When your coach and your algorithm disagree, who gets the final word?

The disagreement starts under the bar

An algorithm should never outrank the person who bears the consequences—but neither should human instinct get an automatic veto.

It is 6.10 on a Tuesday morning. The original plan is five sets of five squats at 82.5 per cent. After two shorter nights of estimated sleep and a sustained change in resting heart rate, the wearable recommends reducing the session. The athlete feels unusually strong and wants to progress. The coach sees controlled warm-up repetitions but knows the sleep record contains a sensor gap, so argues for maintaining the planned load. Measured trend, felt experience and observed performance now point in three directions.

The coach adds two submaximal checks. The adapted response is to hold the planned load, reassess movement quality and effort after each working set, and shorten the accessory block if the earlier signals begin to appear in performance. The explanation is explicit: no single metric justified a reduction, but unresolved disagreement did not justify progression. Pain, faintness, deteriorating execution or another concerning response would stop the session and may warrant qualified professional advice. The decision remains reversible.

Three seats at the table, no single throne

The athlete holds authority over participation, effort and personal goals. Felt experience can reveal soreness, stress, apprehension or unusual effort before those factors appear in a dataset. Athlete autonomy is not a courtesy. A large sport and exercise meta-analysis found strong associations between autonomy-supportive coaching, autonomous motivation and wellbeing, while also noting that much of the evidence is correlational.

The coach contributes situated judgement: exercise history, technique, behaviour, recent life context and what is happening in front of them. The model contributes consistency and pattern detection across records that humans may forget or selectively notice. Good readiness-based training uses all three forms of information without pretending they are interchangeable.

The practical hierarchy is clear. Safety constraints come first. The athlete retains consent and can decline an action, although they cannot compel a coach to supervise an option the coach considers outside agreed boundaries. The coach guides within their scope. The model advises. This adapts a shared decision-making principle: combine available evidence with the person’s values and goals, retain the option of no change and document material trade-offs. Our research approach treats that hierarchy as a reasonable transfer from authoritative and medicine-adjacent work, not direct proof for every training setting.

The conflict matrix: authority follows stakes

The deciding question is not simply who is right. It is which source holds relevant information, what can be verified and what happens if the call is wrong. Use this matrix as an escalation order rather than a scoreboard.

Decision cost runs in both directions. The possible cost of progressing too aggressively includes greater fatigue, avoidable risk, reduced adherence or a session that undermines subsequent work. The possible cost of reducing unnecessarily includes slower progress, a weaker training stimulus or a missed opportunity to train productively. Blindly following a model can automate a bad call. Reflexively overriding it can hide a slow pattern that is difficult to see session by session. The purpose is not to maximise training or remove every exposure; it is to choose the most appropriate action from imperfect information.

Three cases, three different winners

Case one — the humans are right. A sleep estimate falls sharply because the wearable was removed for several hours. The model recommends delaying an easy run. The athlete reports normal sleep, the coach confirms there has been no unusual recent load, and the warm-up remains easy at the intended pace. They maintain the run while retaining its easy ceiling. The device output is not ignored; it is discounted because the underlying record is incomplete and the possible cost of an unnecessary delay outweighs the alert’s weak informational value.

Case two — the athlete is right. The model recommends progressing a pressing exercise because recent workload and recovery records appear stable. During the warm-up, the athlete reports new, localised discomfort that changes with the movement. The coach cannot reproduce that felt information from a dashboard. They stop the movement, substitute only if a comfortable option is available, and seek qualified advice if the issue persists or is concerning. A positive model output cannot overrule immediate human feedback with safety implications.

Case three — the model catches a blind spot. The athlete feels strong after an excellent night’s sleep, and the coach is anchored to a crisp opening set. The model identifies three weeks of accumulating session load, declining self-reported recovery and rising effort at previously routine workloads. They check the records and confirm the pattern. The key movement stays, but working sets are reduced and the secondary high-intensity block is rescheduled. The model wins this argument because it exposes a coherent pattern neither human had assembled—not because it has superior status. These cases are illustrative judgement calls, not universal rules.

Decision confidence

Decision confidence belongs to the process, not a mood or score. High confidence requires reliable inputs, agreement across signals, a stable trend and context that fits the recommendation. Medium confidence means useful agreement remains, but information is incomplete or one source conflicts. Low confidence follows poor sensor contact, missing reports, outliers or unresolved disagreement. It should produce a conservative hold, reduction, delay or reassessment rather than certainty theatre.

In the gym-floor case, the signals that mattered were the resting-heart-rate change, disrupted sleep estimates, the athlete’s felt state and observed warm-up quality. The composite score and one good repetition did not matter enough to decide. Full progression was rejected because disagreement remained; automatic reduction was rejected because the sleep record had a gap and current observations were stable. Keeping the original recommendation is justified when inputs are sound, material signals do not converge on change, safety boundaries are clear, and the rationale and revisit trigger are documented. Flex Force X readiness signals are designed to be interpreted together with context, confidence and safety constraints.

Override rules: leave a decision trail

Every override—whether human-led or model-led—should answer five questions. What changed? Which signals mattered? Which were discounted, and why? Why was this option selected over maintaining, progressing, reducing, substituting or delaying? What will trigger another review? An explanation should be concise, relevant to the person making the call and proportionate to the stakes. More detail is not automatically better: reviews suggest explanations can support appropriate trust, but can also make an incorrect recommendation more persuasive or impose extra cognitive burden.

The minimum record is the original recommendation, the conflicting information, the chosen action, who approved it and the revisit condition. Record only what is necessary and handle it consistently with privacy expectations. Repeated overrides in the same direction should trigger a review of data quality, model fit, coaching assumptions or athlete reporting—not quiet normalisation.

Flex Force X adaptive workout plans are designed to adapt training recommendations when relevant context justifies a change. The system philosophy favours clear explanations for recommendations and changes. Safety checks and appropriate human oversight are design constraints, not guarantees that all risk can be removed. Read together, these principles support a disciplined sequence: connect the available information, assess its quality and agreement, apply safety rules, decide whether adaptation is justified and explain the reasoning. Technology does not replace qualified professional advice; the medical disclaimer defines that boundary.

The research supports restraint rather than loyalty to either side. Automation bias is documented, particularly in complex decisions. Wearable research indicates that some runners prefer bodily experience over device guidance, while readability, personal relevance and useful dialogue can shape continued use. Sport ethics reviews repeatedly identify transparency and interpretability as concerns. Much of the relevant work is observational, laboratory-based or medicine-adjacent, so this hierarchy is an evidence-informed operating model rather than a universal law.

The closing challenge

At the next disagreement, do not ask whether humans or algorithms are better. Ask who bears the consequence, whether a safety boundary has been crossed, which source holds information the others lack, what the cost of error is in both directions and what evidence would justify changing the call. If the answer cannot be explained and recorded, the decision is not ready.

The final word belongs not to the loudest authority, but to the best-governed decision.