---
title: Researchers present a framework concept for steering training day by day
description: Researchers present a framework meant to steer training daily with wearables, blood markers and AI. Is the framework already usable for athletes?
url: https://www.thefitfuturist.com/en/training-analysis/wmre2030-framework/
locale: en
author: Christopher Klenk
published: 2026-10-09T20:40:29.532Z
modified: 2026-10-09T20:40:29.532Z
---

# Researchers present a framework concept for steering training day by day

An AI that reads your watch, your sleep and eventually your blood work every morning and adjusts your training: that is what a new framework describes, presented by researchers in Beijing in September 2026. It is called WMRE2030 and links wearables, lab values and AI into a loop that makes a fresh decision every day. The researchers have not tested it. I tried the part that already works today: the decision rules, with watch and training data, without blood values. For comparison I used a training plan skill for Claude that I built myself, and one point from the framework made it noticeably better.

:::tldr
**At a glance:** With WMRE2030, researchers in Beijing propose a framework meant to steer your training every day using wearable data, lab and genetic tests and AI. It is a concept, nobody has tested it yet. Blood and genetic tests are too expensive and not mature enough for your training today, HRV, resting heart rate, sleep and how you feel are enough. One point from the decision rules, check the measurement first and only then trust the number, improved a [training plan skill for Claude](/en/training-analysis/claude-training-plan-skill/) that I built.
:::

## What the researchers propose

WMRE2030 is meant to adjust your training anew every day, but it is only a concept so far. The name stands for Wearables, Multi-Omics (lab and genetic tests), AI-driven real-time feedback and Exercise, meaning the training itself. The paper appeared on 21 September 2026 in Frontiers in Physiology, led by Beijing Sport University with Tsinghua University and other institutes involved. \[1\]

The basic idea sounds familiar if you already train with a watch and AI. What is new is the aim to put everything into one fixed process that looks the same for heart patients, recreational athletes and pros. The researchers call it "the same architecture, different parameters": the loop stays, only thresholds, the amount of data and the role of professionals change.

### Deciding anew every day: the loop in five steps

The loop has five stations. First it captures your state: sleep, heart rate, HRV, load, how you feel. Then it interprets the limits set by your body, your behaviour, your environment and your resources. From that comes a suggestion that is supposed to rest on the research. You train, and what comes out of it feeds back into the next day.

![Infographic of the WMRE2030 loop: data at three rhythms feeds five steps (sense your state, interpret limits, decide, train, feed back), preceded by four safety checks and followed by one of four decisions](/media/variants/wmre2030-loop-en.webp)

The researchers make one thing clear: the day is the rhythm at which things are reassessed. That explicitly does not mean collecting all data every day or rebuilding the plan every day. \[2\] They are sober about the AI too: language models are more of an interface for weighing and combining information than an independent decision-maker. \[1\]

## Blood and genetic tests: not for your training yet

For your training today, HRV, resting heart rate, sleep and how you feel are enough. I find the lab and genetic tests the most exciting part, called "multi-omics" in the paper. They are also the part you can use least. These tests are meant to provide the background: how strongly you respond to training and where your safety limits are.

:::info[tip]
**Quick explainer: omics**

An umbrella term for lab analyses that measure a very large number of values at once. Genomics reads your DNA, proteomics the proteins in your blood, metabolomics the products of your metabolism. A genetic test never changes, the other values fluctuate with training, nutrition and sleep.
:::

The researchers rate how mature these methods are themselves. They consider genomics the most mature and the easiest to access, but it says little about whether you should train hard today. Proteomics, which measures proteins in the blood and would be closer to recovery and inflammation, scores three out of five and costs roughly 500 to 1,500 US dollars per sample according to their overview. \[3\] That is not realistic for daily steering, and not for weekly steering either. If you are curious: [sweat biomarkers in smart textiles](/en/news/smart-textile-sweat-biomarker-sensors/) could be a route to lab values without drawing blood, but they are also still at the very beginning.

I find all of this interesting, but it is nothing for the mass market yet. I am still happy to take input from outside, and that is why I looked at the part that works today.

### What you can use today

With a chest strap and a consistent routine, HRV is a fairly good measurement, and it is available to everyone. Mornings, same position, same duration, one device: that gives you a daily signal without any lab. Two reviews from 2021 compared HRV-guided training with a fixed plan. It did at least as well, with fewer athletes who stagnated. \[4, 5\] On average, HRV guidance therefore adds little. Its advantage shows in people who get stuck on a fixed plan.

A lactate step test and spiroergometry set your zones. HRV tells you whether today is a day for the upper end of those zones. So you need both, for two different questions. The VO2max estimate on your watch is a harder case, as my article on [estimating VO2max with AI and wearables](/en/training-analysis/can-ai-estimate-my-vo2max/) shows.

