Using AI to decide a vote

One voter's experiment for Israel's 27 October 2026 Knesset election. A 42-minute voice note was turned into a scoring rubric, research agents gathered evidence on all 37 lists on the ballot, and every list was scored against that voter's priorities. Adjust the weights below to see how the ranking moves.

How it worked

1 · Voice noteDaniel records what he thinks is broken and what he cares about (26 Oct, 42 min).
2 · TranscribeAssemblyAI; raw transcript kept, plus a lightly edited version.
3 · RubricClaude turns the note into 15 criteria with weights, each cited to the note.
4 · ResearchFour Claude Sonnet agents gather sourced evidence on every list, on election day.
5 · ScoreOne Claude Opus agent scores all lists 0–5 on every criterion, with written reasons.
6 · RankA script combines the scores, applies the voter's exclusion rule and checks the polls.

Full repo (voice note, transcripts, evidence, scores, PDFs): github.com/danielrosehill/Israel-Election-2026-Vote-Decision. Party lists: Israel-Election-2026-Candidate-Lists. Scores are AI judgements against one person's stated positions, not neutral ratings of the parties.

Result with the current weights

Weights

Defaults are the rubric's, inferred from the voice note. Stream A = everyday pain points; Stream B = the divisive national issues (each B item 0–5 = alignment with the voter).

Ranking

Raw scores (0–5)

Why each list scored as it did

Findings (default weights)