voice-corpus

Turns raw exports into corpus.jsonl: one line per message the user actually wrote, quoted replies and signatures cut, automated mail and assistant leftovers dropped, third-party emails/phones/URLs/names redacted, each sample tagged with a register (email-internal, email-external, chat-dm, chat-group, doc). Everything downstream (profile, write, check, calibrate) reads this file.

The failure it fixes: a Sent export is often mostly other people's text (quoted threads, signatures, calendar mail, forwarded content) plus AI-assisted drafts the user hit Send on. Measuring that measures a chimera, and the profile then "solidly" describes someone else. Practitioner logs put the foreign share of a raw Sent export around half.

When to use

• "here's my Takeout mbox / my Sent.mbox / these .eml files, build my writing corpus"
• "import my WhatsApp / Slack export, only my messages"
• "clean my sent mail export: strip the quoted replies and signatures"
• "prep my emails for the voice profile, redact names"