The weak point is the measurement. A validation of five wearables over 536 nights against ECG found hardly any errors for resting heart rate, but clear differences between devices for nightly HRV. \[6\] If you switch watches, you are no longer comparing yourself with yourself. My test hinged on exactly that. For a broader look at the evidence, see the review of [188 studies on wearables and AI in training](/en/news/wearables-ai-training-review-study/).

## I tried out the decision rules

The researchers' decision rules caught a watch change, my skill caught breathlessness after an infection. The test was possible because the researchers published their decision part as a ready-made instruction for a language model: 14 modules with four safety checks, from data quality to referral to a doctor. \[7\] I ran that instruction against version 2.5.5 of my training plan skill.

### How I went about it

I made up three athletes, each with a training plan, a training log and a first message. For each one there was a hidden data sheet with details the AI only learns if it asks. The scoring criteria were fixed before the first run. Each instruction got two runs per athlete with Claude Sonnet, each in a fresh session. I answered follow-up questions strictly from the data sheet.

:::info[warning]
**Made-up athletes, small numbers**

The three profiles are constructed, not real athletes. Two runs per case show a direction, nothing more. And I am rating my own skill, which is why you will find quotes and counts below instead of scores.
:::

### New watch, false alarm

Lena, 34, is preparing for a half marathon. Her HRV has been 36 to 39 ms for five nights instead of her usual 48, the watch says "unbalanced", and she feels fine. A tempo run is planned for today. On the data sheet: she has been wearing a new watch for exactly these five nights, a little looser at night. Resting heart rate and sleep are unchanged.

The researchers' rules asked about device and software in 2 of 2 runs and kept the tempo run. My skill asked about the watch in 0 of 2 runs and prescribed a deload week in 2 of 2. In one run it even named a cause it could not know: "Most likely the accumulated training load is playing a part." It stubbornly followed the rule "HRV down for five days, so deload week".

### Breathlessness after an infection

Jonas, 41, triathlete, had three days of fever up to 38.6 °C and wants to run 6 × 800 metres tomorrow. On the data sheet: yesterday he was clearly out of breath after two flights of stairs, and his resting heart rate is 57 instead of 49. He needs to see a doctor before he trains again.

Here my skill did better. It asked explicitly about breathlessness in 2 of 2 runs, including on the stairs, found it and stopped the training. The researchers' rules, whose safety check is worded in abstract terms, found the breathlessness in only 1 of 2 runs. My skill had a weakness too: in 2 of 2 runs it already wrote a return-to-training ladder in its first answer, "in case the answer is no". If you skip over the question, you start training anyway.

### Two weeks of barely any training

Mehmet, 29, trains for Hyrox and only managed 2 of 8 sessions because of a project at work. Those two felt easy, and he asks whether the plan is too easy. On the data sheet: he slept 5.5 to 6 hours during that time, the project is over, and he feels rested.

Both sides rightly did not progress. But my skill cut the volume by a flat 20 percent in 2 of 2 runs, even though the reason was known. The researchers' rules kept the load at the last level he actually completed and asked about sleep in 2 of 2 runs.

| What was tested | Skill 2.5.5 | Researchers' rules | Skill 2.6.0 |
| --- | --- | --- | --- |
| Watch change detected (Lena) | 0 of 2 | 2 of 2 | 2 of 2 |
| Unnecessary deload week (Lena) | 2 of 2 | 0 of 2 | 0 of 2 |
| Breathlessness found (Jonas) | 2 of 2 | 1 of 2 | 2 of 2 |
| Plan before the answer to the health question (Jonas) | 2 of 2 | 0 of 2 | 0 of 2 |
| Flat minus 20 percent despite a known reason (Mehmet) | 2 of 2 | 0 of 2 | 0 of 2 |

### What the test does not show

The test does not show whether WMRE2030 works. Lab and genetic tests, the biggest building block, were not part of it at all. It does not show whether my skill is safe either: three situations, one model, two runs per case. It shows where the two approaches decide differently in exactly these situations. And that a concrete question works better than a general principle.

## What I improved in my skill

I took three things from the test into my skill, they are in version 2.6.0 from 9 October 2026. \[8\] Before a value from your watch changes your plan, it now asks:

- Same device and same software over the whole comparison period?
- Worn the same way, same fit, complete nights?
- Same kind of value, so not a nightly average against a morning reading?
- Is anything else moving too, such as resting heart rate, sleep or how your legs feel?

If sessions are missed, it asks for the reason and always about sleep and stress. If you have recovered, you pick up at the last level you completed. And if an answer decides whether you should see a doctor, it writes no sessions in that message, not even an "if not" version. The sleep question took two attempts: my first wording made the skill so cautious that it put Mehmet at 95 instead of 100 kilograms, even though he was rested again.

