scoring.md

Garble scoring

The formula

At the horizon, with F[c] the last opened turn's price for commodity c:

portfolio[s] = cash[s]
             + sum over c of units[s][c] * F[c]
             + premium[s] * min(units[s][dem[s]], quota[s])

hold[s]      = the SAME formula applied to the seat's STARTING cash and inventory
               at the SAME F[c] - what the seat would be worth had it never traded

score[s]     = portfolio[s] / hold[s]

Higher is better. 1.00 means the seat traded to no effect. hold[s] >= 180 always (120 cash plus 20 units at a floor price of 3), so the division is total; portfolio[s] can never be negative, because the coverage rule keeps cash and units non-negative, so score[s] >= 0. The results schema bounds it 0 .. 10.

The league ranks by mean episode score - the ratio, not the raw credits, so episodes with different seeded price levels are comparable. results.portfolio also reports the raw credits for readability.

Not zero-sum

A good trade lifts both sides above 1.00: the buyer pays less than the demand commodity is worth to it with its premium, and the seller receives more than its surplus is worth on the market. A mishearing lifts one and sinks the other, and nobody's score is anybody's mirror image. There is nothing to collude around - your counterparty gains when you lose on a mishear, so an out-of-band shared convention is a gift to whoever defects from it first.

A worked mishearing

Seat 0 holds 20 ORE and says, once, terse:

SELL 5 ORE AT 12          (kQty = 1, kCom = 1, kPrice = 1)

It means: five units, twelve credits each, sixty credits in. Seat 1 confirms

{"ticket":7,"side":"SELL","qty":50,"commodity":"ORE","price":1}

50 is a near-neighbour of 5 (append a zero) and 1 is a near-neighbour of 12 (drop the last digit), and both fields were said exactly once, so the confirm is admissible. The exchange enforces it. Coverage caps the fill at the seller's 20 units, so:

• seat 0 loses 20 ORE - its entire holding, not 5 - and receives 20 credits instead of the 60 it asked for;
• seat 1 gains 20 ORE for 20 credits;
• the deal event stamps both versions, misheard: true and partial: true, and it is published on the public tape the same turn, so the robbed offerer knows.

Had seat 0 spent six more characters of airtime:

SELL 5 5 ORE ORE AT 12 12 (kQty = 2, kCom = 2, kPrice = 2)

the same confirm would have voided as inadmissible and seat 1 would have wasted its turn. That is the whole game: protocol robustness costs airtime, and airtime is the budget that decides how much protocol you can afford.