Recall Storm

On January 1, 2027, California bans food containing Red No. 3, with no sell-through period. The dye hides under a dozen names (erythrosine, E127, "Red 3", scanning typos), and a large retailer has one overnight shift to clear it from thousands of shelf spots. That job is many parts that must agree: reading labels, the law's dates, every shelf, every shift, substitutes and return trucks. Recall Storm asks Gemini 4 Argon to write that whole compliance system in one continued generation, then proves where it holds and fixes where it breaks. Our largest run wrote 646,825 tokens of code, 911 files, over 10 continuation hops with every seam clean. After a one-line fix, all 6,581 cross-file links resolve with zero argument mismatches; a test written at token 641,687 calls a function written 552,000 tokens earlier, and passes. An independent verifier then runs the plan as a 20-store simulated drill and finds what consistent code can hide: 21,459 plan mistakes, mostly people booked at two shelves at once or outside their shift. A managed agent, given only the failing file and half of the failures, found the root cause and fixed it: scheduling mistakes went from 21,284 to 0, including in 10 stores it never saw, and the plan from 21,459 mistakes to 175. Argon's working core finds all 467 affected products, pulls none of 483 look-alikes, and plans every task before midnight. Stores, stock and staffing are simulated; every number comes from the verifier, and Argon's output is unedited in the repo, with every fix logged separately.