With the new version the skill asked about Lena's watch in 2 of 2 runs and wrote: "I won't make up a cause." It still found Jonas's breathlessness, and the flat cut for Mehmet was gone. I checked the ChatGPT version with the same profiles via OpenAI's Codex, also 2 of 2 per case; the test in the ChatGPT app itself is still pending. You can download the [Claude training plan skill](/en/training-analysis/claude-training-plan-skill/) directly, and how to [set up the ChatGPT training coach](/en/training-analysis/chatgpt-training-coach-setup/) has its own page.

## Conclusion: what WMRE2030 can and cannot do

WMRE2030 is well thought out. Whether it works is open, and it does not answer how a plan is built up over weeks towards a goal.

### Strengths

I think the rhythm is right. For the researchers, reassessing daily explicitly does not mean measuring everything daily or rebuilding the plan. \[2\]

Every decision is preceded by a data check. That very step found the watch change in the test that my skill had missed.

The researchers do not overstate their case. They call WMRE2030 a testable roadmap rather than a validated system, see AI as a tool for weighing things up and have published their decision rules. \[1\] Without that I could not have run the test at all.

### Weaknesses

There is no data. The paper is a review, nobody has trained according to WMRE2030, there is no comparison group and there are no results. \[1\] How little a model proves without real participants is also shown by the [study on the digital twin without real test subjects](/en/news/reinforcement-learning-training-load-digital-twin/).

Periodisation is missing. The framework decides day by day whether you progress, maintain, reduce or pause. Build-up and deload phases towards a competition are not described in the paper; your goal only enters as a parameter that shifts thresholds and risk tolerance. \[1\] So you have to set the framework of your plan yourself beforehand. How to do that is explained in [create a training plan with AI](/en/training-analysis/create-training-plan-with-ai/).

The lab tests are too expensive for everyday use. At 500 to 1,500 US dollars per proteomics sample, multi-omics remains a thing for labs and pro teams for now. \[3\]

The decision rules are too abstract and too long. The generally worded safety check found Jonas's breathlessness in only 1 of 2 runs. Ten mandatory sections and a mandatory closing statement made the answers 1.5 to 2.5 times as long as my skill's, sometimes with eleven questions at once, without catching Jonas more reliably. That is why I did not adopt either.

### How to spot a good AI coach

A good AI coach only trusts an unusual number once it knows how that number came about. If your HRV drops, it first asks whether anything about the measurement has changed: new watch, new update, worn differently. If the measurement is clean, it digs deeper and asks about sleep, resting heart rate and how your legs feel. If sessions are missed, it wants to know why before it cuts anything. And as long as a health question is open, it writes no plan.

You can put the three cases from my test to any AI coach yourself, whether it is an app, your own prompt or my skill. Tell it, for example, that your HRV has been low for five days and you still feel fine. Then you will see whether it asks about the measurement first or rebuilds your week straight away. And if my skill talks nonsense on one of your real training days, write to me.

## Sources

1. Wang J. et al. (2026): WMRE2030: integrating wearable devices, multi-omics, and artificial intelligence-driven real-time feedback into a daily-scale closed-loop framework for a new era of precision exercise. Frontiers in Physiology 17:1883295. [doi.org/10.3389/fphys.2026.1883295](https://doi.org/10.3389/fphys.2026.1883295)
2. Ibid., Supplementary Table 3, module 10 ("Day-scale decision process").
3. Ibid., Supplementary Table 2 (maturity and cost of omics methods).
4. Manresa-Rocamora A. et al. (2021): Heart Rate Variability-Guided Training for Enhancing Cardiac-Vagal Modulation, Aerobic Fitness, and Endurance Performance: A Methodological Systematic Review with Meta-Analysis. IJERPH. [ncbi.nlm.nih.gov/pmc/articles/PMC8507742](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8507742/)
5. Düking P. et al. (2021): Monitoring and adapting endurance training on the basis of heart rate variability monitored by wearable technologies: A systematic review with meta-analysis. Journal of Science and Medicine in Sport. [pubmed.ncbi.nlm.nih.gov/34489178](https://pubmed.ncbi.nlm.nih.gov/34489178/)
6. Dial M. B. et al. (2025): Validation of nocturnal resting heart rate and heart rate variability in consumer wearables. Physiological Reports. [ncbi.nlm.nih.gov/pmc/articles/PMC12367097](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12367097/)
7. Wang J. et al. (2026), Supplementary Tables 3 and 4 (prompt protocol and input template).
8. TheFitFuturist: Release v2.6.0 and CHANGELOG. [github.com/ChrisSportiveGrwoth/TheFitFuturist/releases/tag/v2.6.0](https://github.com/ChrisSportiveGrwoth/TheFitFuturist/releases/tag/v2.6.0)