File a few cards and watch the right-hand column. The model stays silent until it has learned something — a bad suggestion costs more than no suggestion. Once it speaks, accepts. Undoing with does not merely put the card back: it unlearns the example.
—
Nobody decrees who is right: every member is asked before learning, and its weight follows its past mistakes.
The same code as bun tools/bench.ts, on a synthetic corpus of 800 cards where the zone mostly depends
on the combination of domain and tag. That is the case naive Bayes cannot see: each feature on its own
votes the same on both sides.
| configuration | top-1 | top-3 | silent | features |
|---|
On that corpus, crossing features buys eight to nine points for Bayes and for the linear model, and kNN moves ahead. On a real corpus of 3,412 links across 72 folders, the same crossing buys half a point — marginal signals dominate there — and the ensemble takes the lead instead. Hence the bench: measure on your corpus, do not copy someone else's numbers.