In the same summer Washington let an unauthorized AI therapist bill Medicare, five states wrote that exact service into law as illegal — and a FOIA suit just surfaced CMS's other AI experiment, where one prior-auth request sat 83 days
The US government is currently doing two opposite things to medical AI at once. On September 15 the Electronic Frontier Foundation published roughly 1,000 pages of internal CMS documents obtained through a FOIA lawsuit, covering WISeR, Medicare's AI-assisted prior-authorization model: two vendors denied more than 20,000 requests in the pilot's first three months, one of them denying more claims than it approved, and a single request went 83 days without a response — against a promised 72-hour window (EFF, 2026-09-15). In the same stretch, FDA's TEMPO pilot let four not-yet-authorized generative-AI devices reach Medicare beneficiaries legally, among them a voice agent that phones older adults and delivers 20-minute CBT sessions (FDA). Meanwhile at state level, Maine, Vermont, Tennessee and Colorado have all made "AI independently delivering psychotherapy" unlawful or an unfair trade practice this year (Becker's). What this issue works through is that contradiction: one service treated simultaneously as a policy experiment and as an illegal act.
01 — Top Stories
A FOIA suit pulls up 1,000 pages: in the first three months of Medicare's AI prior-auth pilot, two vendors denied over 20,000 requests and one request sat 83 days
EFF filed a FOIA lawsuit against CMS in March 2026 and on September 15 published roughly 1,000 pages it obtained — vendor contracts, internal weekly status reports and provider feedback forms — covering WISeR (Wasteful and Inappropriate Service Reduction), which went live across six states in January 2026. The records show the promised 72-hour turnaround was missed routinely, with one request going 83 days unanswered; two vendors together denied more than 20,000 requests in the first three months; Virtix denied more claims than it approved over that period, prompting CMS to require corrective action; and vendor Innovaccer told CMS a month before launch that its software was not fully functional and had not been fully tested — the model launched on schedule anyway. The whole thing went from planning to live in about six months (EFF, 2026-09-15; STAT, 2026-09-15).
This is the first set of primary documents — not a vendor deck, not an agency white paper — showing what a public-sector AI decision system actually looks like in operation. The incentive structure is the striking part: vendors are paid in connection with denials, and while CMS ties payment to quality scores, poor performance reduces compensation by only 5–10% — not enough to offset the pull toward denying one more claim. One provider feedback line quoted by EFF reads: "Patients calling our offices crying in pain," referring to procedures delayed while approval was pending (EFF). The governance lesson: regulatory debate fixates on whether the model is accurate, but what failed here was procurement incentive design and launch-schedule pressure, not model precision.
EFF is the plaintiff in this case and has a position on how the documents are selected and read; what is quoted is a small slice of 1,000 pages and the rest cannot be assessed. Same-day coverage in STAT and Medscape sits behind paywalls — this report confirmed direction only from publicly visible headlines and standfirsts, and takes its figures from EFF's public write-up. No full CMS response to this document set was visible at the time of writing.
FDA's TEMPO roster fills to four: unauthorized generative-AI devices can legally touch Medicare patients, including a voice agent that phones people to do CBT
FDA's TEMPO (Technology-Enabled Meaningful Patient Outcomes) pilot now lists four selected participants: SonderMind's SACA app (behavioral health, depression and anxiety in adults 22+), Limbic's Unpacked (an AI voice agent delivering CBT), Cadence Solutions' HypertensionOS (helping clinicians manage Stage 2 hypertension), and Dexcom's Glucose Health Program (metabolic monitoring and prediabetes screening). The first was announced July 22, 2026; the two behavioral-health firms joined August 24 (FDA; MedTech Dive). The mechanism is enforcement discretion over certain requirements, conditioned on the devices being offered through participants in CMS's ACCESS model. Limbic's Unpacked runs as scheduled 20-minute phone calls, reachable on a landline or a basic handset, with a theoretical addressable population of the 34 million Medicare beneficiaries covered by ACCESS (Fierce Healthcare, 2026-08-26).
