How the AI works
Why a model at all
A rules engine can check a box. It cannot read “rotator cuff repair, cleared four weeks ago, still pretty tender at the top of the range” and decide what gentle movements to offer and at what dose, because that judgement lives in language, not in a conditional.
The second part is harder. Writing two coaching cues for a seated ankle circle that a person three weeks out of knee surgery can follow on a Tuesday morning, when they are tired and a little unsure, is a plain-language writing problem. That is what a language model does, and it is the part Reva is built on.
So the model is the engine. Everything else is what keeps it safe: the code screen that runs first, the schema that shapes the output, and the fallback chain that keeps a session available when one provider has a bad afternoon.
The four ideas
Reading 'knee surgery six weeks ago, stiff in the morning but no sharp pain' and knowing what kind of gentle session to build from it is language work. A rules engine can match keywords. It cannot read the sentence the way a person would, weigh the detail against what it knows about post-surgical recovery, and decide the right level. That judgement is what a modern language model does, and it is what makes Reva possible.
The model is given a system prompt with hard safety rules, in order. Screen first. If the description sounds like it needs a clinician, return no movements. If it is early or uncertain, stay very light. Only then build the session, and always within comfortable pain-free range. The output is JSON, validated against a strict schema before you see it. A session that does not fit the shape is discarded, not shown.
This is a separate routed step, because adjusting is a different job from building. The model is given the last session focus, level, and feedback, and asked to progress gently if it felt good, hold or ease if it was sore, and ease clearly with a professional note if it hurt. Adaptive sessions are on Plus and Care, because adapting requires storing session history.
Before any model call, a deterministic keyword check runs on what you described. Chest pain, trouble breathing, numbness, tingling, calf swelling, a recent fall, dizziness, and several more. This layer is code. It cannot be reasoned around. If it fires, the session stops and you are told to see a professional. The model also screens, as a belt-and-braces check, but the code screen never depends on the model to say stop.
Routing
This is not a diagram somebody drew. It is rendered from the same routing table the session engine reads, so if it is wrong here it is wrong in production.
| Step | Model | Provider | Tier | USD per M tokens |
|---|---|---|---|---|
| Build the gentle session for today | gpt-4.1-mini | openai | balanced | $1.60 |
| Adjust the next session from how the last one felt | gpt-4.1-mini | openai | balanced | $1.60 |
| Screen the described state for red flags | gpt-4.1-nano | openai | fast | $0.40 |
Longer or more involved descriptions escalate from the balanced model to the frontier one, because a description that mentions several things going on at once, or an early post-surgical state, benefits from more careful reading. Descriptions are capped at 1,200 characters in one pass.
If the routed model fails or returns nothing, the call walks a chain of candidates from other tiers and other labs before giving up. One provider having a bad afternoon should not leave a person without their session, and it does not.
The session the model returns is validated against a strict schema before you see it: required fields, string lengths, movement count, care level. A result that does not pass is discarded and reported, not shown. This is the part of the pipeline the model does not get a vote on.
Candidates
Every model below sits behind the same interface. Adding a provider is one case in one file, which is the point of building it this way.
openai · 1,000,000 token context
openai · 1,000,000 token context
openai · 1,000,000 token context
anthropic · 200,000 token context
google · 1,000,000 token context
meta · 128,000 token context
In this deployment the OpenAI and Anthropic adapters are written and the OpenAI one is live. The rest are declared with their real model names and switch on when their key is present. No model is trained on what you describe, by us or by our providers.