Pipeline

How a verdict gets made

1. You paste or upload. The parser handles WhatsApp .txt exports ([date, time] Name: msg), iMessage-style pastes, Discord logs and loose Name: message transcripts. System lines — "added you", "missed call" — are filtered out.
2. Rex counts what counts. Message volume per sender, who writes paragraphs vs one-worders, response gaps, initiation patterns. You see these stats before generating.
3. A truncated excerpt goes to the model. Long chats are sampled across the whole arc — beginning, middle, end — capped around 14k characters. The full file never leaves your browser.
4. The persona does the reading. A fixed analyst persona ("Rex") with strict rules: cite evidence, distinguish pattern from proof, no generic advice, flag thin samples.
5. You get sections, not a wall. The Read → Who Cares More → Turning Points → The Pattern → The Play → Sign-off.

Why a persona instead of a generic summary

A raw LLM summary of a chat is boring and hedged. Rex exists because the product is the voice: unsentimental, evidence-first, funny when the receipts are funny. The structure stays consistent so reports are comparable — run the same person twice in six months and the delta is the story.

Credits and limits

Reports are billed in credits through the standard ZShip billing path. Generation fails cleanly on insufficient balance — you are never charged for a broken report.