A token count is the cleanest number in the building. It arrives daily, it has an invoice behind it, and nobody argues about the definition. The other side of the ratio needs a person to say what changed and for whom, and there is no feed for that. The Cutlefish post's image catches the mood: an organisation that used to eat at a buffet is now buying each dumpling separately and re-litigating the value of every one [5]. What the finer-grained bill does not fix is the older problem the same post raises about hours, that where effort lands tells you nothing about the quality or efficacy of that effort, and that when the whole system is waiting on one specialist or one review, spending time elsewhere has negative leverage [8].
The arithmetic is worth doing slowly. Take the post's switching example: alternate between two tasks every ten minutes across a six-hour day, and at least three hours go to switching and reorienting [9]. Three of six is half the day gone [10], while the timesheet still books six hours against two tasks, because time accounting adds where coordination multiplies [9]. The post's other comparison lands in the same place: ten people at ten hours each and one person at a hundred hours both total 100 [11], and the ledger cannot tell them apart [8].
Token reporting has that property too. Spend per seat climbing tells you the meter is running. It does not tell you whether the same people were still inside that workflow six weeks later, or how long a new user took to get one output they were willing to ship. Those numbers are harder to pull, and they are the R.
Here is the forcing function I would use. Draw two axes: how precisely you can measure the input, and how precisely you have defined the return. Tokens land where hours and points landed, with the input measured to four decimal places and the return described in a sentence nobody will put their name to. The fix is not a better meter, it is deliberately inverting which axis gets the rigour. In practice that means a token budget does not clear review without one named workflow, the people working inside it, and an observable change with a date attached. If nobody will sign that sentence, the meter is doing the justifying.
That matters more than usual right now because the post is candid that it cannot tell what sits underneath the current enthusiasm, a genuine attempt to understand how AI augments people, or a calculation about how many people can safely be let go, or both [15]. Under that ambiguity, an undefined R does not stay undefined. It gets filled in later by whoever is holding the invoice.
