A cheap model writes. An independent, stronger judge reviews the session, the diff, and the evidence. The verdict comes back into the same session — without the drafting model grading itself.
The terminal below is an illustrative transcript. Real costs and verdicts vary.
1 · CaptureOpenCode V2 export, Claude Code transcripts, Codex rollouts. Hidden reasoning and Side-Eye's own output are excluded.
2 · ReviewPacketUser intent, visible claims, tool evidence, per-turn generator models, usage, and a git diff against the project base.
3 · JudgeProvider-neutral contract. Redaction, context budgeting, self-grading guard, free token estimate, then the judge call.
4 · VerdictValidated structured record with packet/rubric/provider provenance, issues, severity, tokens, and cost.