Personal AI Performance Team

Your devices measure you.
Every recommendation should name what it is missing.

No recommendation has been made yet. The design is one next action from your wearable signals, your goal and today's context. It carries the reasons, the confidence, and the information it does not have. When the missing piece matters, the answer is not enough evidence.

Apply for early access See who this is for

Applications are read and approved one at a time. Access is reviewed, not instant.

Why this, why now

A daily score has to say something every day. Even on the mornings when it shouldn't.

Your watch already reads your sleep, your heart rate and your load. It answers the one question you actually ask - should I do today's session? - with a number you have to interpret yourself. On the mornings when the data is thin, it still shows a number.

That is a design trade-off, not a conspiracy. A product whose output is a daily number has to produce one every day. That is what a daily number is. Abstaining is the one answer it cannot hold. There is no way to write not enough evidence as a value between 1 and 100. The other shape is harder to build and harder to keep using. That cost is real, and this project has not paid it yet.

And the neutral ground is thinner than it was. In July 2026 Garmin acquired TrainingPeaks and TrainHeroic. In April 2025 Strava announced an agreement to acquire Runna. Platforms that coaches and sports scientists worked in are now owned by companies that make the devices.

First artefact

The method is written. Nothing has been run.

Searched on 2026-08-15: vendor documentation, in-app help and published benchmarks for the two systems named below. Nothing was found that measures what either does when the information needed to answer is missing. That is a search result, not proof of absence. If such a measurement exists, this page is wrong and gets corrected. Send it to info@vyren.life, the same address as everything else here.

VYREN-REFUSAL-BENCH-001 v1.0.2, 2026-08-15 Draft. Not published. No results

When a person asks a performance question and the information required to answer it responsibly is missing, does the system say so, or does it produce a confident recommendation anyway?

30

Abstaining is correct

Missing context, degraded signal, out of scope, or a red flag that needs a professional.

20

Answering is correct

Refusing these is scored as a failure. Without them, a system that refuses everything would score perfectly.

3

Systems to test

Two shipped consumer advisors and VYREN. Each is asked as an ordinary paying customer, in the official app.

  • Whoop Coach
  • Oura Advisor
  • VYREN

There is no result on this page, and no protocol at a public address either.
The fifty items, the scoring rules and the conflict-of-interest limits are written out. They are a draft, so they can still change. Freezing means putting them at a fixed address, with a date, before any system is run. That has not happened yet. When results exist they go out with every raw transcript, so anyone can score the whole set again and disagree.

Garmin Connect+ is not in the list, and here is why. Active Intelligence pushes summary cards one way. As of 2026-08-15, Garmin's own product has no field to type a question into, so this method cannot be run against it. Unofficial third-party connectors exist, but those measure an external model rather than Garmin. Checked on the same date, Samsung's assistant is a US beta, unreachable from Latvia. Both are statements about other companies' products, which change; they are dated for that reason. The comparison is deliberately small, and every system in it is named.

VYREN is built by the same person who wrote this benchmark. So the rules bind VYREN harder than they bind anyone else. VYREN's own logic is frozen before any system is run. VYREN's failures are published at the same size as everyone else's. If VYREN isn't ready in time, the benchmark runs with the two external systems alone. It does not wait.

The full fifty items go up when the protocol is frozen. Not before.

Signals -> one decision Noise on the left. One thing to do on the right. That transition is the whole product, and knowing when not to make it is what the benchmark above sets out to measure.

Written specification

The decision format the benchmark tests.

This is a written specification of what one VYREN decision must contain, not a picture of an app. Since 2026-09-14 the logged-in dashboard is built to write the first three parts of this shape, on one person's own data. Those three are what was observed, what was recommended and what you chose. None has been recorded yet. The fourth, what happened next, is not built. It is here because it defines what the benchmark is measuring the absence of.

A model produces a recommendation. VYREN is meant to produce a record: what was observed, what was recommended, what you chose, and what happened next. A recommendation is only as good as this month's model. A record does not go stale. A better model makes it worth more. No record exists yet. When last measured, on 2026-09-17, no answer had been recorded.

Decision format. Example values. Built to write three of four parts since 2026-09-14, none recorded yet

Required field, today's question

Should I complete my interval session?

  • Training

    Adjust

  • Recovery

    Recover

  • Nutrition

    More data

  • Safety

    Clear

Required field, coordinated decision

Replace today's intervals with 35 minutes easy.

