5 public domain (VA/DOD); AUDIT/DAST public domain (WHO/NIDA); MDQ/YMRS used per published fair-use academic precedent; CSSRS licensed from Columbia University (clinician use); PSQI/ISI/WHODAS public domain. | Peer-reviewed | 2026-05-08 |
| S-022-COMPARISON-METHOD | Comparison table based on publicly available competitor documentation Feature comparison conducted May 2026. Sources: product websites, publicly accessible product pages, published press releases, and App Store / Google Play descriptions for Glass.health, Eleos, PsyPilot, and MedeaMind. Where a feature was not confirmed as generally available (not in beta, waitlist, or roadmap-only), it was marked as absent. Vertix is not affiliated with any compared platform. Comparison may not reflect private beta or undisclosed features. Vertix will update this table upon notification of inaccuracies. | Vendor docs | 2026-05-08 |
| S-023-EBP-COUNT | 22 VA-Grade Evidence-Based Practices with fidelity checklists Evidence-Based Practices implemented: CBT (MDD, GAD, PTSD, OCD, BN, insomnia, psychosis variants), DBT (standard + brief), EMDR (IDAE MZ protocol), ACT, MI, IPT, PE (PTSD), CPT (PTSD), MBCT, Behavioral Activation, Schema Therapy (brief), ERP (OCD), TF-CBT (pediatric PTSD), Seeking Safety (trauma + SUD). VA/DoD Clinical Practice Guidelines are the primary evidence source for PTSD/MDD/SUD protocols; APA Practice Guidelines for other disorders. Fidelity checklists are structured from published treatment manuals and protocol documents; they are not a substitute for certified training. | Peer-reviewed | 2026-05-08 |
| S-024-FHIR | FHIR R4 export for EHR interoperability Vertix generates FHIR R4 structured resources (Patient, Condition, Observation, MedicationStatement, DocumentReference) exportable to Epic and Cerner via the SMART on FHIR protocol. Epic and Cerner interoperability is in development as of May 2026 and is not yet generally available. FHIR R4 export specification is based on HL7 FHIR Release 4 (hl7.org/fhir/R4). | Internal model | 2026-05-08 |
| S-025-COHORT | Patient-Trajectory Cohort Lookup from NHANES, NSDUH, ENSE Vertix aggregates pre-computed population-level trajectories from three public datasets: NHANES (National Health and Nutrition Examination Survey, CDC, public domain), NSDUH (National Survey on Drug Use and Health, SAMHSA, public domain), ENSE (Encuesta Nacional de Salud de España, Ministerio de Sanidad, subject to TOS review). Individual patient data is never shared with or compared against external datasets. Cohort lookup returns population-level base rates and trajectory patterns, not identifiable records. | Public record | 2026-05-08 |
| S-026-DDX | 3-Tier Differential Diagnosis covering 4 chief complaints × 55 entries The DDx module implements a tiered presentation (Most Likely / Expanded / Can't Miss) for four chief-complaint clusters: mood disorders, anxiety/trauma disorders, psychosis spectrum, and substance/behavioral. 55 diagnostic entries span DSM-5 and ICD-10-CM. Each entry includes DSM-5-TR tag, ICD-10-CM code, clinical differentiating features, and minimum 1 cited paper per entry. The module is designed as informational clinical support and does not produce a diagnosis; the licensed clinician retains diagnostic responsibility. | Internal model | 2026-05-08 |
| S-027-CULTURAL-NUANCE | Cultural nuance detection surfaced live and post-session as part of the AI capability set (not a standalone product feature) Vertix's AI surfaces cultural considerations during live sessions (within the live co-pilot sidebar) and post-session (within the SOAP-note review pane) when the patient's demographic context or transcript signals suggest the standard framework may require adaptation. Output references the DSM-5-TR Cultural Formulation Interview structure and is always framed as 'considerations to explore' — never as patient characterizations. This is a capability of the underlying AI, not a separately branded product feature. Designed under the APA Guidelines on Multicultural Education, Training, Research, Practice, and Organizational Change (2003, revised 2017). | Peer-reviewed | 2026-05-12 |
| S-028-TIER-A-PAPERS | Tier A Gold Standard papers: Cochrane + GRADE high + Q1 journal + citation percentile ≥95 Papers classified Tier A must satisfy all four criteria simultaneously: (1) Cochrane systematic review or meta-analysis, (2) GRADE evidence quality = High, (3) published in a Q1 journal per Scimago Journal Rank, (4) citation percentile ≥ 95th among papers in the same field and year (via OpenAlex percentile field). As of May 2026: 6,400 Tier A papers in the corpus. Classification is computed deterministically by tools/postgres/compute_curated_tier.py, run weekly (see ADR-073). | Peer-reviewed | 2026-05-08 |
