The mistake is upstream of the plan
Ask for a sub-3:30 marathon training plan and you will usually get one. The sessions will be sensibly structured, the long runs will build, the taper will be there. And the paces will be derived from 3:30, because that is the only number in the conversation.
That is the failure. A plan should pace you off what you can currently do, then aim you at the goal. A plan built the other way round paces you off an outcome you have not achieved yet, which means every session is prescribed at the fitness you hope to have rather than the fitness you have.
Why it is easy to miss
Nothing about the output looks wrong. The paces are internally consistent with the goal, the structure is orthodox, and the first week feels hard in a way that is easy to interpret as the plan working.
The gap shows up later and indirectly. Sessions start failing in the last repetition. Easy days drift faster than easy because the prescribed easy pace was calculated from a threshold you do not have. Then the long runs get uncomfortable in a way that has nothing to do with endurance and everything to do with running them slightly too quickly for months.
The distance makes it worse
Over 5K a pace that is two per cent optimistic is a hard session. Over a marathon build it is sixteen weeks of small overreaching, and the cost compounds rather than showing up on any single day. That is why this failure is most expensive in exactly the plans people most want AI to write for them.
It also runs in the other direction. A goal you set conservatively produces a plan that never asks anything of you, and the sessions feel comfortable all the way to a race you were capable of running faster.
The fix is one extra number
Give it a current benchmark alongside the goal, and say which is which. Keeping those two as separate kinds of record is the whole reason Athmex stores goals apart from measurements. A recent race result or a threshold you actually tested lets it derive training paces from where you are, and treat the goal as the target to close the gap on.
A goal with no benchmark is a wish. A benchmark with no goal produces maintenance. The pair is what produces a plan, and the pair is what most prompts are missing. The wider set of things worth telling it before you ask for anything is covered separately.
Cyclists hit the identical problem with FTP, where percentages of a wished-for number rescale every session in the block.
Where Athmex fits
Athmex keeps the two apart on purpose. Your goals and target races are stored as goals, with dates and priorities. Your thresholds and personal bests are stored as dated measurements. They are different kinds of record, and an assistant reading your context receives them as different kinds of claim rather than as two numbers in a paragraph.
It also receives how old the benchmark is, which is the other half of this problem: a current goal paired with a benchmark from last season produces the same error by a slower route.
Common questions
How do I tell if my plan has this problem?
Check whether the easy pace it gave you matches the pace you actually run easy runs at. If the plan's easy pace is faster than your habitual easy pace, the paces were probably derived from the goal.
What if I have no recent race?
Say so. An assistant told that your benchmark is old, or that you are estimating, can hedge appropriately. The damage comes from it treating an aspiration as a measurement, not from you lacking a perfect number.
Does Athmex set my training paces?
No. It stores the benchmarks and the goals you confirmed. Claude or ChatGPT derives the paces, and it does so from values you checked.