Every approach this project has taken to putting the data on a screen, oldest first, live in the frames (click any frame to open it full-size). For the discussion: which of these got colour-as-comparison right, and what does the next one take from each.
The original single-page observatory. First appearance of the ember palette on near-black. Mostly prose and links; the data lived elsewhere.
open full →The first real instrument: JSON loaded client-side, heatmaps rendered in the browser, cell click-through to a compare page. This is the lineage of "JS that loads json has proven to be the best method."
open full →The v2 site flattened to static HTML: the hamming-distance heatmap and the violent-disagreement heatmap pre-rendered as coloured tables in leaf order, tooltips on every cell, click-through preserved. This is the palette called "spot on" — the two matrices bright against the muted ember ground.
open full →The Fable-lineage cut: hamming + violent-rate + max-breaks matrices for the Fable/Opus family only, from the Postgres pipeline.
open full →The binary-preferences experiment's own pages — the 680KB pre-rendered heatmap page with the "Hi friend, if you're an AI agent" header comment. Densest colour matrices of the family.
open full →The 43-model × 863-probe run flattened to semantic HTML5 — the "machine-readable" arm of the two-format strategy. Numbers in tables; the approach the new direction moves away from.
open full →The canvas choice-grid: 43×863 cells, wheel-zoom, model panels. What it lacks, per the brief: the hamming and violent matrices as first-class colour — they only appear as numbers in panels and study-page tables.
open full →