Approval Gates and the 7-Set Rule
Why AI coaching should default to constrained autonomy, not delegated control
Approval Gates and the 7-Set Rule
Why AI coaching should default to constrained autonomy, not delegated control
A 2019 Nature Machine Intelligence study found people were less willing to accept an AI decision after seeing the same system fall below a human in just one comparison; the mechanism is trust calibration. That matters for fitness coaching because the failure mode is not only a bad plan, but a plan that keeps adapting after it should have been stopped. The falsifiable thesis here is simple: in AI fitness coaching, the right default is approval-gated automation for any change that affects training dose, exercise selection, or weekly progression, because autonomy should be constrained until a human has explicitly accepted the next action.
The temptation in AI coaching is to let the system "run the process" once the first plan looks decent. That works fine for low-stakes reminders. It is much riskier for body composition work, where small repeated errors compound. Justin Harris’s teaching on training programming points in the same direction from a different angle: with low-volume, high-intensity approaches, "any miscue makes the whole session a failure." His preference in that context is not more automation, but more volume and more room for error. The underlying lesson is structural: when the cost of a wrong turn is high and the system cannot self-correct perfectly, you want checkpoints, not open-loop control.
Approval-gated automation is the practical version of that principle. The AI can draft the next training block, flag when volume is drifting, and propose a progression. But it does not execute changes on its own. A coach approves, edits, or rejects. That sounds slower. It is slower. That is the point. In coaching, speed is only a virtue when the downstream decision is already reliable. Most of the time, the valuable thing is not faster output, but fewer unreviewed mistakes.
This is especially true when the AI is making decisions that sound small but accumulate. Add one set here, swap one movement there, push load too aggressively for two weeks, and the athlete has a different recovery bill. Justin’s programming guidance repeatedly comes back to the idea that training quality is not just about intensity; it is about having enough volume and enough structure that one imperfect set does not wreck the whole session. Approval gating mirrors that logic at the software layer: one bad suggestion should not become an automatic policy.
A constrained workflow also gives the coach a cleaner job. Instead of checking everything manually, the coach reviews only the decisions that matter. That is the real productivity gain. The AI can sort, summarize, and recommend. The human stays responsible for the threshold call: is this change worth making now, for this athlete, in this phase? Once you move the AI from recommendation to execution without that gate, you have not built a better coach. You have built a faster intern with write access.
Good gates are specific, not vague. If the model proposes a new weekly set target, the coach has to see the old target, the reason for the change, and the expected tradeoff before approval. If it suggests a movement swap, the coach should know whether the swap is for joint stress, weak-point emphasis, or compliance. If it wants to alter cardio, fatigue management, or exercise order, it should explain the mechanism in plain language. The approval step is not just a rubber stamp; it is the place where the human checks whether the recommendation is coherent with the plan.
This matters because AI systems are good at confident synthesis and mediocre at context. They can assemble a recommendation from many inputs, but they do not automatically know which input is the real constraint. A coach does. In physique coaching, that constraint might be recovery, adherence, exercise tolerance, travel, or the athlete’s ability to train the weak parts without breaking the rest of the week. Justin’s material on coaching mindset and programming keeps returning to the same practical truth: bodybuilding progress is constrained by execution, not by the prettiness of the template. Approval gating respects that reality.
There is also a trust benefit, and it is not sentimental. When athletes see an AI making unchecked changes, they do not experience sophistication; they experience drift. They stop knowing whether the plan they are following is still the plan they agreed to. Approval gates create a visible chain of custody. The athlete can see that changes are intentional, reviewed, and bounded. That makes the system easier to audit and easier to stick with when the process gets boring, which is most of the time.
The best automation pattern for coaching is therefore not "autonomous unless the coach intervenes." It is "propose, review, then act." Leave open-loop automation for low-risk admin: reminders, logging prompts, summaries, missed-check-in alerts. Put high-impact decisions behind explicit approval: progression changes, deload decisions, exercise substitutions, recovery load shifts, and any recommendation that changes the stress budget. The more the AI touches the athlete’s actual training dose, the more the workflow should resemble a control room than a self-driving car.
That distinction matters because bodybuilding coaching is a long game. The worst failures are not dramatic. They are the subtle ones: a session that was just a little too much, a week that accumulated avoidable fatigue, a change that looked smart in isolation but was wrong in sequence. Approval-gated automation is the antidote to that kind of failure because it forces sequence back into the hands of a human who understands the athlete’s context.
So the useful standard is not "Can AI coach?" It is "Where should AI be allowed to act without asking?" For serious coaching, the answer is narrow. Let it automate the low-risk, repetitive work. Keep the consequential work behind a gate. If a system can’t explain why the next action is worth approving, it should not be allowed to take that action on its own.
Sources Used
wiki/youtube-primary-training-programming-05.mdwiki/youtube-primary-training-programming-06.mdwiki/youtube-primary-coaching-mindset-06.mdwiki/troponiniq-kb.md
Sources Used
- /Users/justinharris/TroponinIQ/kb/supertrop/wiki/youtube-primary-training-programming-05.md
- /Users/justinharris/TroponinIQ/kb/supertrop/wiki/troponiniq-kb.md
- /Users/justinharris/TroponinIQ/kb/supertrop/wiki/youtube-primary-coaching-mindset-06.md
- /Users/justinharris/TroponinIQ/kb/supertrop/wiki/youtube-primary-training-programming-06.md