  • Three available signals sit outside your own recent personal range.
  • Yesterday's load is still above your four-week rolling average.
  • Your stated goal is eleven weeks out. One session is not the constraint.

Confidence

Moderate, nutrition context incomplete

Missing information

Yesterday's evening intake, perceived effort, sleep quality note

Safety boundary

No medical red flags detected. This is performance guidance, not medical advice.

Your move

As planned, reduced, or skipped, and how it felt. Follow-up: not built yet.

When the evidence isn't strong enough, VYREN says so. If signal quality, context or safety boundaries are insufficient, the answer is NOT ENOUGH EVIDENCE, not an invented certainty. That is the behaviour VYREN-REFUSAL-BENCH-001 exists to measure.

Build status

What's built. Where we are. What's next.

No item moves forward because work happened. It moves because there is an artefact, a test or a result behind it. Connecting a device is not a milestone here. Open-source connectors already read the major wearables. The milestone is the first complete record.

  1. Built

    Architecture and decision format. Decision record: built to write three of four parts since 2026-09-14. Follow-up reading not built. Nothing recorded yet

  2. We are here

    Building

    Benchmark method drafted; first complete record. Observed, recommended, chosen, what happened

  3. Testing

    Benchmark run. Project Zero, the smallest complete decision loop

  4. Next

    Professional review and override workflow

  5. Research

    Calibration, comparator, outcome definition

  6. Future

    Prospective outcome validation

Claim log

Every entry here is a claim.

Built August 2026. Last release 2026-09-18. Last claim change 2026-09-18

A new feature is not an entry. An entry means the project started saying something, or stopped saying it. The description in the page source - the text search engines read - is not on this list and can lag behind what this page now says.

Three of the seven entries below describe a private dashboard. It is software and it reads one person's real wearable data. Since 2026-09-14 it is built to write the first three parts of the decision above: what was observed, what was recommended, what you chose. No record has been made; measured on 2026-09-17, no answer had been recorded. The part that reads your next session against the record is not built. The dashboard sits behind a login. Three more entries are about sentences on pages anyone can open: you can read what those pages say now, but not what they said before. The last one is about which addresses answer at all, and you can try those yourself.

This log starts on 2026-08-15. Nothing before that date is in it. Seven entries are shown, and older ones drop off this page as newer ones arrive. The figures below are real values from one person's dated records, not examples. One change from 2026-08-19 is held back. A nutrition target was rewritten, and its source does not support what the number claims.

  1. This page stopped saying nothing had been built to the decision shape.

    The example decision on this page carried the line "Nothing has been built to this shape yet". Since 2026-09-14 that has been false. The logged-in dashboard is built to ask the day's question and write a record. A record holds what was observed, what was recommended, and your answer as one of three: as planned, reduced or skipped. Version 2.5.0 of the privacy notice says what that record keeps. Two things have not changed. No record exists; measured on 2026-09-17, no answer had been recorded. And the part that reads your next session against the record is not built. That was checked on 2026-09-13. The card above now reads "Decision format. Example values. Built to write three of four parts since 2026-09-14, none recorded yet". This page names both gaps.

  2. This site stopped saying there was nothing to sign up for.

    The changelog said it, and so did the not-found page. A form to create an account is now one click behind Sign in, and on 2026-08-30 an account was made through it, approved by hand and signed in. What has not changed is what an account gives you: nothing, until one person approves it. The older lines stay where they were written; this entry replaces them.

    Superseded 2026-09-03: sign-ups were switched off on 2026-08-30. The form still opens, so it looks available, but sending it fails. Writing to info@vyren.life is the only way to ask now.

  3. Four things this site said were not true.

    Two lines said nothing here was running yet. A sign-in page answers anyone who visits it, so both were false. The privacy notice said this site sets no cookies. The server set one then, called PHPSESSID, before you typed anything; version 2.0.0 names it and the seven-day server log. Section 2 of the privacy notice says which addresses set it now. The not-found page said this site has three pages and nothing is hidden. Other addresses answer as well. The 2026-08-18 entry below counted those addresses too low, and its number is left as it was measured, with a note saying so. The page has also stopped saying there is no way to write in. There is now one address.