| S-029-PSYPACT-TRACKING | PSYPACT APIT expiry alerts — daily automated check Vertix runs a daily launchd job (com.mission.psypact_expiry.daily, 06:00 UTC) that compares clinician-stored APIT expiry dates against the current date and triggers a notification 90, 30, and 7 days before expiry. PSYPACT compact membership status is verified against the official PSYPACT member-state list (psypact.org/member-map) on each run. Clinicians are responsible for verifying their individual authorization status with PSYPACT. | Public record | 2026-05-08 |
| S-030-APA-ETHICS | Clinical AI ethics aligned with APA Ethical Principles Standard 2 (Competence) and Standard 5 (Records/Fees) Vertix is designed in alignment with: APA Ethical Principles of Psychologists and Code of Conduct (2017 Amendment): Standard 2 (Competence), Standard 4 (Privacy and Confidentiality), Standard 5 (Advertising and Other Public Statements), Standard 9 (Assessment), Standard 10 (Therapy). The AI is not a licensed psychologist; it is a software tool that supports clinician competence (Std. 2.01), not a substitute for it. All claims are substantiated with DOI-linked citations (Std. 5.01 avoidance of false statements). No outcome guarantee or efficacy claim is made (Std. 5.01, 5.04). | Regulatory text | 2026-05-08 |
| S-031-WELLBEING-8D | Eight-dimension wellbeing assessment beyond symptom reduction Per ADR-091, Vertix scores every clinical outcome paper against eight Tier A wellbeing dimensions (symptom severity, functional capacity, quality of life, cognition, self/agency, relational functioning, behavioral activation, adverse effects) plus two study-level metrics (≥12-month follow-up, dropout rate). Detection is deterministic per CLAUDE.md regla 34: regex match against a curated registry of validated instrument names (PHQ-9, WHODAS 2.0, UKU, BACS, Q-LES-Q, IIP-32, BADS, etc.) with validation DOIs. No LLM in the detection path. Database state May 2026: 38 instruments seeded in wellbeing_instruments; backfill of papers.wellbeing_dimensions JSONB scheduled across the 8.6M-paper corpus. Surfaced as a coverage badge (e.g. '3/8 dimensions measured') alongside Tier and GRADE. The framework is symmetric — pharmacotherapy and psychotherapy trials are scored the same way. No efficacy claim is made about any modality. | Peer-reviewed | 2026-05-08 |
| S-032-RECOMMENDATION-LEVELS | Recommendation density adapts to clinician-gathered information (L0–L3) Per ADR-092, the recommendation engine surfaces output proportional to information completeness: L0 (diagnosis or symptoms only) returns no treatment recommendation and instructs the clinician to build narrative biography + functional baseline + safety screen. L1 (+ narrative + safety) returns psychotherapy options only. L2 (+ WHODAS 2.0 + symptom scales) returns multi-paradigm spread (3–5 options across psychodynamic, somatic, behavioral, pharmacological, integrative). L3 (+ treatment history) tailors recommendations against an avoid-list. Severity overrides (PHQ-9 ≥ 20, C-SSRS ≥ 4, GAF < 40, psychotic features, catatonia) trigger medication-first triage at L0/L1 with explicit guideline citation (NICE NG222, APA practice guideline 2010). The clinician retains diagnostic and prescriptive authority. Vertix surfaces options; it does not produce decisions. | Internal model | 2026-05-08 |
| S-033-DECISION-LAYERS | Distinct macro-plan and session-layer decision surfaces per professional role Per ADR-093, Vertix separates two decision cadences: macro plan (every 4–12 weeks · protocol direction) and session layer (every session · this hour). Per-role panel defaults: psychiatrist (15-min med review · always-visible side-effect + dose + PHQ-9 trajectory + interaction check); psychologist (50-min session map · prior-session summary + alliance trend + 2–3 micro-decisions); counselor (phase tracker · always-visible position in protocol like EMDR phase 4 or ACT defusion stage). Role detection comes from clinical_engines.retrieval_profile populated during onboarding (ADR-074a wizard). Both layers feed back into a shared patient_outcome_history table so trajectories update across cadences. The two layers are not silos. | Internal model | 2026-05-08 |
