Blood Sugar Feedback Loops in 3 Rebounds

Blood Sugar Feedback Loops in 3 Rebounds

What Justin Harris’s coaching logs say about fast feedback, slower interpretation, and the real tradeoff between precision and noise in AI fitness coaching.

The strongest signal in the Kahunas coaching logs is brutally simple: in a rebound phase, the same insulin dose that worked one week earlier can start dipping blood sugar enough to force meal timing forward by 30 minutes, and Justin Harris responded by reducing the dose and tightening the meal interval. That is a classic feedback loop problem, not a motivation problem. In practice, the mechanism is insulin sensitivity changing faster than the plan changes around it. The falsifiable thesis here is straightforward: AI fitness coaching will be useful only when it can detect those short-cycle performance shifts and recommend the right adjustment speed; if it cannot, it becomes an expensive note-taking layer on top of bad timing.

This matters because most coaching failures are not dramatic. They are small mismatches between a plan and the body’s current state. Joe Webb’s high day example is a clean one: the same dose that had been routine the week before produced a noticeable low enough to move the next meal up. He did not “fail” the day. He did not overeat. He simply had to compress the spacing between meals and then planned to lower the dose further on the next high day. That is what good feedback looks like: a short loop, a specific signal, and a proportional response.

The coaching value is not in the signal itself. It is in the decision tradeoff the signal creates. If the dose stays the same, the plan may keep drifting into under-fueled or mistimed territory. If the dose gets reduced too aggressively, the athlete may lose the effect the high day was supposed to produce. The better move is not “be more disciplined.” It is to notice the response, adjust one variable, and preserve the rest of the structure. That is why feedback loops beat static plans in real training: the athlete’s state is moving, so the plan has to move too.

Justin’s reply in the same log also shows a second, often missed layer: performance feedback is only useful if you interpret it relative to confounders. He told the client that he was sick and holding water, so he was not going to look as good as he otherwise would, even though body fat had dropped. In other words, the visual output was noisy while the underlying trend was real. That’s a coaching problem AI systems frequently mishandle. They are good at pattern matching on surface inputs, but real training decisions often depend on separating the transient from the structural.

That distinction shows up again in the broader coaching notes. In the curated nutrition knowledge base, Harris says that if a client did all carbs as fruit, he would be surprised if the difference over a year was noticeable because most results come from nailing macros. But he also says the last few percent lives in the details, and he puts a number on it: a 5% difference over a year where you gain 10 lbs of muscle is half a pound, which sounds trivial until it compounds across years. This is the same feedback principle at a different timescale. Most of the outcome is governed by the big levers. The smaller levers matter when they are repeatedly applied and when the system is stable enough for those differences to accumulate.

That creates the practical tradeoff coaches and AI tools need to respect. If you try to optimize everything, you overload the system with false precision. If you ignore the small signals, you miss the compounding effect of repeated decisions. Harris’s fruit guidance is a useful example. Pre- and post-workout, fruit is fine up to about 50% of the carbs in those meals, with no downside. On medium days, fruit can appear in other meals too as long as it is not every meal and at least half the carbs per meal come from more complex sources. On high days, sugar barely matters because carbs are so high and insulin is elevated all day. That is not a blanket “fruit is good” claim. It is a context-dependent rule based on how the day is structured and what problem the day is trying to solve.

AI coaching gets into trouble when it treats every check-in like a universal truth instead of a state estimate. A check-in is not the same thing as a conclusion. It is a measurement with error bars. In the logs, the athlete’s blood sugar response on a high day told them something specific: the current dose was now stronger relative to the client’s condition than it had been before. That is actionable because it is tied to an immediate outcome and a concrete change. But the interpretation would be much weaker if the same number were ripped out of context and treated as a permanent rule.

There is also a behavioral reason feedback loops matter. Justin explicitly notes how easy it is to forget whether a topic has already been covered and that he repeats things a dozen times because he is paranoid clients will think he is not paying attention. That is the hidden cost of coaching at scale: human memory is imperfect, repeated topics get messy, and the coach’s job is partly to preserve continuity across many small adjustments. AI can help here, but only if it is built to retain the relevant state without turning every prior note into a rigid constraint. The goal is not to memorize everything. The goal is to surface the right prior signal when the same pattern reappears.

For coaches, the decision rule is simple enough to use tomorrow. Prefer fast feedback when the variable is acute and reversible, like meal timing on a high day or a training-day recovery issue. Prefer slower feedback when the outcome is noisy, like visual changes during illness or water retention. Keep the big levers dominant and use the smaller levers only when the system is stable enough to measure them. And do not confuse more data with better decisions. Better decisions come from shortening the gap between signal, interpretation, and adjustment.

That is the real promise of AI fitness coaching, and the real trap. The promise is a tighter loop: detect, compare, adjust. The trap is pretending that every input is equally important or equally trustworthy. The logs here point in the other direction. The best coaching is not the most data-rich coaching. It is the coaching that knows which feedback deserves a change, which feedback deserves patience, and which feedback is just noise until the next measurement.

Sources Used

• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w13-18m/transcripts/joe_webb___members-rksigkykimaxwmo_t4_e8nwvbtc2j0etleutkyysads.md
• modules/03-knowledge/kahunas-coaching-deep-nutrition.md
• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w19-24m/transcripts/rory_lazowski___members-c5balaovjbdoeefqmfuqdhh2tbpmfdu16lnf0tnrtmw.md