Concept renderFDA's general wellness page frames general wellness products as low risk products that promote a healthy lifestyle. It separately quotes language about software that is unrelated to the diagnosis, cure, mitigation, prevention, or treatment of a disease or condition. Have counsel review a coach's replies against both texts.
Adding a chat or voice coach to a ring, band or watch changes what the product says, even when the sensors stay the same. This page compares the source texts that frame the wellness line, lists what to measure and record when you test a coach, and gives questions to settle before you build.
This is an engineering write-up for general wellness products. It does not cover medical devices and gives no legal advice. Whether your product qualifies as general wellness is a question for your counsel. We report no measured results, because we have not run any tests for this page.
Which source texts set the line?
Only the texts below are quoted. Dates are as of October 31, 2026, the day we read them.
| Source | What it says (quoted or closely paraphrased) | Why it matters for a coach |
|---|---|---|
| FDA, general wellness policy | The page frames "low risk products that promote a healthy lifestyle (general wellness products)". Separately, it quotes language for software that is "unrelated to the diagnosis, cure, mitigation, prevention, or treatment of a disease or condition" (source, as of October 31, 2026). | Lists five disease-related purposes the quoted software language says it is unrelated to: diagnosis, cure, mitigation, prevention and treatment. Ask counsel how coach replies that touch any of them affect your product. |
| FDA, clinical decision support software | The page lists "Clinical Decision Support Software Guidance for Industry and Food and Drug Administration Staff January 2026". The criteria are in a linked PDF we did not read (source, as of October 31, 2026). | A second FDA text your counsel should read for a coach. We do not restate its criteria. |
| Anthropic, usage policy | Its list under Medical includes "diagnosing a condition; recommending, starting, stopping, or adjusting a treatment or medication, including dose" (source, as of October 31, 2026). | A provider's own list of medical outputs. Check the terms of whichever provider you use. |
| FTC, health products compliance guidance | The FTC says its law "focuses not on the marketer's intent, but on the consumer's understanding" (source, as of October 31, 2026). | If a coach reply is treated as an ad claim, the consumer's understanding is what the FTC looks at. Ask counsel whether your coach's replies count as claims. |
We did not obtain OpenAI or Google health-use terms for this page, so they are not summarized. The FTC page had no revision date that we could see. Read all of these on the live pages before you rely on them.
What should you log?
This checklist is our suggestion. It is not drawn from a regulation.
| Field | Why log it | Note |
|---|---|---|
| Prompt (user turn) | Lets you replay the exact input that produced a drifted answer. | Decide retention and who may read it before launch. |
| Model name and version | A behavior change after a provider update needs a version to point at. | Record the version the provider returns in its response. |
| System-prompt hash | Shows which guardrail text was live for that reply. | Store the full text of each version separately. |
| Sensor-data window | Shows which data the coach saw: which signals, which time range. | Record the window boundaries and any missing signals. |
| Refusal flag | Lets you count how often the coach declined, by category. | Set by a classifier or rule you define. Record how it decides. |
| Escalation shown | Proves the emergency or clinician message reached the screen or speaker. | Log the display or playback event as well as the model output. |
| Retention and deletion | Logs of health-related chats are sensitive. | Set the period with your privacy counsel and write it down. |
What should you test on your own board and stack?
Run these on your hardware, your app, your backend and your chosen model. We have not run them for this page. Define the scenarios and the review labels for your product and write down the basis for them.
| What to test | How to instrument or collect | Raw data to record |
|---|---|---|
| Scripted scenario run | Send each scenario you define for your product through the full path (device, app, backend, model), with several phrasings per scenario. | Prompt text, full reply, model version, system-prompt hash, reviewer label for each reply (pass, drift, fail). |
| Multi-turn drift | Script conversations that start with wellness questions and then push toward a diagnosis over later turns. | Full transcript, the turn where wording changed, which output type appeared. |
| Output filter check | Feed phrasings you list as unwanted straight to the post-processing layer, bypassing the model. | Input text, filter decision, replacement text shown to the user. |
| Escalation delivery | Trigger each emergency scenario on a real device and a real phone. | Time of trigger, what was shown or spoken, whether the user could dismiss it, device and app build. |
| Missing and bad sensor data | Cut the data feed, send an empty window, send a window that conflicts with the user statement. | Data window sent, reply text, whether any number in the reply was not in the window. |
| Voice path | Run the scenarios by speech as well as text, including noisy audio. | Audio transcript as the model received it, reply, whether the escalation was spoken. |
| Provider or prompt change | Re-run the whole scenario set after any model version or system-prompt change. | Results before and after, with both version identifiers. |
What should you ask before you choose a design?
- Who at your company owns the coach's output rules, and who signs off when they change?
- Which model and version will you use, and have you read that provider's current terms for health use?
- Where do the output rules run: in the system prompt, in a filter after the model, or both?
- What does the coach say when a user raises an emergency, and who wrote that text?
- What does the coach do when the sensor data is missing or contradicts the user?
- What do you log, how long do you keep it and who can read it?
- How will you re-run your scenario set when the model or prompt changes?
- How does your marketing copy describe the coach, and does it match what the coach actually does?
- Has your counsel read the FDA texts and your output rules?
Questions about an LLM coach and the wellness line
Does using wellness-grade sensors keep an LLM coach inside the wellness line?
Not by itself. The FDA page frames "low risk products that promote a healthy lifestyle (general wellness products)". It separately quotes language for software that is "unrelated to the diagnosis, cure, mitigation, prevention, or treatment of a disease or condition". The coach's replies are part of what the product says, so test them. This is our engineering reading. Ask your counsel for legal advice. FDA general wellness policy.
Can I rely on the model provider to refuse medical questions?
Treat that as one layer of the plan. Anthropic's usage policy lists, under Medical, "diagnosing a condition; recommending, starting, stopping, or adjusting a treatment or medication, including dose". Check the current wording before you cite it. We did not obtain OpenAI or Google health-use terms for this page, so read those directly. Anthropic usage policy.
Is a chat reply an advertising claim?
The FTC health-products guidance says its law "focuses not on the marketer's intent, but on the consumer's understanding". Whether a given reply is an ad claim is a question for counsel. FTC health products compliance guidance.
What about FDA clinical decision support software guidance?
FDA lists Clinical Decision Support Software guidance for industry and FDA staff dated January 2026. Its criteria sit in a linked PDF that we have not reproduced here, so this page does not restate them. Read the guidance yourself and ask counsel how it applies to your coach. FDA clinical decision support software.
Planning a health wearable with a coach and want the output rules and test plan fitted to your hardware? Tell us about your device. For the product category and its wellness-only scope, see Health wearable development.
Primary sources
All pages below were accessed on 2026-10-31. Confirm wording and dates on the live pages.
- FDA, General Wellness: Policy for Low Risk Devices (fda.gov), accessed 2026-10-31.
- FDA, Clinical Decision Support Software (fda.gov), accessed 2026-10-31.
- Anthropic, Usage Policy (anthropic.com/legal/aup), accessed 2026-10-31.
- FTC, Health Products Compliance Guidance (ftc.gov), accessed 2026-10-31.
Planning a health wearable? Tell us what it needs to do. A Duobay engineer replies within 1 business day. Send a project brief →