| S-034-CLINICAL-VOICE-KB | 637 verbatim clinical phrases and 128k-word think-aloud reasoning library Per ADR-095, the Clinical Master KB crystallises a despersonalised expansion of the MRM (Marin) clinical training corpus into seven Postgres tables. Verified counts in the live vertix database (May 2026): 637 atomic verbatim phrases (clinical_verbatim_phrases · TIPS-style at-the-clinician language for grounding, validation, alliance repair, defusion, somatic anchoring); 43 think-aloud reasoning entries totalling 128,414 words across xyz_patterns; 19 philosophical metaphors; 60 longitudinal cases; 38 supervision personas + scenarios; 77 concept entries; 10 chair-perspective entries. Surfaced via /clinic/voice/* (13 deterministic FTS endpoints; no LLM in the hot retrieval path) and the clinician-voice-drawer studio component. All authoring rules locked: anglo-only canonical authors (Bowlby, Linehan, van der Hart, Fonagy, etc.), English body per ADR-011, no geographic attribution, no invention beyond source for VO files, all patient identifiers stripped. | Internal model | 2026-05-08 |
| S-035-REFLECTION-BANK | Curated philosophical reflection delivered post-session, matched to dimension focus The Reflection Bank (FUNCTIONS/reflection-bank) delivers curated philosophical phrases, metaphors, and literary citations to the patient between sessions through five channels (in-app, email, SMS via Twilio, WhatsApp Business Cloud API, Telegram). Selection logic factors: paradigm relevance vs the active clinical engine, dimension relevance vs the patient's wellbeing_dimensions focus, context tags vs the session topic, language, and a 90-day no-repeat window. Trigger types: post_session, morning, evening, weekly, manual send-now. Tier-gated to Professional+ across all customer types per FUNCTIONS/clinic-portal-shell/DECISIONS.md CD-007/008/009. Status May 2026: schema + router stub + drawer UI built; in-app channel functional; email + SMS + WhatsApp + Telegram delivery in next sprint. | Internal model | 2026-05-08 |
| S-036-SECURE-MESSAGING | Async patient messaging — text-only Y1 · 90-day retention · structured templates · clinician triage queue Patient secure messaging (Y1: text-only) is delivered inside the authenticated patient portal and the clinic studio; the channel is not email or SMS, both of which are not HIPAA-compliant by default and live outside the audit perimeter. Default retention 90 days from sent_at; clinician extension to 365 days or indefinite is logged in audit_log. Patient may revoke any of their own messages per HIPAA right-to-revoke; physical purge is scheduled by tools/postgres/purge_expired_messages.py. Structured templates cover: homework reminder, between-session check-in, crisis-plan-confirmation, appointment-confirmation, custom. Triage queue surfaces patient-initiated messages with priority sort (acute indicators first per session_layer_state · ADR-093). Tier-gated to Professional+. BAA chain: Vertix is the Business Associate; all message storage encrypted at rest with AES-256 referenced to KMS. | Internal model | 2026-05-08 |
| S-037-AUDIO-MESSAGING | Async voice memo (≤5 min) auto-transcribed with Whisper · 90-day retention · BAA-aligned Per FUNCTIONS/audio-messaging/SPEC.md: asynchronous voice messaging between clinician and patient, not real-time call. Maximum 5-minute message length; webm (browser MediaRecorder) or m4a (mobile native, planned). Transcription via Whisper large-v3-turbo running on the local vertix_models daemon (port 7778) — audio and transcript never leave the server in production (CLAUDE.md regla 23 · regla 28b PHI never leaves server). Auto-detect ES/EN/mixed language. Default retention 90 days from sent_at; clinician extension to 365 days or indefinite is logged. Patient revoke supported per HIPAA. Encryption at rest: AES-256 with audio_encryption_key_id referencing KMS. Access requires JWT bound to the thread's patient_id; no cross-patient access. Audit log retained 6 years per HIPAA Security Rule § 164.312(b). v1 limitations: no native push to patient (email-only notification), no end-to-end encryption client-to-client. Tier-gated to Professional+. | Internal model | 2026-05-08 |