This is the first time FDA has decoupled premarket authorization from patient access — and the cost of the decoupling does not sit with FDA, it is bolted onto CMS's payment machinery. Lower regulatory threshold in, government as the payer. Limbic CEO Ross Harper frames it as "existing regulatory frameworks were not written for generative AI in healthcare — they predate that technology," and "the very first example of payers paying for outcomes delivered, not services rendered" (Fierce Healthcare). For Taiwan and other markets that track US precedent, this is a pathway worth watching: it moves real-world evidence collection from post-market to pre-market, and the first cohort bearing that shift is elderly Medicare beneficiaries.
The "34 million beneficiaries" figure is the theoretical denominator of the ACCESS model, not Limbic's actual caseload — the gap is enormous. The exclusions are substantial too: Limbic's service excludes non-English speakers and people without a telephone, and SonderMind's app is contraindicated for patients with suicidality, manic episodes, psychosis or PTSD (MedTech Dive). Limbic's claim of deployment across 12 US states and the UK supporting 850,000 patients is company-reported and not independently audited.
On the same day, CMS moves to expand the ACCESS model — the very pipe that pays for TEMPO devices is widening
On September 15, Medicare announced it is expanding the ACCESS (Advancing Chronic Care with Effective Scalable Solutions) pilot to include more technology-supported chronic disease management options (STAT, 2026-09-15). ACCESS was announced December 23, 2025, runs for ten years, and covers telehealth, remote monitoring through FDA-authorized devices, digital coaching and asynchronous care; applications are rolling, with the first cohort closing April 1, 2026 for a July 1 start and later applicants folded into a January 1, 2027 cohort (ArentFox Schiff). CMS greenlit more than 150 participants on April 13 (STAT, 2026-04-13).
TEMPO only has commercial meaning to the extent ACCESS is wide. FDA granted enforcement discretion, not money; the step that converts discretion into revenue is ACCESS. ACCESS ordinarily requires "FDA-authorized devices," while TEMPO's selectees are precisely the ones lacking authorization — the seam between the two models is the central clause of the US digital-health business model for the next decade. Any medical-AI company eyeing the US should be reading this paired design of market exemption plus payment linkage, not just the 510(k) route.
The primary report on the September 15 expansion is behind STAT's paywall; this report can confirm only the direction — that it is expanding — and has no citable public detail on which items, what sums or what timeline. The per-patient payment bands for ACCESS (a $90–$420 range has circulated in secondary coverage) could not be verified against primary CMS documents and are therefore not used here.
Five states have already made "AI independently delivering therapy" unlawful — the voice agent Washington just waved through is an unfair trade practice in Maine
Five states have enacted restrictions on AI therapy in 2026: Vermont (effective June 17, barring AI systems from independently delivering mental health services), Tennessee (July 1, prohibiting development or deployment of AI systems that hold themselves out as qualified mental health professionals), Maine (July 29, making AI therapy offered without a licensed professional an unfair trade practice), Colorado (August 12, HB26-1195), and Rhode Island (effective January 1, 2027) (Becker's Behavioral Health). Colorado sets the highest bar: AI may not engage in therapeutic communication with a client unless "the provider, the AI system, and the client are all participating in a synchronous, real-time interaction" — the provider must be present throughout, not reviewing afterwards; AI-generated therapeutic recommendations require provider approval before reaching the client; recording or transcribing with AI requires written, revocable consent. Violations carry administrative fines up to $5,000 and civil penalties up to $20,000 under the Colorado Consumer Protection Act (Snell & Wilmer; Colorado HB26-1195).
Read this alongside the previous item and today's real question comes into focus: what FDA waves through with enforcement discretion is not binding on state practice acts or consumer protection law. TEMPO's paperwork is FDA discretion over premarket requirements, not a state's licence to say who may deliver psychotherapy. Colorado's synchronous-presence clause all but cancels the economics of the Limbic model — one clinician supervising a large volume of AI conversations — and that economics is the entirety of what Harper means by "less scale-limited by clinician supply." This is not a theoretical conflict; it is two legal instruments giving opposite answers about the same service. Colorado also has HB26-1139, effective January 1, 2027, governing payers using AI for coverage decisions: determinations must be grounded in the individual patient's medical history, denials must be routed to a competent clinician, and AI use must be disclosed to state regulators — aimed squarely at systems of the WISeR type.
