If you are building a panel-size or revenue forecast for the CMS ACCESS Model, there is a line in the eligibility criteria that changes the arithmetic, and it is easy to read past because it looks like the other exclusions around it.
It is not like the others. The other exclusions describe something already true about a patient — their coverage, their other programs, their existing alignment. This one describes something CMS decides after you ask.
Everything below is quoted from CMS documents with the page number attached, fetched July 26, 2026.
From the ACCESS Request for Applications, version 1.1, page 16, under Beneficiary Eligibility → Program Exclusions:
> 3. Program Exclusions: > a. Not be enrolled in Medicare Advantage, Program of All-Inclusive Care for the Elderly (PACE), or the Medicare hospice benefit… > b. Not be actively aligned to another ACCESS Participant for the same track starting within the past 3 months. > c. Not be assigned to the control group. See Section 10: Evaluation, for additional information on the model's randomized evaluation.
Items (a) and (b) are facts you can establish about a patient before you begin. Item (c) is a coin CMS flips.
The evaluation section, RFA pages 39–40, in full on this point:
> The ACCESS Model will be evaluated using a randomized design with beneficiary-level randomization. CMS will conduct the randomization and communicate assignment results to Participants through the CMS Eligibility and Alignment APIs. Before querying the APIs to verify a beneficiary's eligibility or enroll a beneficiary, Participants must inform the beneficiary that the service is offered as part of a new CMS payment model test and that, as part of model evaluation, their data may be shared with CMS subject to federal privacy and security protections; that they may be randomly assigned to a comparison group and therefore ineligible to participate in the clinical track while continuing to have access to all usual Medicare benefits and services; and that such assignment will not affect their Medicare benefits, rights, or coverage. If a beneficiary is assigned to the control group, they will not be eligible to participate in the model for the following 12-month period. To preserve the integrity of the evaluation, Participants will be required to use standardized language provided by CMS when informing beneficiaries of assignment to the control group.
And the next sentence, which is the one that makes this sizeable:
> The randomization rate will start at 90:10 intervention-to-control ratio in the first year and may be adjusted or eliminated in future years depending on the number of beneficiaries enrolled in the model and the evaluation's ability to detect impact of the model on improved quality and reduced Medicare spending.
CMS's Technical FAQ says the same thing to patients, in patient language:
> Because the model's new payment approach is being tested, a small share of people who try to enroll in a specific ACCESS track may be randomly assigned to a control group for that track. This process helps CMS evaluate the model's impact… Individuals in a control group continue to have full access to all regular Medicare services and can work with their usual health care providers.
One: your ceiling is not 100%, it is 90 — in year one. The randomization rate starts at a 90:10 intervention-to-control ratio in the first year, so in that first year the share of your eligible, willing, fully-worked-up patients who can actually align is bounded above by about nine in ten. Not because of anything you did.
Carry the hedge with the number, always: 90:10 is a year-one parameter that CMS reserves the right to adjust or eliminate in future years. A forecast built on it is a year-one forecast and should say so on its face. Do not extend it across a multi-year model without saying that you have assumed CMS left it alone.
Two: the loss is not recoverable next quarter. A beneficiary assigned to the control group "will not be eligible to participate in the model for the following 12-month period." That is not a patient you re-approach after a cycle of outreach. For forecasting purposes they leave the addressable pool for a year.
**Three: there is a disclosure obligation, and it fires before the API call.** Read the sequencing in the passage above: "Before querying the APIs to verify a beneficiary's eligibility or enroll a beneficiary, Participants must inform the beneficiary…" The obligation is on the Participant, it is pre-check rather than post-result, and it has required content — model test, data sharing, possible comparison-group assignment, no effect on Medicare benefits.
And when a result does come back control-group, the language is not yours to write: "Participants will be required to use standardized language provided by CMS." CMS publishes that text in its ACCESS Beneficiary Notices Reference. We are deliberately not reproducing it here — it is a string CMS requires be used unmodified, and the only sensible place to copy an unmodifiable string from is CMS. Have whoever owns your patient-facing scripts pull it from the source.
Four: it lands on the workflow, not just the spreadsheet. Every one of these patients has been identified, screened for a qualifying condition, and had a conversation, and then comes back ineligible through no clinical fact about them. If your enrollment workflow has no branch for that outcome, it has two branches where CMS describes three.
The RFA uses two different scopes in adjacent sentences. The disclosure sentence says a beneficiary may be randomly assigned to a comparison group "and therefore ineligible to participate in the clinical track." The next sentence says they "will not be eligible to participate in the model for the following 12-month period." The Technical FAQ sides with the narrower reading — "a control group for that track."
