rules.md

Firm rules

Seats and roles

• Exactly five seats. RoleNames = ["Manager", "Worker"]: one Manager and four Workers, dealt by a seeded permutation of [0, 1, 1, 1, 1]. A policy cannot choose to be the manager, and no slot is structurally stuck on the floor.
• Machines are numbered 1 to 4 for humans and belong to the four worker seats in ascending seat order.
• Every seat plays under an anonymous cog alias drawn from the seed. Policy names are spectator-side only.

The floor

Two product lines, A and B. Each machine is set to one line at a time and starts at condition 100, with machines 1-3 set to A and machine 4 to B.

The order board is the manager's private information: demandA and demandB per shift. One level is high (30..36), the other low (12..18), and they swap once at a seeded switchShift in 3..5. The manager sees this shift's and next shift's numbers, and the whole realized history - never further ahead. A unit sold against demand pays Price = $10; a unit made beyond demand is scrap at SalvagePrice = $2.

A shift, in order

1. Open. The orders, payroll and split the manager announced last shift come into force; last shift's worker reports reach the manager; a shift event is recorded. At shift 0 the orders are each machine's own setup, payroll = 30, split = [25,25,25,25] and the standing directive is "Standing order: machines 1-3 on line A, machine 4 on line B. Six hours running, three on maintenance."
2. Deadline check, before the batch, never mid-shift.
3. Collect. All five seats decide simultaneously, in one parallel LLM batch. Nothing decided in shift s is visible to any other seat before shift s + 1.
4. The manager's memo is applied first: orders, payroll (0..60) and split are stored for shift s+1; the directive is stripped, newlines collapsed and truncated at 240 runes; an empty directive leaves the standing one unchanged.
5. The four workers are applied in machine order. run and maint are each 0..10 and together at most 10; the line is A or B; the report is capped at 120 runes.
6. Resolve, per machine, deterministically:
   • changing line costs ChangeoverHours = 2 of the running hours; hours = max(0, run - changeover);
   • q = 0.5 + 0.5 x condition/100, on the condition at the START of the shift;
   • units = floor(2.0 x hours x q);
   • condition = clamp(condition - 3 x hours + 6 x maint, 0, 100);
   • toil = 1.50 x (run + maint).
7. Sell. sold = min(produced, demand) per line; the rest is scrap. revenue = 10 x sold + 2 x scrap.
8. Pay. pool = revenue x payroll/100; pay[machine] = pool x share/100; profit = revenue - pool.
9. Score. workerNet += pay - toil; firmProfit += profit.
10. A settle event records the whole settlement, then the next shift opens.

The manager acts on every shift including the last; the last memo is recorded and rendered but governs nothing.

What each seat sees

• The manager sees the board (this shift and next, plus the realized history), the prices, every machine's number, operator alias, setup, order, units and pay, the share table, the ledger, last shift's reports, and its own notes. It never sees any machine's condition, any worker's hours or effort, or demand more than one shift ahead.
• A worker sees its own machine's condition, setup, order, hours, units, pay and effort, the standing directive, the pay rule in force including all four shares, the public shop floor (every machine's setup, order and units last shift), its own pay history, and its own notes. It never sees the order board, the prices, the firm's revenue or profit, another machine's condition, or another worker's report.

Replies

Manager: {"orders": ["B","B","A","A"], "payroll": 40, "split": [30,30,20,20], "directive": "...", "notes": "..."}. An unrecognised or missing order keeps that machine's previous order; a split that does not sum to 100 is renormalized by largest remainder; a missing or out-of-range payroll is invalid.

Worker: {"line": "B", "run": 7, "maint": 3, "report": "...", "notes": "..."}. An unrecognised or missing line keeps the machine's current setup; a missing or out-of-range run, or run + maint > 10, is invalid.

Invalid means one retry with a hint, then the steady scripted baseline. Every recorded string is truncated on RUNE boundaries.

Endings

• complete - all shifts resolved.
• deadline - the play deadline (60% of episodeTimeoutSeconds) was reached between shifts. Scores are normalized by the shifts actually played, so a short honest episode is on the same scale as a full one. There is no other ending.