DOOM MCP vs Structured Output Benchmark

Chris Nager built a fully playable DOOM that runs as an MCP (Model Context Protocol) app, rendering inline inside Claude and ChatGPT without a separate browser tab. The architecture uses two MCP tools — create_doom_session for inline-capable hosts and get_doom_launch_url as a browser fallback — combined with cloudflare/doom-wasm for the game runtime and a signed token system that maintains session state across both surfaces. The result is the same session whether you're playing inline or in a tab. The key technical challenge was avoiding iframe and CSP (Content Security Policy) issues. Rather than embedding a browser page inside the MCP iframe, the DOOM canvas runs directly inside the host's iframe — a subtle but critical distinction that resolved a class of rendering and input-handling bugs. The final implementation is intentionally stripped down: no save/load, no persistence adapters, just stable playable DOOM. Beyond the novelty, this project is a concrete demonstration that MCP apps are interactive surfaces, not just tool-calling JSON endpoints. The progressive enhancement pattern — same signed-token foundation serving both inline and browser modes — is a reusable architecture for any game or interactive experience that wants to live inside an AI assistant. Nager open-sourced the implementation and the blog post is a detailed technical breakdown.

Interfaze's Structured Output Benchmark (SOB) exposes a gap that has been quietly breaking production AI pipelines: models can produce syntactically valid JSON while getting the actual values wrong. SOB measures value accuracy across 21 models using 5,000 text passages, 209 OCR documents, and 115 meeting transcripts — scoring each on seven metrics including value accuracy, faithfulness (grounding vs. hallucination), type safety, and perfect-response rate. The benchmark reveals some sobering findings. Even top models like GPT-5.4 and Claude Sonnet 4.6 achieve ~83% on text but drop to 67% on images and only 23.7% on audio. No single model dominates all modalities — GPT-5.4, GLM-4.7, Qwen3.5-35B, and Gemini 2.5 Flash cluster within one point of each other on text. Perfect response rates (all seven metrics correct) rarely exceed 50% for even the best performers. For developers building data extraction pipelines, agents that read invoices, or any system where "correct JSON" means more than syntactically valid JSON, this is required reading. The dataset is on Hugging Face, the paper is on arXiv, and the playground lets you test your own model's structured output capability directly.