FDA's generative-AI device discussion paper: comments close October 19 — this year's docket most worth filing on
FDA's Digital Health Center of Excellence, within CDRH, published "Considerations for the Regulation of Generative AI-Enabled Medical Devices" on August 19, 2026 and opened it for comment under docket FDA-2026-N-7874, closing October 19, 2026. The paper covers four areas, each with its own questions: risk assessment for generative-AI devices, premarket evaluation approaches, postmarket monitoring strategies, and other relevant topics. FDA states plainly that this is a discussion paper rather than draft guidance and does not establish CDRH's regulatory expectations for future submissions; respondents may answer only the questions relevant to them (FDA; FDA press release).
TEMPO is a case-by-case exemption; this discussion paper is the embryo of the general rule. FDA itself concedes in the paper that generative-AI devices carry "unique risks when compared to traditional software and AI-enabled medical devices" — while simultaneously letting four such devices reach patients unauthorized. Read together, the agency's actual strategy is legible: collect operating data through small exemptions first, then write the general rule. With a month left on the clock, this is one of the few windows in which a non-US manufacturer — Taiwanese device and digital-health firms included — can shape future rules without a US footprint.
Florida's AG proposed criminal liability on September 9 — and plans to write chatbot practice rules with the state Board of Medicine
Florida Attorney General James Uthmeier, alongside FDLE Director Mark Glass, laid out a liability package for AI chatbot companies on September 9, 2026: criminal sanctions, heavy fines and victim payments, court-ordered monitorship, and suspension of activity in Florida. It targets firms with "practical control" over a chatbot's design, training, deployment and safety settings; OpenAI is currently under investigation by his office. Two strands bear directly on healthcare: Uthmeier said chatbots dispensing unlicensed medical guidance has produced adverse health outcomes and that he intends to work with Florida's Board of Medicine on rules, and he cited chatbots "encouraging users to take their own lives" (WFSU, 2026-09-09).
The four items above concern what counts as a device, who may deliver therapy, and who pays. This one concerns where liability lands. Handing rulemaking to a state Board of Medicine pulls general-purpose chatbots into the jurisdiction of medical licensure, rather than handling them through product liability or consumer protection alone — and if that path holds, it reaches beyond medical-AI companies to every general model that answers a health question in conversation. This remains a proposal; nothing has been enacted.
The EU pushed AI obligations for device-embedded AI to August 2028 — and narrowed what counts as a "safety component"
EU institutions reached provisional political agreement on the Digital Omnibus on May 6, 2026, confirmed by member state representatives on May 13. Three points matter most for devices. First, high-risk obligations for Annex I (AI embedded in regulated products, medical devices included) move from August 2, 2027 to August 2, 2028; Annex III (stand-alone high-risk systems) moves from August 2, 2026 to December 2, 2027. Second, for Annex I products already covered by EU sectoral safety legislation such as MDR/IVDR, the Commission may limit the application of specific AI Act requirements where the sectoral rules already impose equivalent obligations, cutting duplicative regulation. Third, the "safety component" definition was narrowed: AI used solely for user assistance, performance optimization, service efficiency, automation, convenience or quality control no longer automatically qualifies as high-risk unless its failure would endanger health and safety (Gibson Dunn).
This is the third piece of evidence that the three major jurisdictions are moving the same way this year: the US granting early access through enforcement discretion, the EU deferring obligations and narrowing definitions, the UK institutionalizing its sandbox. The practical implication for manufacturers is that compliance pressure between 2026 and 2028 comes mainly from MDR/IVDR rather than the AI Act. But the narrowed "safety component" definition also means classifying your own AI now leans harder on a case-by-case argument about whether failure endangers health and safety — more documentation work, not less.
This is the only item here that is not from the past week; it is included because it is the coordinate needed to read the other six. Details are drawn from law firm analysis rather than the EU's official legal text — article numbers and final wording should be checked against the Official Journal.
NMPA opens a draft clinical-evaluation guideline for AI-assisted diagnostic devices — and published 57 device guidelines in one batch in July
China's NMPA issued a draft Guideline on Clinical Evaluation of AI-Assisted Diagnostic Medical Devices, covering Class III software products and focused on systems that characterize lesions as benign or malignant in medical imaging — pulmonary nodule, thyroid nodule, breast nodule or gastrointestinal polyp evaluation systems. Separately, on June 30, 2026 NMPA published two guidelines on brain-computer interface devices covering product classification and nomenclature (non-invasive stroke rehabilitation BCI devices get a Class II exception), and on July 21 released 57 final device guidelines in a single batch spanning type testing, non-clinical evaluation, clinical assessment and submissions (China Med Device, 2026-08).
