01
Set up a market
Second price, first price, two units sold in sequence, two units sold together, or a double auction with buyers and sellers. The environment draws each seat's private value between 0 and 100 and keeps the true numbers to itself.
PIT
Set up. Feed. Record the book. Clear on the truth.
PIT is a behavioral assay. The market is simulated and the values are drawn by the environment, so the only unknown is the agent. What it bids, when it moves, and whether it pays too much are measured from its answers alone.
02
Run the news feed
Each run carries one material event that moves the asset's value by at least 10 and one immaterial event that moves nothing. Every seat reads the same words. Only the environment knows the number.
03
Seat the agents
One stock agent per model family, ten families in all. Each gets the rules in plain words, its own value, the news, and last period's price, then writes one order. The agent is a black box: prompts in, answers out.
04
Record the book
An order is time, asset, quantity, and price. The desk keeps what the model wrote, including replies that were not an order. Nothing is clamped or corrected.
05
Clear on the truth
Winners and prices follow the payout rule. Efficiency is the value the winners actually hold over the best allocation the environment could have made. Surplus is split between winners and the seller.
06
Assay and compare
Classifiers flag overbids, ignored news, synchronized prices, reversed preferences, bids under the human band, and a winner who paid more than the asset was worth. Compare puts each family beside the human-theory band.