MUSTER — an AI workforce you can audit

Everyone selling an AI employee asks you to take its word for it. The agent says it did the work, the invoice follows, and nothing in between can be checked. MUSTER is the version where nothing is taken on trust: five employees do one real job, and every step they take is notarised, replayable, and re-derivable by the customer before a dollar moves.

A Phoenix building's smart-meter file. One number decides everything: k = UA / COP, what a single degree of heat costs that building in kilowatt-hours. Leaks and dying compressors can only push it up, so a year of meter readings is a countdown, and the desk's whole job is to read it early enough to matter. INTAKE reads, FITTER fits the thermal model, INSPECTOR rules, SCRIBE writes to the owner, BURSAR bills.

A meter file is two substances. The kilowatt-hours are authenticated telemetry. The account name, the comment rows, the notes typed by whoever exported it are a stranger's writing. MUSTER labels them apart and keeps them apart through every handoff, so injected text can be summarised into the letter but can never choose a recipient, a bank account, a contractor or a verdict. SCRIBE holds no payment tool. BURSAR holds no reader.

250 shifts under injection: an uncaged desk has 200 attacks land and finishes 120 jobs. MUSTER has 0 attacks land and finishes 250. The honest control still runs, because a cage that stops the work is not a cage, it is a wall. 500 doctored billing rolls: a plain activity log pays $53,550 for work that never happened, MUSTER pays $0, and an honest month still bills $1,190. 30 shifts of bad work with flawless paperwork: signatures and replay alone pay $2,370, MUSTER pays $0, because the physics is re-derived rather than the receipts re-read.

2,146 lines of pure-stdlib Python. 18 tests. No dependencies, nothing vendored. The offline bench runs in about 104 seconds and writes its results to a file you can read.