AI Outcome ResponsesCopy Anchor Link

During a live round, the speech timer's menu (the ⋮ next to the timer) has an AI Outcome Responses item. It reads the whole round so far and suggests three ways to give the speech you are on, then plays each one forward.

• Where: live round workspace, speech timer menu
• Package: debate-round (UI and context), debate-speech-writer (prompt)

What it showsCopy Anchor Link

Three candidate approaches side by side. Each one has:

• a success rate: the AI's estimate that this approach wins the round for your side, with the best bet marked;
• a speech-ready outline that extends cards already read and answers the opponent's newest arguments (copy button included);
• the opponent's likely answers in their next speech;
• the judge decision: who the predicted judge votes for on that path, and why;
• the issues that could lose it (drops, missing evidence, judge fit);
• cards to find: search queries that run against CARDS in place and show the highlighted text of the top hits, with a copy button and a link to open the full search.

You can override the judge paradigm and add an optional focus ("go for the DA") before regenerating.

Where the context comes fromCopy Anchor Link

• Prior speeches: every speech column before the current one. Each speech's linked editor document is used when it has one, otherwise the flow's own speech doc. Highlighted card text is kept marked so the model can tell what was read aloud.
• Flow: every flow sheet in the round, one line per argument row with each speech's entry, unanswered rows flagged.
• Judge: the round judge with the most ballots in Judge Profiles, voting under that profile's most-tagged paradigm. With no saved profile for the round's judges, the paradigm most common across your saved profiles is used, else a flow judge.

Data flowCopy Anchor Link

SpeechRecordingMenu (debate-timer) actions prop
  → dialogs/AiOutcomeResponsesDialog.tsx
round/outcome-responses-context.ts   (prior speeches, flow text, likely judge)
round/outcome-responses-client.ts    (linked docs, /api/reason-ai, /api/search)
debate-speech-writer prompts/speech-outcome-responses.ts
  → SPEECH_OUTCOME_RESPONSES_SYSTEM_PROMPT (built on speechToResponsePrompt)
  → buildSpeechOutcomeResponsesUserPrompt / parseSpeechOutcomeResponsesResponse

Known gapsCopy Anchor Link

• Results are not saved; closing the dialog keeps them only until you switch speeches.
• Success rates are model estimates, not calibrated against real ballots.
• On the free plan the AI reply is capped at 2,048 tokens, which can cut a long three-candidate reply short; the dialog then shows a parse error.
• Long rounds are trimmed to about 60,000 characters of speech text, oldest speeches first.