Check that replies copy the researched Answer line

[devbadya]

Research already extracts a name or figure onto the first line of the digest (Answer: Friedrich Merz.). The 0.8B model still invents a different one, or quotes the definition of the office instead of the incumbent. That is the same failure the calculator already had, and it had the same fix waiting: a deterministic check against the tool result.

reviewAnswer now treats that line the way it already treats 6748 * 9 = 60732. If the draft drops it, one correction asks for it back. The check stays shy:

• A surname standing in for the full name (Merz for Friedrich Merz)
• A missing accent (Antonio for António)
• A German decimal (13,96) or a figure rounded the way extractFigure already treats as one reading (14 million for 13.96 million)
• A biography digest with no Answer: line (Who is Elon Musk?)

When the wind-down round comes back empty, budgetFallback hands over the same one-liner instead of an apology.

This is the highest-leverage remaining research fix. #50 and #51 put the extract on the digest; they did not check that the model copied it. Lengthening the system prompt, adding more tools, or asking the model to grade itself have already been measured here and made things worse.

pnpm check is green (1,324 tests). Generation was not run: this VM has no GPU adapter.

What this does not fix

• Follow-ups like und der von Frankreich? after a chancellor question — carry-over already routes them, but the query still has to name the new subject
• Definitions and ties, where the digest stays silent on purpose
• Whether the 0.8B actually applies the correction — that still needs /?eval on a real GPU