The Reasoner uses the same 11 MB model as the Minter, but asks a sharper question. Given a typing, derive its weaknesses from a type chart that was never handed to it as a lookup. The page grades every answer live against the real chart.
Most of the type chart should be memorised. Electric weakTo Ground has no
derivation; every player just knows it. Only one thing here is compositional: a dual type's
weaknesses are the product of the two single-type charts. That product needs
cancellation, knowing that Steel's weakness to Fire dies against Dragon's resistance
to Fire.
So the model was taught all 18 single types and 123 of the 153 dual typings, but 30
dual typings had their weaknesses withheld from every document. Those held-out combos
(the demo flags them ◆) are the real test: their answer can't be retrieved, only
derived.
On typings it saw, it is close to perfect. On the held-out ones, it mostly does the union, adding the two weakness lists together instead of multiplying them, so it over-claims: it keeps a weakness that the second type should have cancelled. Exact-match on held combos sits near the union baseline (~0.23), not the ~1.0 a learned rule would give.
This is the collection's argument in one place: fluency is not reasoning. The model learned the chart as a 105-entry table, one level up from the entity table the Memoriser learned, and a table, however large, doesn't generalise to the row you removed.
◆ held-out combo (the hint tells you). Watch it over-claim, and the
grader mark the weakness it should have cancelled.