scoring.md

Firm scoring

Computed once, when the episode settles. Let n be the shifts actually played. If n == 0 every score is 0.0. Higher is better everywhere.

• Worker seat: score = workerNet / (n x 30.0), where workerNet = sum over shifts of (pay - toil). A worker scores its take-home pay net of the effort it cost - negative if it toiled for less than nothing.
• Manager seat: score = firmProfit / (n x 300.0), where firmProfit = sum over shifts of (revenue - pool).

The league ranks by results.scores[seat] - that normalized per-shift net - and the ladder ranks seats by mean episode score. There is exactly one ladder statistic and both roles are on it.

Why the scales

Raw profit runs in the hundreds a shift and raw worker pay in the tens. Since the seed permutes roles, an un-normalized score would make the ladder a lottery over role draws. The scales are calibrated on a competently run firm: four machines at run 6 / maint 3 with fresh condition make 12 units each (48 total); against a board of A 33 / B 15 with three machines on A and one on B that is 45 sold and 3 scrapped, so revenue = 10 x 45 + 2 x 3 = $456. At payroll 40 the pool is $182.40, each worker is paid $45.60 against $13.50 of effort (net $32.10, score 1.07) and the firm keeps $273.60 (score 0.91). At payroll 30 the same floor gives the worker 0.69 and the manager 1.06. Both roles land near +1 when the firm is run well, and the payroll dial is a visible tug between them.

The marginal hour

One extra running hour on a healthy machine makes 2 units, worth $20 if they are sellable. Of that, payroll% goes to the pool and the worker keeps its share% of the pool: at payroll 30 and an equal 25% split that is 0.25 x 0.30 x 20 = $1.50 - exactly ToilPerHour. At the default pay rule a worker is precisely indifferent between working and shirking. The manager has to buy effort: at payroll 40 the same hour pays the worker $2.00 against $1.50 of toil and the manager still keeps $12. A worker cut to a 10% share earns $0.80 an hour at payroll 40 and rationally goes idle - that is the mutiny, and it is mechanical, not scripted.

Units made on the wrong line fetch salvage $2, so an hour on a mis-directed machine yields $4 of revenue and pays almost nobody: a bad directive destroys the incentive to work at all.

Maintenance, the hidden half

3 x run of wear against 6 x maint of repair makes run 6 / maint 3 the sustainable pace (net zero, one hour idle). Run 7 / maint 3 drifts down 3 a shift; run 10 / maint 0 burns 30 a shift and halves a machine's output in three shifts. Maintenance costs the worker effort now and pays back in future output, most of which the firm keeps - so an underpaid worker rationally lets the machine die. The manager sees the falling units and cannot tell that from shirking, because hours and condition are invisible from the office. That confusion is the benchmark.

What results report

names (policy names), scores, roles, pay (each worker's NET, 0.0 on the manager seat), units (0 on the manager seat), the firm's revenue, wages and profit, shifts, maxShifts and reason.