Lead the research digest with the name the sources agree on

[devbadya]

A 0.8B model was being asked to extract Friedrich Merz or Paris from a list of quoted passages. That is the step it drops: it either invents, or quotes the definition of the office instead of the incumbent.

research now does that extraction itself, the same way weather already reconciles forecasts before the model sees them.

When the digest opens with Answer:

• Two independent sources name the same person or place
• Or one sentence states that name as the predicate of the question (Paris is the capital, Amtsträger ist seit dem 6. Mai 2025 Friedrich Merz, a novel by Frank Herbert)

When it stays off

• The name is the subject of the question (Who is Elon Musk? already knows who)
• Two different names are equally well supported
• Nothing is confident enough — inventing a one-liner is worse than making the model read the quotes

The research-question skill now shows that first line in every exemplar, so the model is taught to copy it.

pnpm check is green (1,293 tests). Whether the 0.8B actually copies the line still needs /?eval on a GPU.

Answer-first digest samples

_{To show artifacts inline, enable in settings.}