Experience

User experience

The public site is laid out as a newspaper front page: a lead story with hero image, two card stories, a grid of shorter items, then two sections that cut across the archive — "Where & When", a Leaflet map teaser sampling four located articles (re-sampled hourly, so the front page rotates), and a "Picture Desk", a justified image gallery whose row layout is computed with a Knuth-style dynamic-programming line-breaker so every image keeps its exact aspect ratio. A dedicated Where/When? page plots every article on a world map above a zoomable canvas timeline of event dates. Articles are also browsable by month of event date. Article pages open with a full-bleed hero treatment when a featured image exists, and the whole site has a light and a dark theme.

The making side is the Wagtail admin plus the API documentation pages (Swagger and a Redoc-style view served from the OpenAPI schema). The daily workflow normally involves no interface at all: the generator runs, the article appears. The admin is where inspection and repair happen — reading the ai_context field to see what assignment produced a piece, fixing a caption, moving a map pin.

Current limitations and open questions

The generated images remain the weakest layer: current image models approximate period processes convincingly at a glance and unconvincingly on inspection, and no automated step verifies an image against the article it illustrates. The historical reasoning is only as good as the model's — the proportionality rule constrains the shape of a timeline but cannot fact-check it, and errors of period detail survive when the critic pass misses them. The event pool currently covers English and French Wikipedia only, so the articles alternate between two press cultures; the Dutch scraper configuration and style exemplars are in place but unused. And the anti-convergence machinery — banned motifs, recent-article digests, the mechanical scanner — is reactive by construction: each entry on the banned list is there because the tell had already appeared often enough to notice, which suggests the list will keep growing.