LIM is Built on Evidence, Not AI Hype

Last Thursday, 3 AM, third shift. I am watching a patient on suicide precautions when my phone buzzes. Hacker News, number one post: "I believe there are entire companies right now under AI psychosis." 1,359 points. 669 comments.

The irony is not lost on me. I run security at a psychiatric hospital. I run an AI fitness coaching company on the side. I lost 112 pounds — 308 down to 196 — without an influencer deal, a supplement sponsor, or a pre-workout cult. And here is the front page of the internet telling me what I already knew: the AI industry is hallucinating in production, and the people running it are starting to believe their own hallucinations.

So let me say this plainly. Legacy In Motion is the opposite. Here is the four-point case.

The anti-hype position

Most AI-fitness products right now run on three things: a chatbot wrapper, a stock photo of a six-pack you will never have, and citations the model invented at 2 AM. We do not run on any of those.

We run on evidence. And we run on a guy who actually did the work.

Proof 1 — The citation purge

On 2026-05-18 — earlier today, actually — we pulled 13 posts off the LIM blog for forensic review. The model that drafts our nutrition explainers had a habit of producing citations that looked right. Right journal. Right author. Right shape of conclusion. Wrong year. Sometimes a wrong journal entirely. The kind of error a credentialed reader catches in 20 seconds and a casual reader never catches at all.

We caught it. We fixed it. Specifically:

• A claim attributed to "Singh 2026, *Cell Metabolism*" was actually **Singh et al., 2022, *Cell Reports Medicine*** — DOI [10.1016/j.xcrm.2022.100633](https://doi.org/10.1016/j.xcrm.2022.100633). Real paper. Real year. Real journal. Adjacent enough that the model felt confident. Wrong enough that it was a lie.
• A cardiovascular-risk number got corrected to the **2026 ACC/AHA Guideline**, *Circulation*, DOI [10.1161/CIR.0000000000001423](https://doi.org/10.1161/CIR.0000000000001423).
• A claim about post-treatment cardiovascular outcomes was rerouted to **Meernik et al., 2026, *JACC***, DOI [10.1016/j.jacc.2026.02.5122](https://doi.org/10.1016/j.jacc.2026.02.5122).
• A physical-activity dose-response claim now anchors to **Ekelund et al., 2026, *The Lancet***, DOI [10.1016/S0140-6736(25)02219-6](https://doi.org/10.1016/S0140-6736(25)02219-6).

One post got killed outright. We could not find a real source that supported what the model had written, so the post is gone. No quiet edit. No silent rewrite. Gone.

If you have ever wondered why a fitness brand needs forensic audit discipline — this is why. Because the alternative is shipping confident lies to people who are trying to change their bodies.

Proof 2 — Brand safety is structural, not aspirational

Every message any of our coaching bots sends to a client passes through a content-gate before it lands in their inbox. Not a vibes check. A piece of middleware in front of the send endpoint that scans the outbound payload for two specific failure modes: distorted eating discourse (DED — language patterns that mirror disordered-eating cues) and scaffold leaks (the model accidentally exposing its system prompt, its tool calls, or its internal reasoning).

If either fires, the message is held, logged, and routed to me for review before the client ever sees it. Not "we promise we trained the model carefully." A pipeline. With an off-switch. Logged.

You cannot build this on top of a chatbot wrapper. You have to mean it from the architecture up.

Proof 3 — The photo on the transformation post is me

There is exactly one rule we do not negotiate on first-person transformation content. The photo is real. My photo. Before and after. 308 and 196. Same human, separated by years of work, lifting heavy things, eating in a deficit my hospital cafeteria did not make easy, and sleeping when third shift would let me.

We do not generate AI hero shots for transformation posts. We do not composite a fake "after." We do not even use AI to retouch the "before," because the "before" is the whole point — a guy who looked like a lot of you, before he did not.

The industry standard for AI-fitness right now is a Midjourney render of a person who has never existed, captioned with a meal plan a model wrote at 2 AM. That is the version of fitness that earned the Hacker News headline. It is not ours.

Proof 4 — I am not an influencer

I work third shift at a psychiatric hospital. Security supervisor. I have a badge and a real schedule and patients whose names I will not say. I did not lose 112 pounds in a studio. I lost them in a hospital break room, in a parking-lot gym at 7 AM after a 12-hour shift, in a kitchen with a scale and a notebook, over years.

That is the credential. Not a follower count. Not a verified check. Not a brand deal. A long, boring, documented outcome that I did to myself with my own two hands, and that I now know how to help other people repeat — because I had to figure it out the slow way first.

What this means if you are considering LIM

If you came here for an AI fitness app that promises a 30-day six-pack and tells you what you want to hear, we are not that. The model will tell you when your protein is short. The plan will get harder when you adapt. The citations under every claim will be real ones. If we ever catch ourselves shipping a fake one, you will get a post like this one — not a stealth edit at 3 AM.

That is the position. Evidence over hype. A real coach behind the AI, not a render. A pipeline that catches its own mistakes before you see them.

If that sounds like the kind of operation you want in your corner for the next year of your life, the 30-day free trial is open. No card up front. Cancel any time inside the app.

Come build something real with us.

— Jake Founder, Legacy In Motion

---

The data behind this

• Singh et al., 2022, *Cell Reports Medicine*, DOI 10.1016/j.xcrm.2022.100633.
• 2026 ACC/AHA Guideline, *Circulation*, DOI 10.1161/CIR.0000000000001423.
• Meernik et al., 2026, *JACC*, DOI 10.1016/j.jacc.2026.02.5122.
• Ekelund et al., 2026, *The Lancet*, DOI 10.1016/S0140-6736(25)02219-6.
• LIM internal: 2026-05-18 cite-audit pass; 13 quarantined posts; 1 deleted; content-gate middleware on `/client/app-message` + `/bot/send`.