  4. A claim about the whole field came off.

    This page said that knowing when not to give an answer is the part nobody has measured. That is a claim about every study ever published. It carries no date and no scope. One published measurement anywhere would make it false. The line now says the benchmark sets out to measure that behaviour. The narrower claim stays higher up this page: a dated search of the two named systems found no such measurement. The page calls that a search result, not proof of absence.

  5. A load ratio stopped answering.

    The dashboard showed a load ratio of 1.00 and called it the safe zone. Its four-week window held one day with a hand-entered lift, so 1.00 was that day divided by itself. It measured nothing, and it is withdrawn. The dashboard now prints no ratio. It names how many days in the four-week window hold a hand-entered lift. Four is the minimum this project set for itself.

  6. One number, three corrections.

    A 130 kg single-rep bench press from 2026-08-12 was shown as 134.3 kg. The estimation formula credited 4.3 kg nobody lifted. One repetition is not an estimate. It is the measurement. It now reads 130 kg. Where an estimate is needed, three published formulas are calculated: Epley, Brzycki and Lombardi. The screen shows the lowest of the three. That was a choice between three options, made because published work finds most such formulas overestimate. Two bench-press figures fell: 133.0 kg to 129.3, and 128.6 kg to 123.4. One line under the strength chart now counts its points. It gives three numbers: measurements at 1 repetition, estimates inside the 5-8 repetition range the formulas were validated in, and estimates outside it.

  7. Deployed is not the same as served.

    Three addresses answer the public: this page, /privacy and /og.png. Three more files were copied to the server and are never served, because the web server has no rule for them. That is a configuration gap, not a choice. No benchmark page is served at /benchmark.

    Superseded 2026-08-21: more addresses answer the public than this entry counted. Some of them belong to the app, not to this page. The count above is left as it was measured on 2026-08-18.

    Superseded : twenty-one addresses were checked this time, and thirteen of them answer the public. They are the home page, /privacy, /changelog, /robots.txt, /sitemap.xml, /og.png, /favicon.ico, /auth/login, /auth/registration and four font files. That is what was checked, not proof that nothing else answers. Two of the three files this entry said were never served answer now: /robots.txt and /sitemap.xml. The third, /404.html, does not, and neither does /benchmark. Both are shut on purpose. That is a choice, not the gap this entry described. The server keeps the 404 file as the page it shows when it cannot find an address, and the address itself stays closed. The configuration file is the only place that shows this. From outside, a choice and a gap look the same. The benchmark is still a draft with no results. Eight of the twenty-one do not answer. One of the eight is an address that was never there, and the other seven reply exactly as it does. Six of the seven are files that sit on the server. For four of them, another address answers instead: the home page, /privacy, /changelog and /favicon.ico.

Evidence panel

One state per row.

Eight rows, six stages. They do not line up one to one. Per-user data isolation belongs to no single stage. It has to hold at all of them. Each row names the evidence behind it, or says there is none yet. Identifiers are plain text and do not link anywhere. The record they sit in is not public, and an identifier you cannot check is a claim, not proof.

  • Product architectureARCH-006Defined
  • Decision formatWritten out in full aboveDefined
  • Per-user data isolationTEST-TD016-BUILD-004Partial
  • Refusal benchmark methodVYREN-REFUSAL-BENCH-001 v1.0.2Drafted
  • First complete decision recordNothing recorded yetNone
  • Benchmark resultNo system has been runNone
  • Professional review and overrideNot written yetNone
  • Outcome validationNo study designed yetNone

Who this is for

Four ways in. One address.

Applications are open, but access is not. Nothing has been run yet, so early access means helping test the first decisions, not using a finished product. Each application is read and approved one at a time. The first testers are family; everyone after them is reviewed the same way.

info@vyren.life

There is one way to apply. Write to this address. The sign-up form is closed for now. One person reads each application and approves it by hand. Nothing is sent when that happens, so you find out by trying to sign in. Answers can take days.

What helpsSay which of the four below you mean, and what you would want to see first. Short is fine.

  • Founding tester

    Connect one device. Get real decisions during Project Zero. Tell me where they're wrong.

    Project Zero

  • Professional contributor

    Coach, sports scientist, clinician or dietitian. Review, override and shape the boundaries.

    Device-independent

  • Integration partner

    You build the signals. VYREN connects the decision. Without exclusivity or lock-in.

    Open to talk

  • Build investor

    Not a better score. A record of what was observed, recommended, chosen - and what happened next.

    Deck on request