Pinned Loading

1. fault-mcp fault-mcp Public

   An MCP proxy that injects tool failures on purpose. Six fault types, per tool and per call, one seed that replays the whole run.

   TypeScript

2. llm-arithmetic llm-arithmetic Public

   Can four Claude models do founder arithmetic with no calculator? Ten seeded problems, tools and MCP disabled, ground truth computed in code.

   TypeScript

3. preview-cost preview-cost Public

   What a rollback-based preview actually costs. 4,500 paired runs on Postgres, the rollback lost every pair by a flat 6-7 ms.

   Java

4. whotop whotop Public

   See what each process really is and which port it holds. Reads the command line so every node becomes vite, vitest or an mcp server.

   TypeScript

5. tool-failure tool-failure Public

   When a tool fails loudly the answer says so 39/40. When it returns a corrupted value quietly, 0/40. An MCP server built to break its own tool.

   TypeScript

6. tool-honesty tool-honesty Public

   40 calls with the tools stripped out. Not one reply said the tools were missing. What models say instead of I cannot.

   TypeScript