While the US concentrates on generative AI and payment mechanics and the EU defers, NMPA is moving in a visibly different direction: writing the clinical-evaluation standard for existing discriminative AI imaging products in finer detail, while pre-building classification and nomenclature for frontier categories like brain-computer interfaces. It is a lay-the-plumbing-first cadence. For Taiwanese firms the practical consequence is that clinical-evaluation expectations for Class III AI imaging software entering China are becoming both clearer and more demanding.
This item comes from an English-language regulatory consultancy roundup rather than NMPA's own notice; the draft's exact publication date and comment deadline are not stated in that source. Anyone acting on it should work from the original NMPA posting.
02 — Product Analysis
Limbic Unpacked
AI voice agent delivering CBT by phone · Limbic Inc. (London, UK)
Function and position. Unpacked runs 20-minute CBT sessions delivered by an AI voice agent over scheduled phone calls under clinician oversight, deliberately built to work on a landline or a basic handset so it reaches older adults who do not use apps. It is not sold to individual consumers but to providers inside CMS's ACCESS model, paid for through Medicare's outcomes-based payment (Fierce Healthcare, 2026-08-26).
- Strength : distribution ahead of technology. It is among the first AI mental-health products granted FDA enforcement discretion to reach Medicare patients without authorization, and that status comes bolted to a payment channel — a combination no competitor currently holds (FDA).
- Strength : low-tech reach. Using the phone rather than an app routes around the single most common adoption barrier for digital health products among older adults.
- Concern : state law may dismantle the business model outright. Colorado's HB26-1195 requires synchronous, real-time participation by provider, AI and client together, cancelling the economics of one clinician overseeing many AI sessions; Maine makes AI therapy without a licensed professional an unfair trade practice (Snell & Wilmer).
- Concern : the missing number is efficacy. TEMPO is by design market-first, evidence-after, so there is no citable third-party controlled trial result yet; the company's 850,000-patient figure is usage, not outcome (MedTech Dive).
Cadence HypertensionOS
Stage 2 hypertension management support · Cadence Solutions, Inc. (US)
Function and position. HypertensionOS assists licensed healthcare providers in managing Stage 2 hypertension in adults, sitting in TEMPO's early cardio-kidney-metabolic category. The decisive difference from Limbic is who decides: it places AI inside the clinician's workflow as support rather than having AI conduct therapeutic conversation with the patient directly (FDA).
- Strength : a far safer legal position. Because the terminal decision stays with a licensed clinician, it does not collide with state practice-act thresholds on who may deliver care — exactly what the five state statutes target (Becker's).
- Strength : the endpoint is hard. Blood pressure is objective and repeatably measurable — easier to settle under outcomes-based payment than depression or anxiety scales, which matters in a model like ACCESS that pays on the share of patients hitting targets (ArentFox Schiff).
- Concern : the safe position is also the crowded one. Remote hypertension management is among the most contested segments in digital health, and positioning AI as clinician support leaves far less differentiation room than Limbic's substitutive design.
- Concern : public information is thin. FDA's listing gives a single line of intended use, and this report found no public source for the product's model architecture, training data or clinical validation. That gap should be read as a gap, not as evidence either way.
