Wiki Harness Racing

Guide your local LLM from the Start Page to the Target Page in as few clicks as possible, using only links in the article body.

• Pre-Race Planning: The harness fetches the target summary first. The LLM devises a strategic plan to reach it *without knowing* the starting page.
• Target Context Retention: The target article summary and the strategic plan are kept permanently in the model's system prompt context to maintain a clear destination.
• Memory Pruning: To conserve context space, older link lists are omitted from the chat history, leaving only the immediately preceding turn's link options.
• Strict Role Alternation: Combines tool results and system alerts to keep alternating user/assistant message roles.
• Interactive Self-Correction: Non-fatal errors (e.g. hallucinated links) are fed back to the LLM as warnings, allowing it to self-correct.
• Lenient Cycle Detection: Normal backtracking is permitted, but tight back-and-forth loops (like A ➔ B ➔ A ➔ B) are blocked. Stream repetitions are monitored and cut off if the model repeats sentences (3x), phrases (4x), or words (5x).
• Banned List Pages: Article titles matching List of ... are strictly banned. The engine automatically filters these from links, random generations, and manual inputs.
• Redirect Exemption: If a clicked link redirects to a target article (e.g. UK ➔ United Kingdom), the agent transitions to the target, but the step count is not incremented.
• Turn Timeout: Each turn is governed by a timeout. If the model doesn't finish within this window, it counts as a failure.
• Mistakes Tracker & History Benchmarking: Reasoning/formatting errors are tracked on the active page and logged. Previous runs are grouped in history, allowing interactive SVG path comparison overlays with zoom, pan, node times, and mistake overlays.