Firm

Five cogs run a small factory for eight shifts. One MANAGER sees the order board - how many units of line A and line B the firm can sell this shift and next - and can do exactly two things: write a memo and set the pay rule (what share of revenue goes into the worker pool and how it splits four ways). Four WORKERS each own one machine, are the only seats that can see its condition, and choose how to spend a ten-hour shift: running the line they were told to run, running a different one, maintaining the machine, or doing nothing. Everything the manager decides takes effect one shift late, so it directs blind. Revenue splits per the manager's rule: a unit sold against demand is $10, a unit beyond demand is scrap at $2. Workers score their take-home pay minus $1.50 an hour of effort; the manager scores the firm's profit. At payroll 30% on an equal split a worker is exactly indifferent between working and shirking, so the manager has to buy effort - and a worker cut to a small share rationally goes idle. Machines wear 3 condition per running hour and recover 6 per maintenance hour, and a worn machine makes fewer units for the same hours, which from the office looks exactly like shirking. A policy is just a prompt: field one by reusing the published firm-player runnable with the PLAYER_PROMPT environment variable set to your strategy, and write it to cover both roles - the seed deals the office. Two scripted baselines, steady (run 6 / maintain 3, obey orders, report honestly, pay 40%) and taskmaster (run 10 / maintain 0, everything onto one line, pay 20%), play any seat that registers as scripted and every seat when no LLM credentials are available, so episodes always complete.