03 — Companies & Competition
| Company | Recent state & numbers | Position & moat |
|---|---|---|
| Limbic AI mental health, London |
Selected into FDA TEMPO on August 24, 2026, the first AI mental-health firm in the pilot; its product Unpacked delivers 20-minute CBT by phone; the company reports deployment across 12 US states and the UK supporting 850,000 patients (Fierce Healthcare). | The moat is regulatory status plus payment linkage, not the model. The weakness is that the status only holds federally and evaporates against state statutes like Colorado's and Maine's. |
| SonderMind Behavioral health platform, Denver |
Also selected into TEMPO on August 24, with the SACA smartphone app for reducing depression and anxiety in adults 22 and over; contraindicated for patients with suicidality, manic episodes, psychosis or PTSD (MedTech Dive). | Positioned as an adjunct to existing therapy or medication rather than a replacement, so its state practice-act exposure is lower than Limbic's — at the cost of losing Limbic's "not limited by clinician supply" scale story. |
| Cadence Solutions Remote chronic care, US |
A TEMPO selectee whose HypertensionOS assists licensed providers in managing Stage 2 hypertension in adults, classified under early cardio-kidney-metabolic (FDA). | An assistive play: low legal risk, narrow differentiation. Its real competition is every incumbent remote blood-pressure management service, not the other TEMPO selectees. |
| Dexcom Continuous glucose monitoring, US |
The only established large device maker among TEMPO selectees, with its Glucose Health Program monitoring metabolic status and aiding prediabetes and type 2 diabetes screening, spanning both early and standard cardio-kidney-metabolic categories (FDA). | It does not need TEMPO for market access — it needs the ACCESS payment channel. Its installed base and distribution are the moat, and startup rivals have no equivalent chip in this square. |
| Innovaccer Healthcare data platform, WISeR vendor |
FOIA documents show Innovaccer notified CMS one month before WISeR's January 2026 launch that its software was not fully functional and had not been fully tested; the model launched on schedule regardless (EFF). | Government contracts are both the moat and the risk: they expose the company's technical readiness to FOIA, a form of disclosure no private-sector customer creates. |
| Virtix Prior-auth vendor, WISeR |
Records show Virtix denied more requests than it approved in WISeR's first three months, prompting CMS to require corrective action; two vendors together denied over 20,000 requests in that window (EFF, 2026-09-15). | Under a contract where compensation tracks denials and quality scores move pay by only 5–10%, its "performance" runs opposite to the public interest; Colorado's HB26-1139, which from 2027 requires denials to be routed to a competent clinician, is aimed straight at this model. |
Today's competition is not between models but between legal statuses. The moat for the four TEMPO firms is a sheet of enforcement discretion plus a payment channel, not technology; and what went wrong for the two WISeR vendors was contract incentives, not model precision. Who gets to sell medical AI in the US currently turns on satisfying federal exemption conditions and state practice acts at once — and those two now point in opposite directions.
04 — Taiwan Angle
(1) Taiwan's instrument today is a guideline, not a pathway. On May 29, 2026 the Ministry of Health and Welfare issued the Guideline on the Use of Generative AI in Healthcare Institutions (ref. 1151663164), covering public and private hospitals and clinics across chart-writing assistance, clinical decision support, administrative documents and patient communication — and expressly leaving out autonomous AI agent systems as higher risk. The guideline is administrative guidance, not a binding rule, and institutions may adapt it to circumstances (Lee and Li). Against today's theme there is a gap worth noting: Limbic's Unpacked, the very thing TEMPO is waving through, is precisely the class of autonomous agent Taiwan's guideline sets aside. Taiwan is not a step behind so much as deliberately leaving that square blank — but blank also means that if such a product applies here next year, no assessment framework exists in the current documents to apply to it.
(2) The half Taiwan is missing is payment, not review. TFDA's Smart Medical Device Project Office has run since May 7, 2021 as a single window offering one-stop consultation for AI/ML devices, along with information and matchmaking platforms (Ministry of Health and Welfare). So Taiwan already has a dedicated body on the "how do we review this" side — but the leverage in this week's US news sits on the other side: without ACCESS paying for it, TEMPO is just a sheet of discretion. The long-standing bottleneck for medical AI adoption in Taiwan has likewise been pinned on NHI reimbursement and clinician willingness rather than review speed (DIGITIMES). On that reading, the document Taiwanese regulators should be studying closely is not TEMPO but ACCESS: a ten-year model that pays on the share of patients hitting targets and writes technology-supported services into the benefit schedule.
(3) For Taiwanese firms selling into the US, this year the states matter as much as FDA. FDA authorization or enforcement discretion does not mean you can operate in fifty states: Colorado's synchronous-presence requirement for psychotherapy AI under HB26-1195, Maine's unfair-trade-practice designation, and Colorado HB26-1139's disclosure and human-review requirements for payer-side AI from 2027 are all product-design constraints, not post-market paperwork. Any design that treats the clinician-supervision ratio as its growth lever should undergo a state-by-state compliance audit before entering the US. And the FDA generative-AI docket (FDA-2026-N-7874) closes October 19, 2026 — one of the few windows for shaping the rules without a US presence (FDA).