This matters if you run more than one track, because the RFA elsewhere (page 16) is explicit that "beneficiaries can simultaneously align with the same or different participants in multiple different tracks." Under the narrow reading, a control-group assignment in one track leaves the others open; under the broad one, it does not.
We are not going to tell you which is right, because we do not know. Two of the three sentences favour the narrow reading and one is written broadly, and none of them is the Participation Agreement. If you are modelling a multi-track panel, this is a question to put to CMS directly rather than to infer from us.
This is the part we would rather not write, so it goes near the top of the second half rather than in a footnote.
Our own ROI calculator asks you for the "achievable enrollment of eligible patients" as a percentage, and until this week it let that percentage run to 100. Every panel figure we have published downstream of that input carried an unstated assumption — that every eligible, willing patient converts to an aligned one — and that assumption is not available in year one, because roughly one in ten of them is assigned to the comparison arm before anyone gets a say.
The direction matters. The error runs toward participation: it makes the model look better, not worse. That is the direction with no natural adversary, which is our best explanation for why it survived in our own file for weeks without anyone here catching it — we did not go looking for a reason our own number was too generous.
The size, honestly: at the 70% default the calculator ships with, there is no arithmetic error at all — 70 is below 90, and the figure a first-time visitor sees is unaffected. At a 100% enrollment assumption, a year-one figure is overstated by about 11%. Above 90%, the scenario is not merely optimistic; in year one it is unavailable.
The practical instruction, which is the same one we are applying to ourselves: for a year-one scenario, keep your enrollment assumption at or below 90%.
We are telling you this because a forecast you cannot check is worth nothing to you, and because the alternative — quietly correcting it and saying nothing — would make every other number on this site a thing you have to take on faith.
We have read two CMS surfaces on this and they agree on the mechanism, the 12-month duration, the year-one rate and the disclosure obligation. They differ on track-versus-model scope, which we have flagged above rather than resolved.
We have not read the CMS Beneficiary Notices Reference first-hand this week and we are not quoting from it. We have not seen an ACCESS Participation Agreement — we looked and could not find one published for this model, which is not the same as its not existing. If one is published and it defines any of this differently, it would outrank the RFA and this post would need revising.
CMS has moved an ACCESS parameter by web page before without reissuing the PDF, and this particular parameter is one the RFA explicitly says may change. So the standing instruction: check the Technical FAQ and the model page on the day it matters to you. Our reading is current as of July 26, 2026.
And the ordinary disclaimer, which is not boilerplate here: this is a summary of published CMS documents, not legal, financial, billing, or clinical advice. CMS decides eligibility, assignment and payment. We have given you the page numbers so your own review can start from the source.
Stated plainly, so you can discount it: we build reporting infrastructure for ACCESS participants. A randomized comparison arm is on our blog because it produces a third terminal state in an eligibility workflow — eligible, not eligible, control — and a terminal state that a workflow does not have a branch for is the shape of problem we work on. It is also the reason we found the flaw in our own calculator, which is a fair summary of what our week looks like.
That is not a pitch. Nothing in this post requires a vendor; all of it is free to read at the links below.
ACCESS is a randomized evaluation with beneficiary-level randomization. A share of the patients you try to enroll will be assigned to a comparison group instead, will be ineligible for the next 12 months, and CMS decides which ones. The RFA publishes the rate: a 90:10 intervention-to-control ratio in the first year, which CMS may adjust or eliminate in later years. So a year-one panel forecast is capped near 90% of your eligible, willing patients — including in our calculator, where it was not, until this week.
If you want to walk your own panel forecast against this, we are happy to do it with you: [email protected].
---
Outcome Rail builds reporting infrastructure for participants in the CMS ACCESS Model. This post summarizes publicly published CMS documents and is not legal, financial, billing, or clinical advice. It states no measure target values and makes no claim about clinical outcomes. CMS decides payment.
Sources: CMS ACCESS Request for Applications (PDF) — v1.1, last modified 2026-02-12; Beneficiary Eligibility / Program Exclusions p.16, multi-track alignment p.16, Evaluation and randomization pp.39–40 · ACCESS Technical FAQ — Who can use ACCESS services? · CMS ACCESS Model page. Both quoted surfaces fetched 2026-07-26.
Device, lab, and PROM data in; compliant FHIR submissions out — validity windows, cadence clocks, and provenance rules enforced before CMS ever sees the bundle. We're onboarding a small founding cohort of design partners this quarter.