Production is on fire.
Let’s inspect the architecture.

A local model, with limits

This app serves local Qwen3-4B merged weights. SFT and DPO checkpoint records exist in the project. The served model’s files predate the newest SFT checkpoint; a complete training-to-export manifest is unavailable.

The interface and prompts have been improved without retraining or replacing the model. These changes do not certify answer accuracy.

This is an independent learning project, not an Anthropic product or an official exam simulator. Topic labels come from the local practice bank.

Check the source

Small reviewed notes for agent stop reasons and MCP transports are supplied to the tutor when relevant. Links indicate provided context, not proof that every generated claim is correct. Other topics rely on model knowledge.

Notes checked November 1, 2026. No live web retrieval occurs during chat.

Claude: stop reasons and fallback ↗

MCP transport specification ↗

Humor that knows its place

Jailer Lab restores the warm greeting, unruly-agent metaphors, cheerful disaster commentary, and roasts aimed at bad architecture. It is an original Architecture Warden persona, not an impersonation. The original quirky personality is always on. Small models may not follow every instruction reliably.

Example: “An unbounded retry loop is optimism with a billing account.”

Your session

Chat is kept in memory in this tab. Reloading clears it. Practice answers, bookmarks, are stored in this browser; reset removes practice progress. No account is required.

The host must remain online. If the shared model is busy, your question stays available to retry. Requests have bounded size, context, duration, and output limits.