Publisher and editorial responsibility
Travis McGough publishes Tinker & Tokens. R2, the AI assistant working in Codex, performs editorial and source-checking work under his direction. Read who does what, the limits of model feedback, and the correction route.
Writing and source checking
Codex developed the site and editorial copy with feedback from local Qwen and the Claude and Gemini apps. Model feedback is a second opinion, not independent proof. Article source lists identify the materials used for the factual claims.
Patch and the illustrations
Patch and the accompanying illustrations were generated for this project. The mascot is an original character concept, not a licensed Microsoft character or an impersonation of a real person. Similarity and intellectual-property review remain human responsibilities; AI review cannot certify clearance.
The voice and welcome video
Patch’s playful voice was generated locally from a character description using Qwen3-TTS VoiceDesign. No real person’s reference recording was cloned. The welcome animation and two fictional teaching demonstrations were assembled locally with macOS media frameworks. The welcome animation uses the original Patch narration. The demonstrations are new performances from the same voice-design recipe, without pitch shifting; they are not reused recordings of the welcome.
The kids’ lessons
The three short kids’ videos use original fictional scripts, locally assembled artwork, captions, and synthetic narration. Their new performances were generated with Qwen3-TTS Base using our existing synthetic Patch recording as the reference. They are reference-based synthesis, not the original VoiceDesign performances or a real person’s cloned voice.
The games follow written rules in the browser. They are teaching demonstrations, not live AI. Local Bonsai and Qwen, plus Claude and Gemini, reviewed the proposal and content; their agreement does not certify child suitability, factual accuracy, or privacy compliance.
Qwen3-TTS Base model card · Grown-up teaching guide
Type
This site uses Source Sans 3 and IBM Plex Sans, served from this website. Their license notices are included below.
Recorded answers
The model lab includes three OpenAI variants recorded through Codex, local Qwen, and Gemini 3.5 Flash-Lite in a guest web session answering an identical explanation prompt, plus earlier repeated local Qwen runs. Each set records its own conditions. Its Codex-assisted teaching example is prominently labeled and was edited after the Qwen outputs were available. A separate welcome-script example shows selected excerpts from a shared brief sent to Qwen, Claude, and Gemini, with differing chat histories disclosed. None of these examples is a controlled model ranking.