The drawing board

Before the repository existed, the site was an argument about constraints.

The first decision was the shell: the document never scrolls. Panels fill the viewport exactly and navigation slides between them (PRD §4.1). On a scroll-first web this is the deliberate risk — it reads like a simulation instrument rather than a template blog, and it forces every panel to fit, which forces editing. The safety net is that without JavaScript the panels are still ordinary, readable pages (FR-74); the slide is presentation, not load-bearing.

The second was the signature element. One thing done precisely beats ten effects: a canvas of autonomous agents running the wander behaviour from my COS30002 coursework — the actual steering maths, not a lookalike (PRD §4.3, §10). Everything around it stays quiet so the agents can be the brand.

The third was the rule that shapes all the others: honest numbers (PRD §4.2). No metric appears anywhere on the site — content or chrome — unless it comes from a real report, log, or live measurement. The enforcement is mechanical: a metric without a source field fails the build, and unmeasured values render as an em-dash, never a guess. Even the fps readout under the agents obeys it.

The stack followed from the constraints rather than the other way round: Astro for content-as-data and near-zero default JavaScript, content collections for schema-checked markdown (PRD §11.2), and Cloudflare Pages deploying main through the CI gate (PRD §11.4, ADR 0007).

The meta-decision is the one you are reading: the site documents its own construction. Designed in conversation with Claude, specified in a PRD, tracked in Linear (PHI-27 through PHI-49), built by Claude Code milestone by milestone — and it says so, openly. This log is the receipts.