American Alchemy
Two things live here. A research instrument built over the complete back catalogue of a documentary show — every claim extracted, classed as fact, inference or speculation, scored against an auditable rubric and checked against primary sources. And the products behind it: real interfaces, lifted from the repositories that shipped them, running on their own data with identities removed.
Built for a specific application. The corpus is the show's canonical episodes, and every citation deep-links to the exact second of source video. Nothing on this page is a figure someone typed — the numbers above are read live from the file the pipeline generates.
What the catalogue says, and what it never said out loud. The site-signature that splits into two disjoint families, the Cydonia coordinates in a 1984 tasking document, the evidence ladder, and the document index with its honest retrieval rate — including the documents that were named but could never be found.
The interactive graph. Every entity, claim and relation, filterable by class and confidence, with the claim-class breakdown drawn as a ring on each node.
Briefings for guests the show has not booked yet: every point where each person already appears in the catalogue, the questions only a full-catalogue reader could ask, and the claims a fact-checker should pre-check.
Claims with their eight scoring components open to inspection, threads adjudicated against retrieved documents, and the distinction between a genuine session document and evidence about Mars.
Each case embeds the interface that actually shipped. Where the real screen holds other people's data it is rebuilt with the same layout and mechanics on anonymised records, and the page says so. Nothing here is a mockup of a product that never existed.
People explorer, the hackathon event report, impact timelines and company ROI — all lifted from the shipped product, on anonymised records.
The approvals workspace, the real weights on the page, the automation workflow rendered from its own export, and one production prompt annotated section by section.
The research dashboard, task creation and investor circles. Profile pages are labelled reconstructions on invented data rather than passed off as real.
The leaderboard and inspector, with the seed set and the full catalog counted separately rather than summed into one flattering number.