Approval Gates for AI Coaching: 3 Failure Modes From Real Clients

Approval Gates for AI Coaching: 3 Failure Modes From Real Clients

Autonomy is useful until it starts making irreversible decisions faster than a human can audit them.

Justin Harris told one client that if GI distension was the issue, a daily serving of Dulcolax in the final 8-10 weeks of prep was the move; in the same breath, he connected the pattern to constipation, reduced water intake, and backed-up output from travel. That is the mechanism worth caring about here: state changes accumulate before they are obvious. In AI fitness coaching, the right lesson is not “let the model run.” It is that automation should be constrained by approval gates whenever the next action can compound error, distort interpretation, or create a bad decision path that the coach cannot easily unwind.

The hype version of AI coaching imagines a system that notices patterns, updates the plan, and keeps the athlete moving. The better version is narrower: AI can draft, sort, and flag, but humans should approve the actions that alter inputs, timing, or escalation. That thesis is falsifiable. If a system is allowed to act autonomously in the parts of coaching where feedback is delayed or noisy, error will show up as accumulated fatigue, missed attribution, or overcorrection. The transcripts in the KB show those failure modes plainly.

1) The first rule: if the input changes the signal, gate it

The strongest argument for approval-gated automation is not philosophical; it is procedural. In the Harry Sims thread, the client noted two daily caffeine doses: a black coffee with meal 1 at 9:30 a.m. for 150 mg, and a white Monster with meal 2 at 1:30 p.m. for another 150 mg. He also volunteered that he was sleeping about 8:30 hours on average and felt well rested. That is the kind of context that makes a coaching decision legible: intake, timing, sleep, subjective readiness.

A sloppy AI coach would treat caffeine as a generic lever and respond with more output, more corrections, or a schedule change based on a single metric. A constrained system does the opposite. It pauses and asks for approval before changing the lever that could contaminate the read on recovery, appetite, or energy. Why? Because once the input changes, the output no longer means the same thing. If you increase, decrease, or re-time stimulants automatically, you are no longer observing the athlete’s baseline behavior. You are editing the experiment.

That matters because good coaching depends on attribution. If the athlete feels worse next week, was it the diet, the schedule, sleep, the second caffeine hit, or the coach’s own change? Approval gating protects the signal. It keeps the system from making the very thing it is supposed to measure less interpretable.

2) The second rule: when the effect is delayed, autonomy invites overcorrection

In the Alex Goracy transcript, Justin Harris says the athlete should get as lean as possible to put himself in the best position to grow on the rebound. He also describes a line where further depletion creates a bigger total weight change from end of diet to end of rebound, but not necessarily a bigger end rebound size; past a point, the athlete is mostly regaining fullness lost during dieting rather than adding new fullness. He puts a number on the tradeoff: around 10 lb of rebound gain may be weight that did not need to be lost to maximize the rebound.

That is an ideal example of a delayed-effect system that should not be fully automated. If you let a model push changes continuously toward a short-term metric, it will often optimize the visible part of the curve and ignore the downstream cost. In physique coaching, the downstream cost is not abstract. It is the difference between useful depletion and unnecessary depletion, between clean rebound and merely refilling what was stripped away.

Approval gates are the antidote because they force a human to ask the important question: is this a real improvement, or just the first half of a longer process? AI can surface the trend. It should not be the entity that decides the athlete is now “better” and pushes the next change. The more delayed the effect, the more the system needs a human to resist false confidence.

3) The third rule: if the mechanism is uncertain, the model should not improvise

The Skip Hill exchange is blunt enough to be useful. Justin says he does not like prepping hard for longer than 16 weeks and would not push too much early if the athlete is already comfortable with conditioning. He also notes that at the doses the athlete is on and the body fat he is at, he is in a great position. In other words: don’t chase action for its own sake when the current state already looks workable.

That is the core danger of unconstrained AI coaching: it is action-biased. Give a model a stream of check-ins, and it will tend to produce a response every time. But coaching is not customer service. Not every message requires a plan update. Not every data point warrants escalation. In a human workflow, approval gates provide friction exactly where action bias is most dangerous.

This is also why the nicotine-or-narcotic-style analogies people like to reach for are less important than the actual observed pattern: constipation, distension, water intake, prep duration, sleep, and dose context. The coach is not guessing at a magic fix. He is checking whether the next move is even necessary. That is what approval-gated automation should imitate.

What approval-gated AI should actually do

A practical coaching system should separate three layers:

1. Detect and draft. The model can summarize check-ins, identify probable issues, and propose a candidate action. Example: “Distension increased after travel and reduced water intake; review hydration and GI consistency before changing anything else.”

2. Require human approval before any action that changes the experiment. That means no automatic changes to calories, cardio, caffeine timing, supplements, sleep targets, or training stress when the model is only working from sparse data. The coach should approve those moves because they change the meaning of the next datapoint.

3. Log the rationale in plain language. If a change is approved, the system should record why it happened and what it is trying to test. Otherwise the athlete ends up in a fog of silent plan edits, and the coach loses the ability to tell whether the system improved or merely intervened a lot.

This design is slower than full autonomy, but it is more honest. It treats coaching as a sequence of controlled experiments, not a stream of automated reactions. That is especially important in physique prep, where small changes can echo for weeks and the body does not always tell you immediately when you have pushed too far.

Why constrained autonomy wins

The reason to constrain AI is not fear of technology. It is respect for causality. The corpus shows a consistent pattern: successful coaching depends on context, timing, and knowing when not to act. Travel plus lower water intake can create backed-up output. Rebound size is not the same thing as total rebound weight. A client who is already in a strong position does not need the model to invent urgency.

So the practical rule is simple: let AI observe, summarize, and propose, but gate every action that changes the athlete’s state. If the system cannot explain what it is testing, if the outcome will be delayed, or if the next move could blur attribution, the coach should approve it manually. That is not a limitation. It is the point.

Sources Used

• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w19-24m/clients/david_lamartina___members-tlssnsjthkmnhfqcscszce25acz_vhdm_x2_xdlpx_i.json
• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w13-18m/clients/skip_hill___members-b7s-_d0oqqz4a1tadlvuztb0y_vzfqlr9c29tpakweg.json
• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w19-24m/transcripts/alex_goracy___members-m1vvgmmbnhazzipv5wqmwwhgjuyotawsompy4kzf6ri.md
• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w19-24m/clients/harry_sims___members-m-iwooxrve6qux4pdtrouhwbkhmvrxoandj1tgzve_m.json

Sources Used

• /Users/justinharris/TroponinIQ/kb/supertrop/raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w19-24m/clients/david_lamartina___members-tlssnsjthkmnhfqcscszce25acz_vhdm_x2_xdlpx_i.json
• /Users/justinharris/TroponinIQ/kb/supertrop/raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w13-18m/clients/skip_hill___members-b7s-_d0oqqz4a1tadlvuztb0y_vzfqlr9c29tpakweg.json
• /Users/justinharris/TroponinIQ/kb/supertrop/raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w19-24m/clients/harry_sims___members-m-iwooxrve6qux4pdtrouhwbkhmvrxoandj1tgzve_m.json
• /Users/justinharris/TroponinIQ/kb/supertrop/raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w19-24m/transcripts/alex_goracy___members-m1vvgmmbnhazzipv5wqmwwhgjuyotawsompy4kzf6ri.md