Approval Gates and the 1-iu Recomp

Approval Gates and the 1-iu Recomp

Why AI fitness coaching should be constrained by human sign-off, not handed full autonomy

The clearest operational lesson in the Kahunas coaching logs is simple: the same insulin dose that was fine one week caused Joe Webb’s blood sugar to dip the next week, forcing meal 2 30 minutes sooner and prompting a 1 iu reduction. That’s not a software bug; it’s a moving target problem, and the underlying mechanism is feedback drift. If the target is changing session to session, a fully autonomous coach will eventually be wrong in exactly the places where small errors matter most, so AI coaching should be approval-gated by default, not autonomous by default.

That thesis matters because fitness coaching is often sold as an automation problem. Log the food, read the wearable, predict the next move, and let the system run. But the real world in these sources looks less like a stable control loop and more like a series of exceptions: appetite changes, insulin sensitivity shifts, illness, fatigue, body-composition changes, and repeated reminders that the coach is still the one interpreting context. The lesson is not that AI should do nothing. The lesson is that the machine should draft, flag, summarize, and compare, while a human approves the part that changes the plan.

What the logs actually show

In Joe Webb’s case, the sequence is instructive. He reports that on a high day, the same insulin dose as the prior week “dipped my blood sugar a little bit noticeably,” not a full hypo, but enough that meal 2 had to come 30 minutes sooner. He reduced the dose by 1 iu and repeated the shot later in the day, and the same issue happened again. The adjustment was immediate and specific: reduce the dose further on the next high day. That is textbook approval-gated territory. The useful part of the system is not “make the change automatically.” It is “detect the pattern, surface the deviation, and hand the decision back to the coach.”

Now compare that with the Rory Lazowski thread in which Justin Harris says he tried retatrutide himself so he could have firsthand experience for clients. After taking 2 mg, he reported “no appetite whatsoever,” more fatigue than normal, and uncertainty about where calories should go if food is added back in. He also said he would pause or reduce the dose if the plan moved toward more food. Again, the point is not that the compound is magic or dangerous; the point is that even an experienced coach can feel the planning consequences immediately, and the right response is not unconditional automation. It is bounded experimentation with a human deciding whether the plan is still appropriate.

These are not abstract anecdotes. They are examples of coaching under volatility. Appetite suppression can make food planning feel easier for a week and harder the next. A small change in sensitivity can turn a previously workable dose into an operational nuisance. A client being sick can hide body-composition changes under water retention. A coach can be “hyper paranoid” about missing something and still need to repeat the same point a dozen times because memory, not compute, is the bottleneck. Autonomy doesn’t remove those problems; it just hides them until they show up as bad recommendations.

Why autonomy should be constrained

The case for approval-gated automation is not anti-technology. It is pro-accountability. In coaching, the expensive failures are usually not dramatic. They are small, confident mistakes: a dose nudged too far, a food change that seemed harmless, a check-in interpreted without context, a client assumed to be “off” when they are just holding water, or a recommendation made on stale data. Once a system is allowed to act without review, it stops being a tool and starts being an author. That is the wrong role for a model trained on messy logs.

Approval gates solve for three specific failure modes.

First, they force context recovery. A model can notice that blood sugar dipped; it cannot know whether that was training load, food timing, illness, or a dose that now needs a smaller increment. In Joe Webb’s log, the decision was not merely to “change insulin.” It was to bring meals closer together on that day and reduce the dose further on the next high day. Those are context-sensitive choices, not generic outputs.

Second, they preserve reversibility. In the Rory case, Justin’s firsthand trial of retatrutide was useful precisely because he was still evaluating it. He was not handing the system control and hoping for the best. He was gathering experience to inform client decisions. Approval-gated workflows keep the human in the loop where the plan can still be paused, reduced, or discarded before the change becomes entrenched.

Third, they reduce false confidence. The community sentiment KB is blunt about TroponinIQ itself: mixed, educational, but criticized for memory issues, contradictions, and inconsistent advice. That matters because the technology stack is only as good as the governance around it. If a coaching assistant has memory problems, then letting it execute changes automatically is a category error. A system that sometimes contradicts itself should not be allowed to behave like a decision engine without review.

What an approval-gated workflow should look like

A useful AI coach should do four things well and stop there.

1. Detect deviations. It should notice that a high-day dose produced a different blood sugar response than last week, or that an appetite signal changed sharply after a new intervention.

2. Summarize context. It should pull the relevant facts into one view: last week’s dose, today’s timing, what changed, what happened after the change, and what the coach already said.

3. Propose options, not actions. It should draft a few human-readable choices, such as reducing the dose, moving meal timing, or holding steady until the next check-in.

4. Require sign-off. The final action should wait for coach approval, with the system recording what was approved and why.

That structure is slower than full automation, but that is the point. In coaching, speed without context creates expensive noise. A slightly slower workflow that requires approval is usually cheaper than a faster one that quietly drifts away from reality.

The falsifiable standard

Here is the test I’d use: if an AI coaching system cannot be trusted to know when not to act, it should not be trusted to act autonomously. That is falsifiable. Show me a model that consistently handles changing sensitivity, fluctuating appetite, illness, and memory constraints without meaningful correction, and the case for wider autonomy gets stronger. Until then, the evidence here supports a narrower design: AI as analyst, human as approver.

That is also the most coach-friendly design. It protects the craft, preserves accountability, and keeps the system useful when the plan is moving under its feet. The best automation in fitness is not the kind that replaces judgment; it is the kind that makes judgment easier to exercise. Approval gates do that.

Sources Used

• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w13-18m/transcripts/joe_webb___members-rksigkykimaxwmo_t4_e8nwvbtc2j0etleutkyysads.md
• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w19-24m/transcripts/rory_lazowski___members-c5balaovjbdoeefqmfuqdhh2tbpmfdu16lnf0tnrtmw.md
• raw/_TROPONIN_SENTIMENT/troponin_community_sentiment_kb.md