One Job, Four Accounts, Nothing Done Twice — a Cross-AI Shard Runner With a Reconciliation You Can Defend

When a batch job is too big for one AI subscription's weekly limit, the tempting move is to open more terminals — and the thing that breaks is not speed, it is bookkeeping: two workers quietly process the same item, write two different answers, and nothing in your system can tell you afterwards which one to trust. This project builds the opposite of more terminals. It gives every unit of work a name derived from its own content, makes ownership a pure function of that name and a roster of legitimately separate accounts across different AI vendors, gives each account its own drawer to write into, and finishes with a reconciliation that proves every unit was done exactly once before a single row is merged. It is deliberately not the two neighbouring projects: uc235 routes one task to the cheapest tier that can be trusted with it, and uc244 runs the same task twice on purpose to find the hole in your specification. This one splits one job into many different, non-overlapping pieces, runs them at the same time on separate accounts, and then proves the pieces add back up — a throughput and accounting mechanism, not a routing decision and not a correctness comparison.