| S-038-WEARABLES | Wearables integration — HealthKit + Health Connect Y1 · Whoop + Garmin Y2 · objective autonomic context per session Per FUNCTIONS/wearables-integration/SPEC.md: data points include HRV (SDNN or RMSSD per vendor), resting heart rate, total + REM + deep sleep hours, sleep quality score, steps, active minutes, stress score, SpO2, and skin-temperature delta where vendor-supported. Y1 (Sept 2026): Apple HealthKit via react-native-health (iOS) and Google Health Connect via react-native-health-connect (Android). Y2 (Q1 2027): Whoop REST API v1 and Garmin Health API with webhook push. BAA chain documented in legal/regulatory/HEALTHKIT_BAA_WORKAROUND.md: Apple does not sign BAAs · HealthKit data flows iPhone → Vertix Clinic app → Vertix Clinic server (BAA-covered) without traversing iCloud. Google Health Connect covered under Google Cloud BAA. Whoop and Garmin BAAs pending Y2 enterprise negotiation. Crisis indicators (RHR > 15% over baseline AND sleep_hours < 5 for 3 consecutive days) surface as informational soft-alert per FDA CDS Criterion 4. Sensitive fields (menstrual_cycle_phase) require additional explicit consent and role-restricted access. OAuth tokens encrypted with AES-256-GCM (MISSION_OAUTH_ENCRYPTION_KEY). | Internal model | 2026-05-08 |
| S-039-DEVILS-ADVOCATE | Every clinical hypothesis returns counter-evidence weighted by Tier A literature with Daubert reasoning chain pre-built Per FUNCTIONS/devils-advocate/SPEC.md: when the clinician records a favorite clinical hypothesis, Devil's Advocate auto-generates 3-5 counter-frames templated through the multi-paradigm engine (k=34) so a single training-data majority paradigm cannot dominate. Each counter is grounded in Tier A papers (GRADE high + Q1 + percentile >=75) per ADR-073, with Tier B included at 1x weight and Tier C excluded by default. Counter-evidence is span-cited with DOI per claim. A bias auditor flags 6 cognitive biases (anchoring, availability, premature closure, confirmation, base-rate neglect, illusory truth) per Croskerry 2003. A Daubert reasoning chain is auto-materialised on every fire (audit_log row containing alternatives_considered, weights, decision_rationale) so the clinician can produce a forensic trail in deposition without reconstructing it from memory months later. Wording is locked to 'consider exploring' / 'differential includes' / 'counter-evidence suggests' · NEVER 'patient is' or 'diagnosis should be' · enforced by a CI lint rule on the prompt template. The panel is informational; clinician must explicitly accept or reject each counter (FDA CDS Criterion 4 alignment). | Internal model | 2026-05-08 |
| S-040-WHATIF-SIM | Switch paradigm or treatment hypothetically · modeled outcome distribution from cohort literature · not a clinical guarantee Per FUNCTIONS/what-if-simulator/SPEC.md: clinician triggers up to 5 candidate branches (less than 2 has no comparative value; more than 5 overloads cognitive bandwidth and pushes toward FDA SaMD scope). Each branch shows modeled response, remission, adverse-event, and discontinuation rates with 95% confidence intervals. Confidence intervals are computed by SQL JOIN over Tier A papers filtered to a cohort comparable to the patient (age band, baseline severity, prior treatment class, comorbidity profile) and weighted by sample size — per CLAUDE.md regla 34, this is a deterministic Python aggregation, not an LLM call. If fewer than 3 cohort papers match, the simulator returns 'insufficient cohort data' rather than fabricating an estimate. Every cell carries a 'modeled · not a guarantee' inline disclaimer enforced by a CI lint rule. Off-label branches are explicitly tagged and pre-fill into orders is BLOCKED per CLAUDE.md regla 28b clinical dose facilitation boundaries. Fires only on clinician trigger — never auto-fires — to stay inside FDA CDS Criterion 4 and avoid burning LLM tokens on cases without a decision pivot. All inputs and outputs are PHI-bearing; processing is local on Aloe-Beta-70B per ADR-083. | Internal model | 2026-05-08 |
| S-041-PRACTICE-FORECAST | 6-month revenue and churn risk projection per patient · resource allocation guidance · modeled from your practice data Per FUNCTIONS/practice-forecasting/SPEC.md: the dashboard projects revenue, active caseload, and per-patient churn risk over a 6-month horizon (3- and 12-month toggles available). Revenue projection is a hybrid of scheduled appointments times fee times (1 - churn_risk) plus an ARIMA(1,1,1) seasonality adjustment on the clinician's past 12 months of receipts. 95% CI is bootstrapped from the clinician's own past variance. Churn risk is a deterministic logistic regression (CLAUDE.md regla 34 · no LLM in the scoring path) over five features: cancellations_30d, days_overdue, cadence_drop_pct, milestone_reached, demographic_match_to_avg_engager. Calibrated against Edlund et al. 2002 J Affect Disord and Olfson et al. 2009 Arch Gen Psychiatry retention norms. Resource allocation narrative (slot saturation, documentation hours, insurance mix concentration) is a Tier S Haiku template fill — judgment-light, deterministic core. Every forecast cell carries 'modeled · not a guarantee' inline. Cross-tenant aggregation is prohibited at the database layer (RLS policy). PHI access by the clinician's drilldown to red-zone patients triggers a HIPAA accounting-of-disclosures audit_log row. Tier-gated to Practice and above per FUNCTIONS/clinic-portal-shell pricing. | Internal model | 2026-05-08 |
