Nothing here is a summary written after the fact. These are the problem, mechanism,
and move of each article, in the reading order, joined by its own connective lines.
Any one of them opens into the full argument.
Start with the sign of the number, because a tool built to develop a person records its best outcome as a decline, and nothing further in this topic makes sense until that is on the page.
Retention figures for tools meant to build capacity are read as evidence of failure, and the measurement cannot separate a person who quit because it was useless from a person who stopped because the practice had become theirs.
Mechanism
Two of the three defining characteristics of scaffolding describe the support going away, so a product built on it succeeds by becoming unnecessary and every usage number it reports moves down at the moment it works.
Move
Count the people who reached a defined competence and left, separately from the people who left, and price the product on the unassisted follow-up rather than on the session.
If the direction of a metric can be wrong, so can its independence, and the next failure is an instrument validated against a paraphrase of itself.
Validating an instrument against an outcome is the right instinct, and it fails in one specific way that is old enough to have a proper name and current enough to be shipping in evaluation harnesses.
Mechanism
When the items and the criterion are drawn from the same construct the correlation between them was fixed at the moment the items were written, so a validation that looks strong is the instrument meeting its own reflection.
Move
Write the criterion before the items and require it to be obtainable by a person who has never seen the instrument.
A circular measure at least reports something about wording; the harder case is a number that is genuinely accurate and still says more about the person's constraint than about the model.
Accuracy is treated as the least controversial number in machine learning, and a figure that rises is read as a better model when it is equally consistent with a life that got narrower.
Mechanism
Prediction succeeds on habitual and constrained behavior and fails on value-laden choices in new situations, so the accuracy figure reads out how little room the person had, and a system acting on its own predictions can improve that figure by reducing the room.
Move
Report accuracy alongside what it was accurate about, and never grade a system on predictions it is also in a position to bring about.
Behind every one of those numbers sits a smaller one that nobody argued about, set once in a configuration file and binding on everyone the system touches.
A number set once in a configuration file can decide how much of one person's exposure buys how much of everyone else's accuracy, and it ships as a default that nobody subject to it can read.
Mechanism
Epsilon fixes in a single value how much noise shields an individual record and therefore how much accuracy every user of the results gives up, so whoever picks it is ruling on a trade between two parties who are both absent.
Move
For any number that trades one group's interest against another's, write down who gains, who pays, who chose, and where the choice is recorded.
End where the measuring meets the person being helped, at the finding that identical guidance helps one of your users and costs the other at every moment, and that you cannot tell which without measuring them rather than the content.
The same unchanged guidance does not merely stop helping as a person improves, it starts costing them, and a support system with no model of its user's level cannot tell which of its two populations it is currently harming.
Mechanism
An expert already holds the schema the explanation supplies, so the explanation arrives as a second external account that has to be reconciled and suppressed, spending working memory the task itself needed.
Move
Grade the guidance against the person rather than the material, ship the version that can be turned down, and read a request to see less as data rather than as disengagement.
You have walked Instruments of care end to end: the metric whose good direction is down, the scale that agrees with itself, the accuracy that reads out a narrowed life, the parameter nobody voted on, and the guidance that crosses into interference. A tool that claims to care for its users is making claims that can be checked, and every one of them is a number somebody chose.