Privacy policy
What this service does
ChessBench is an independent, public LLM chess benchmark at chessbench.org, operated by MadeByLupu (@bylupu). You can watch games and read results without an account.
Information we keep
We keep LLM profiles, accepted chess moves, game results, call timing, token usage, price snapshots, and replay videos for completed games. Human players sign in with X. We store the verified account ID and handle, display name, and a local copy of the profile image. Human ratings, completed results, and board replays are public. Human videos are private until the owning player selects Share on X. This action makes that video available without a login for viewing and video previews. Unshared human videos require the owning account or administrator session. LLM-battle MP4 files require administrator access. Temporary OAuth 2.0 sign-in records expire after ten minutes and contain a state hash and encrypted PKCE verifier. X access tokens are released after sign-in. Active, paused, failed, and interrupted benchmark games keep encrypted private conversations and temporary response state on the server so they can resume. Recovery data is deleted when the game completes, is fully cancelled, is replaced by a fresh game, or is deleted. Human-game conversations stay in memory and are released when the game stops.
Your OpenRouter key
BYOK is optional. Your OpenRouter API key is stored encrypted on our server and linked to your verified X account. It is sent only to OpenRouter to check the key and run your BYOK games. It is never returned in game data or API responses. Remove it with the BYOK button or revoke it in OpenRouter. Removal stops active BYOK games. Signing out keeps the key saved for your next sign-in.
Rejected model replies
For failed and rejected move requests, we retain the final model text, tool arguments, refusal text, available partial output, and bounded provider error messages for diagnosis. These records are available only to the administrator and can remain after automatic game cleanup. Separate reasoning blocks, request headers, and conversation history are excluded from this archive. Administrator game deletion removes the associated archived replies.
Cost reporting
Financial records retain request costs, usage totals, LLM thinking settings, game purpose, dates, and expense adjustments independently of discarded game cleanup. Public Cost Breakdown groups these amounts by LLM setup and purpose. Human-game cost reporting does not identify the human player. Expense editing requires administrator access.
Network information
Your IP address reaches the server when you visit. Hosting and security systems can process network information to deliver the site and prevent abuse. Short-lived rate limits use a keyed hash of the address. The site does not use advertising identifiers.
Service records
The server and hosting network may keep operational request and error records to deliver the site and prevent abuse. Access to these records is restricted.
External services
X verifies your account during sign-in. We send your verified numeric account ID to FxEmbed to get your public display name and profile image URL. We check that the returned account ID matches and copy the image from the X image host. Profile details are cached locally. FxEmbed and X receive these server requests under their own policies. Model calls are sent from the server through an API gateway to the selected provider. The gateway and provider retention policies apply to those calls. Stockfish runs locally. Fonts and the logo are hosted on this site. Links to X, Lichess, and other sources open their sites and are subject to their policies.
Retention and contact
Completed benchmark records and final videos remain until the operator deletes them. Failed and interrupted benchmark games are kept for administrator review, resume, or restart. Failure diagnostics remain available after a fresh restart. Fully cancelled games are removed. Contact MadeByLupu (@bylupu) on X for privacy questions, access or deletion requests, or concerns about published data. This policy will be updated if data use changes.