| S-042-AI-PEER-REVIEW | Pre-submission methodology check + Q1-journal feedback patterns · advisory only · does not replace journal peer review Per FUNCTIONS/ai-peer-review/SPEC.md: the report has five sections — methodology check against the relevant standard (CONSORT, CONSORT-pilot per Eldridge 2016, STROBE, PRISMA, STARD, SRQR, COREQ); citation density vs Q1 norms in the manuscript's domain (SQL JOIN over the Tier A subset of the 240k-paper corpus); top-5 ranked predicted reviewer concerns with mitigation; editorial template format check vs the target journal's published author guidelines (5 Q1 venues at v1: J Traum Stress, JCCP, Psychol Bull, Am J Psychiatry, Lancet Psychiatry); methodological strengths to highlight. Reviewer-pattern training corpus is restricted to journals that publish reviewer reports under open licenses (eLife, F1000Research, Wellcome Open Research, PLOS, BMC, Royal Society Open Research) — no PubPeer or non-licensed scrape. Auto-generated cover-letter snippet aligns with COPE / ICMJE 2024 disclosure guidance; AI Peer Review is NEVER listed as co-author. No 'predicted % acceptance' score is generated — that would be statistically dishonest given the training corpus only contains published papers (survivorship bias) and would create FTC § 5 risk. Tier-gated to Doctoral and above; Master's tier excluded. | Internal model | 2026-05-08 |
| S-043-CE-CREDITS | 1 CE credit per 5 case reads + 1 per 1.5 hr structured paper review + Vertix learning modules · auto-issued APA/ASWB/NBCC accredited (accreditation pending pre-launch) Per FUNCTIONS/ce-credits-tracker/SPEC.md: credits are derived from existing clinical workflow activity rather than purchased course-by-course. Conversion is calibrated to APA/ASWB/NBCC norms — 5 case studies reviewed = 1 hour; 1.5 hours of structured paper review (paper plus 5-question post-test or 250-plus-word reflection) = 1 hour; Vertix learning modules carry explicit hour counts. State-specific requirements engine covers the top 15 states by clinician volume × 4 license types (LCSW, LMFT, LPC, Psychologist) at v1 = 60-cell matrix; remainder of the 51-state × 4-license-type matrix added Q1-Q2 2027. Accreditation pathway: APA-Approved Provider, ASWB ACE, and NBCC ACEP applications filed pre-launch with combined annual cost approximately $7,500 and lead time 3-12 months. UNTIL THE FIRST ACCREDITATION IS GRANTED, the dashboard frames hours as 'Self-tracked CE hours · accreditation pending' — no certificates are issued and the marketing claim 'APA/ASWB/NBCC accredited' is gated by a feature flag tied to grant date. Self-study caps are hard-enforced at certificate issuance (e.g. CA 75% cap) so no uncountable certificates leave the system. Tier-gated to Practice tier and above. | Internal model | 2026-05-08 |
| S-044-MULTILANG-INTERP | Real-time AI-assisted translation across 25 languages + cultural context flags · NOT a substitute for certified medical interpretation · disclosure shown to patient Per FUNCTIONS/multi-language-interpretation/SPEC.md: bidirectional voice + text translation between clinician and patient with end-to-end latency target of 2 seconds or less per turn. Spanish only at v1; v2 expands to Mandarin (Simplified + Traditional), Vietnamese, Tagalog, French (incl. Haitian Creole sibling). Cultural context tags reuse the Cultural Mediator 12-dimension engine and an idiom catalog of approximately 500 hand-curated Spanish entries with regional dialect tags (Mexico / Caribbean / Central America / South America), each with a clinical-anthropology consultant attestation. Critical safety: a protected-phrase library of approximately 300 regex patterns covers suicidal ideation, abuse disclosure, and custody-related phrasings across regional variants. Detection forces human review of the translation before the clinician sees the English version — never trust LLM-only on these. Patient consent flow runs in the patient's language with locked wording: 'AI-assisted translation · NOT