A governing instrument for your own coding agent: every claim of done is independently verified, and every verdict carries its receipt.
Enter the workspace“I have not yet begun to fight.” — JOHN PAUL JONES · 1779
Your agent reports done. Serapis does not take its word. Every claim is put through deterministic verification and either passes, fails, or is marked unknown — and each verdict carries a receipt: what was checked, how, and when. Blind review stands between the claim and the record; the agent cannot talk its way past it.
One chart shows what exists, what is proven, and what still needs a human decision. When evidence goes stale, the verdict is withdrawn — the map never flatters. Beneath it, durable architectural memory outlives any conversation window, so what was decided, and why, survives every session that follows.
Paste one line into the agent you already run — Claude Code, Codex, and Cursor connect through MCP or a plugin. Then say “review my project”. That is the whole ceremony. No new editor, no migration, no week of setup. Your workflow stays your workflow; it simply gains a keel.
The first scan fingerprints your project and derives its atoms: what passed, what could not be proven, and what to correct first. Each finding arrives with its evidence attached and a receipt behind it. You are not reading a summary of confidence; you are reading a chart of what is actually known.
Serapis governs the agent you already pay for. Bring your own AI, your own keys: we never resell tokens and never hold your credentials. The intelligence is yours; the instrument that verifies it stands apart. That separation is deliberate — the party that checks the work should hold no stake in the work’s telling.
When the work demands it, run the control plane inside your own boundary. Your data lives in your database; every record — receipt, verdict, map — exports on demand. The decision points stay with you: AI drafts, it never decides. That line is drawn in the architecture, not in the terms.