Approval Gates and 2-iU Insulin: Why Coaching Autonomy Needs a Brake

Approval Gates and 2-iU Insulin: Why Coaching Autonomy Needs a Brake

The useful AI coach is not the one that acts first; it is the one that can only act after a human has approved the plan, the dose, or the escalation.

The clearest pattern in the Kahunas coaching logs is not that more automation improved decisions; it is that repeated human review prevented the kind of overreaction that comes from acting too fast. In one exchange, the coach explicitly notes that coaching is hardest because the same topics repeat, and that he repeats things “a dozen times” to make sure clients do not feel ignored. In another, a client reports a 1 IU insulin reduction that still brought his next meal forward by about 30 minutes, prompting another adjustment on the next high day. That is the mechanism here: human review as an error brake. The falsifiable thesis is simple: in fitness coaching, the highest-value AI workflows will be approval-gated by default, because constrained autonomy beats fast autonomy whenever the system is managing changing bodies, noisy inputs, and high-cost mistakes.

The strongest signal is not “AI can coach”; it is “coaching is full of corrections”

The raw exchanges show a boring but important reality: execution changes are often tiny, local, and reversible. Joe Webb’s check-in is a good example. He did his high day, noticed the same insulin dose as the week before seemed to dip blood sugar more than expected, reduced the dose by 1 IU, and still had to bring the next meal forward. The correct response was not a grand automated optimization. It was a small correction, followed by another correction on the next high day.

That is exactly the sort of environment where autonomy should be limited. The system is not solving a stable puzzle once. It is revising a moving target under imperfect information. In that setting, an AI that is allowed to “just handle it” is more likely to amplify a small miss than to produce a better outcome.

This is why approval gates matter more than polish. The value is not in the model making a recommendation faster. The value is in forcing a human to say: yes, this change is worth making, and yes, we understand the downside if we are wrong.

Why approval gating beats full automation in coaching

Fitness coaching has three features that make unconstrained automation risky.

First, the inputs are noisy. Scale weight, photos, appetite, fatigue, and training performance all bounce around. Justin’s reply to Rory Lazowski makes that visible: a client can be leaner and still look worse when sick or holding water. If a system auto-reacts to every poor-looking photo, it can chase noise instead of trend.

Second, the stakes are uneven. Some changes are low consequence, like shifting meal timing or swapping carb sources. Others are not. The logs show a coach being cautious even around appetite suppression and fatigue after a 2 mg retatrutide trial, saying he was not sure he liked the idea of forcing appetite lower and might pause or reduce the dose if food increased later. Whatever one thinks of the substance, the coaching logic is clear: when a variable can change appetite and fatigue, the response should not be delegated to an unreviewed automation chain.

Third, the best coaches are already running a “human override” model. They notice patterns, interpret context, and then decide whether to adjust. That is not because humans are magically better at every subtask. It is because humans are better at knowing when the system is being fooled.

Approval-gated workflows are not slower if they are designed right

A bad version of approval gating is bureaucratic: every minor change needs a committee. That is not the model worth defending.

The better version is narrow and tactical. Let the AI draft, summarize, flag, and rank. Let it detect mismatches: a client says energy is down, but body weight is stable; photos look flat, but the client is recovering from illness; insulin timing changed, but the actual issue might be meal spacing rather than dose. Then require approval before any action that changes the plan.

That division of labor is useful because it respects where automation is strongest and where it is weakest. AI is good at:

• summarizing a check-in
• surfacing deviations from prior weeks
• reminding the coach that a client previously responded to a similar issue
• drafting conservative next-step options

AI is not, by itself, good at deciding whether the next step should happen today, wait a week, or be ignored as noise.

If you build the workflow around approval, you preserve the speed advantage without surrendering judgment.

The real failure mode is confident autopilot

The community sentiment KB is blunt about the broader product category: TroponinIQ is viewed as educationally useful but criticized for memory issues, contradictions, and inconsistent gear advice. That combination is exactly why the daily angle matters. The danger is not that AI is useless. The danger is that people confuse fluency with reliability and then let the tool act like a coach instead of a draft assistant.

In practice, the bad workflow looks like this: a model sees a data point, predicts a trend, and outputs a plan that gets executed immediately. The coach only checks later, after the client has already absorbed the consequence.

The better workflow is the reverse: the model produces a proposed action, flags uncertainty, and waits. That wait is not wasted time. It is the safety feature.

What to gate, and what not to gate

Not every AI action needs the same level of control. Coaches should be selective.

Approval-gate anything that changes a client’s plan:

• calorie targets
• meal timing structure
• training volume jumps
• cardio increases
• any escalation that could compound a misunderstanding

Do not over-gate low-risk clerical work:

• check-in summaries
• message drafts
• trend visualizations
• reminders about prior decisions

The rule is simple: if the output can materially change what the client does next week, a human should approve it before it goes out.

That rule scales better than trying to predict every edge case. It also makes the system easier to audit. When something goes sideways, the coach can inspect the proposed action, the evidence used, and the approval that allowed it through.

The coaching standard should be constrained autonomy by default

The logs from Justin Harris do not read like a story of automation winning. They read like a story of attentive iteration: blood sugar expectations discussed in advance, recomp changes interpreted in context, appetite changes observed, and small adjustments made as needed. That is not glamorous, but it is what durable coaching looks like.

The AI version of that should not be “hands off.” It should be “hands on, with assistance.” The model drafts. The coach decides. The system can move quickly, but only after a human has checked the recommendation against the person in front of them.

That is the falsifiable bet worth making: if approval-gated workflows are implemented well, they will outperform autonomous coaching systems on safety, trust, and consistency in real-world fitness practice. If they are not, the failures will show up where coaching always fails first — in the gap between a plausible recommendation and the wrong one.

Sources Used

• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w13-18m/transcripts/rory_lazowski___members-c5balaovjbdoeefqmfuqdhh2tbpmfdu16lnf0tnrtmw.md
• 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