How a machine decides what you are weak at

1Weakness has three faces

Getting it wrong is the obvious one. But a child can be weak at something they are getting right.

• Wrong - not learned yet
• Right, but slow - reconstructed each time, not recalled
• Not met for a while - it was there once; whether it still is, nobody knows

The second is the one that goes missing without a record. A sum answered correctly after ten seconds looks exactly like one answered in a second: both are ticks. This is what a paper worksheet cannot keep.

2Three numbers into one

Each is scaled to a number between 0 and 1, then added. The weights this drill uses:

• 0.5 × (1 − accuracy)
• 0.3 × (mean time ÷ 10 seconds, capped at 1.0)
• 0.2 × (days since last seen ÷ 30, capped at 1.0)

Accuracy carries the most weight: a wrong answer means the thing is not there yet, and that outranks being slow or being stale. The weights are the policy, which is why they are written down rather than left to feel.

3Counting forgetting as part of it

The third term pushes a topic up the list even with a clean record behind it: thirty days untouched scores full marks. What was learned decays if left alone - the same assumption that spaced repetition is built on.

This is not a full spaced-repetition scheduler, though. Nothing here computes a due date; the term simply raises the weight. In a ten-question run there are not enough slots for a schedule to be honoured anyway.

4All weak spots, all the time, does not work

Hand out questions in strict order of weakness and a ten-question run fills up with the single worst topic. Efficient, perhaps. But a child who meets ten problems they cannot do does not come back for a second round.

So two rules are bolted on. No single topic takes more than 40% of a run - four questions out of ten. And nothing drops to zero weight: even a mastered topic keeps a floor of 0.05. The first is there to keep the run bearable, the second to stop mastered material from rusting.

Which is to say: the best schedule is not the one that teaches fastest on paper. If continuing is part of the goal, stepping back from the optimal allocation usually ends in more problems solved.

5A wrong answer outlives the day

A missed problem is kept as it was and slipped back into later runs (up to 30% of a run). It clears only when answered correctly first time. Getting it right immediately after being shown the answer does not count: help was needed.

For the same reason, only first attempts feed the weakness score. The drill shows the answer after two misses and re-asks the question later; counting that re-answer would make the hardest topics look like the most accurate ones. That repetition is practice, not a result.

The records expire after 90 days. Kept forever, a single slip from six months ago would go on bending today's practice.

6Choosing the pool first

Before any of this, the pool is cut down. So that 1 + 1 and 48 ÷ 12 never share a run, there are three bands - and the bands do not overlap: the top band does not reach down into the ones below.

Kinds never attempted are introduced at a modest weight (0.3), but only if they fall inside the chosen band - everything in a band is material for that level, so meeting one cold is reasonable. The full breakdown is in the 42 kinds.