COLOPHON
How this site is built, and what runs underneath it.
Stack
- Framework
- Next.js 15.5 (App Router) on Node 20 LTS
- Language
- TypeScript 5.5, strict mode
- Styling
- Tailwind v4 (CSS-first @theme tokens)
- Motion
- Framer Motion 11, Lenis 1 smooth scroll
- 3D
- React Three Fiber 9, drei 10, three 0.165
- Content
- MDX (@next/mdx) with Zod-validated frontmatter
- Hosting
- Railway (Node, standalone)
- Analytics
- Plausible
Performance
Performance
PEND
Accessibility
100
Best Practices
96
SEO
100
Accessibility, Best Practices, and SEO scores are stable across builds. The Performance number is captured on launch day from the live Railway deploy and updated here.
Decisions
Edition 01
Tetrad: one shared material, four objects
Every object in The Set imports the same `meshPhysicalMaterial`. Roughness 0.78, no metalness, clearcoat 0.15. Variation comes from geometry and how silhouettes catch the rim light, never from per-object material tweaks. The single-material rule made the lighting decisions transferable: the live Tetrad and the v1.1 stills can share one HDRI-free studio lighting setup, and the muted accent on one Tetrad edge resolves a CSS variable into Three.js at runtime, so the brand-system color tokens remain the source of truth for both the DOM and the WebGL scene.
Edition 02
MCP Investment Mentors: protocol layer, not wrapper
The four mentor personas live on a Model Context Protocol server, not behind a custom adapter over OpenAI function calling. The protocol layer is the value. Any MCP-compliant client routes a question to all four personas in parallel, gathers structured refusals where they apply, and presents the disagreement side-by-side without further engineering. A function-calling wrapper would have re-implemented in a year exactly what MCP already standardises.
Edition 03
What If? Magazine: the pipeline is the container
Speculative journalism collapses into either prediction or science fiction when published without a frame. Here the frame is a three-phase pipeline that cannot be collapsed or skipped: research with every source verified, a McLuhan tetrad projected to 2036, then narrative drafted to a house style held in a reusable skill file. Two openly fictional bylines, Caelus Ward and Cy Skewhouse, carry the reporting and are declared as fictions, never as real people. The tetrad is the analytical engine of every piece, yet the magazine never prints McLuhan's name on the page. The concealment is deliberate: the framework should do the work without being invoked as an authority. It is the same tetrad that ships standalone as Edition 01.
Edition 04
Local Frontier: emulate the incumbents, do not invent
The local inference service exposes the OpenAI Chat Completions shape and the Anthropic Messages shape byte-for-byte, including the streaming chunk format and the tool-use event ordering. Any existing client targeting one of those two APIs can change its base URL and run against a 70B model on a laptop with no other code change. A new, cleaner local-first protocol was the wrong move: drop-in compatibility is the only thing that gets the install past the first hour.
Edition 05
Square Mile Platform: MDX in v1, Postgres in v1.2
The platform's content runtime ships in two phases. v1 is MDX in the repository, validated by Zod at build time. v1.2 is Postgres with Prisma behind the same Zod schemas. The order matters. Writing the schemas against MDX first forced every frontmatter field to be defined as content, not as a database column. When the loader contract swaps to Postgres, the components consume the same shape they always did, and no UI work has to be redone.