certified medical interpretation · you may request a certified interpreter at any time at no cost'. Warm transfer to Cyracom or LanguageLine VRI (Video Remote Interpretation) within 90 seconds for clinician-triggered escalation. Translation runs on local Aloe-Beta-70B FT (primary) and NLLB-3B (fallback) per ADR-083; PHI never leaves the server. Azure BAA-covered translation is opt-in only with audit log. 17 US states require certified medical interpreter for clinical encounters under Title VI Limited English Proficiency obligations + state Medicaid contracts; AI is NOT a certified interpreter; disclosure plus offer of certified interpreter at no patient cost is a hard requirement. | Internal model | 2026-05-08 |
| S-045-CULTURAL-MED-V2 | 12 cultural dimensions surfaced as 'consider exploring' · structured citations to peer-reviewed literature · NEVER affirmative claim about a patient Per FUNCTIONS/cultural-mediator/SPEC.md: the engine implements the DSM-5-TR Cultural Formulation Interview (CFI) 12-domain framework, surfacing only the 3-6 dimensions most relevant to the encounter rather than all 12 every time. Wording is STRICTLY locked: 'consider exploring whether the patient frames distress in [X] way' · 'cultural factors that may be relevant include [X]' · 'literature suggests population-level pattern: [X] · verify with patient'. NEVER 'patient is/has', 'suffers from', 'presents with [diagnosis]', or any affirmative claim about a culture-bound syndrome. A CI lint rule runs 50+ prompts on every release and rejects deploy if a banned phrase appears in output. Idiom catalog v1 = 47 entries: 9 from DSM-5-TR Glossary (ataque de nervios, dhat, khyâl cap, kufungisisa, maladi moun, nervios, shenjing shuairuo, susto, taijin kyofusho) plus 38 expansions across East Asian, South Asian, sub-Saharan African, MENA, and Latin American regions. Every entry carries a source DOI and a clinical-anthropology consultant attestation. Each cultural flag carries DOI(s) inline · clicking opens a citation card with paper title, year, journal, and Tier (per ADR-073). A lightweight self-paced onboarding module (~15 min) covers population-level vs individual-level reasoning, the 'consider exploring' framing, APA Multicultural Guidelines (2017), and DSM-5-TR CFI before the clinician can enable the panel. Stereotype-risk red-team budgeted pre-launch. | Internal model | 2026-05-08 |
| S-046-LIVE-COPILOT-POSTSESSION | Within 5 min of session end · auto-suggested SOAP note + 'what struck me' signals + 3 plan-next-session candidates · clinician approves before save · audit log required Per FUNCTIONS/live-copilot/SPEC.md: pipeline target p95 ≤ 4 minutes and p99 ≤ 5 minutes from session_end event to panel ready. Three components: 'what struck me' triage (3 surface signals from the session in 'consider exploring' framing), SOAP/DAP/BIRP note draft (clinician selects format preference), and 3 plan-next-session candidates each cited with cohort papers from the Tier A subset. Approve-before-save gate is non-negotiable: the Save button is disabled until the clinician explicitly ticks 'I have reviewed and approve this note as accurate'. Backend rejects approval payloads with the attestation flag false. The audit log records note_hash_before_edit, note_hash_after_edit, edit_distance_pct, and time_to_approve so rubber-stamping can be detected post-hoc. Hallucination rate target: less than 2% major errors (factually wrong about session content) and less than 5% minor errors (phrasing imprecision · missing nuance), measured by a weekly sample audit of n=20 sessions reviewed manually. Above-threshold triggers a prompt-engineering sprint. Patient may opt out of AI assistance at intake (`patient_preferences.ai_assistance = false`) and the clinician falls back to fully manual workflow. All inference runs on local Aloe-Beta-70B per ADR-083; PHI never leaves the server. Forensic export available via the forensic-daubert function. | Internal model | 2026-05-08 |
Corrections and updates
This register is updated when claims change, data sources are refreshed, or a factual error is identified. To report an inaccuracy or request additional substantiation for any entry, contact [email protected]. We acknowledge all reports within 2 business days. The Disclosures page contains the full responsible advertising statement and conflict reporting protocol.
Audit it yourself. Every recommendation.
We publish the methodology because accountability requires transparency. Independence is engineered, not promised.