How is any of this trusted? (the honest version)
The seeds and their held-out numbers were sealed in two commits (prediction registered before grading) in seed-library, whose CI re-derives the measurements on GitHub's own hardware. One domain (shopper) is REAL data — UCI Online Shoppers Purchasing Intention (Sakar & Kastro, 2018, CC BY 4.0). The other three are REALISTIC-SYNTHETIC: generated from a known rule with seeded noise — an honest stand-in, labelled as such everywhere, never passed off as real-world measurements. The factory's own laws (the scorer, the generated app, the receipt, the zip) are mutation-gated in this repo's CI; the germination you run here is the same gated ribosome, not a demo of one.