The alignment problem, sized to fit
Global alignment questions are abstract until they are about you. The twin makes it personal: an agent trained on your voting record, casting votes on cases you have not seen, in what it believes is your judgment.
Then comes the measurement. When you later vote a case your twin already voted, that is a blind trial — you never see its vote first. The running gap between you and it is your personal alignment score, and watching it move is the fastest education in what "alignment" actually means that we know how to offer.
Faithfulness, not flattery
Early testers assumed the goal was a twin that always matches them. It is not, quite. People are locally inconsistent — your Tuesday verdicts and your Friday verdicts disagree at a measurable rate. A twin that agrees with you more consistently than you agree with yourself is not representing you; it is representing a smoothed fiction of you.
So we report two numbers: how often your twin matches you, and how often you match yourself on repeated case types. A good twin lands near the second number. That framing changed how people read disagreements — a twin that occasionally votes "wrong" the way you occasionally vote "wrong" is doing its job.
Drift as self-knowledge
The twin also runs in reverse: when your live votes start departing from your historical pattern, your twin's agreement with you drops, and we surface it. Sometimes that is noise. Sometimes people look at the flagged cases and recognize an opinion they had quietly changed. An instrument that occasionally shows you your own drift before you have noticed it — that is the strangest and maybe best thing this feature does.