05 — Further Reading
-
New Records Reveal Problems with Medicare's AI Prior Authorization Experiment — EFF Deeplinks (2026-09-15)
This year's most worthwhile read on public-sector AI governance, because it quotes contracts and weekly status reports rather than white papers.
-
Considerations for the Regulation of Generative AI-Enabled Medical Devices — FDA DHCoE (2026-08-19)
The docket is open until October 19; the right way to read it is to hold your own product's classification questions against its four sets of questions.
-
Update: Colorado's Three-Front AI Healthcare Landscape Has Shifted — Snell & Wilmer (2026)
One of the few analyses putting the treatment side, the payer side and the general AI act on one page — the shortest route to understanding how state law routes around FDA.
-
EU AI Act Omnibus Agreement — Postponed High-Risk Deadlines and Other Key Changes — Gibson Dunn (2026)
The Annex I / Annex III deadline split and the narrowed "safety component" definition are the two variables that actually drive EU device-AI compliance planning.
-
衛福部頒布「醫療機構應用生成式人工智慧指引」 — 理律法律事務所 (2026)
The only analysis of Taiwan's sole official document addressing clinical generative AI directly — note in particular the sentence excluding autonomous AI agents.
06 — References
- New Records Reveal Problems with Medicare's AI Prior Authorization Experiment. Electronic Frontier Foundation, 2026-09-15. eff.org
- Medicare's AI prior authorization pilot was rushed and full of problems, new documents reveal. STAT News, 2026-09-15. statnews.com
- New Records Reveal Backlogs, Testing Gaps in Medicare's AI Prior Authorization Pilot. Medscape, 2026-09. medscape.com
- Participants Selected for TEMPO for Digital Health Devices Pilot. U.S. Food and Drug Administration, 2026. fda.gov
- TEMPO for Digital Health Devices Pilot. U.S. Food and Drug Administration, 2026. fda.gov
- FDA adds two behavioral health firms to TEMPO pilot. MedTech Dive, 2026-08. medtechdive.com
- Limbic selected to FDA pilot, will test voice AI therapy with Medicare enrollees. Fierce Healthcare, 2026-08-26. fiercehealthcare.com
- Limbic Becomes First AI-Led Mental Healthcare Company Selected for FDA TEMPO. Business Wire, 2026-08-19. businesswire.com
- Medicare to expand pilot that pays for technology to manage chronic diseases. STAT News, 2026-09-15. statnews.com
- Access granted: CMS greenlights more than 150 participants for chronic care experiment. STAT News, 2026-04-13. statnews.com
- CMS Innovation Center Unveils ACCESS Model to Expand Technology-Supported Care for Chronic Disease. ArentFox Schiff, 2026. afslaw.com
- Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback (Docket FDA-2026-N-7874). U.S. Food and Drug Administration, 2026-08-19. fda.gov
- FDA Seeks Public Feedback to Inform Regulatory Approach for Generative AI-Enabled Medical Devices. U.S. Food and Drug Administration, 2026-08. fda.gov
- 5 states restrict AI therapy chatbots in 2026. Becker's Behavioral Health, 2026. beckersbehavioralhealth.com
- Update: Colorado's Three-Front AI Healthcare Landscape Has Shifted. Snell & Wilmer, 2026. swlaw.com
- HB26-1195: Concerning restrictions on the use of artificial intelligence relating to psychotherapy services. Colorado General Assembly, 2026. leg.colorado.gov
- Florida's attorney general seeks new penalties for chatbot companies abetting crimes. WFSU News, 2026-09-09. news.wfsu.org
- EU AI Act Omnibus Agreement — Postponed High-Risk Deadlines and Other Key Changes. Gibson Dunn, 2026. gibsondunn.com
- NMPA Roundup August 2026. China Med Device, 2026-08. chinameddevice.com
- 衛福部頒布「醫療機構應用生成式人工智慧指引」. 理律法律事務所, 2026. leeandli.com
- 食品藥物管理署智慧醫療器材專案辦公室成立. 衛生福利部, 2021-05-07. mohw.gov.tw
- AI醫療創新加速 醫師採用意願與健保給付成落地挑戰. DIGITIMES, 2026. digitimes.com.tw
- Joint Commission Releases Voluntary Responsible Use of AI in Healthcare Certification. The Joint Commission, 2026-05. jointcommission.org