Real prompts and replies from the pilot run before submission: open-weight language models as black-box traders and news readers, scored by CALIBER. Nothing here was staged; every step is a logged query and its reply.
| Model | C1 update slope (Bayesian = 1) | C8 desk / forum slope | Double auction, no check | With bid check | Losing orders | Sealed-bid auction vs random | Winner profit (equilibrium +) |
|---|
Efficiency is the share of the maximum trading surplus realized. Without the check every model traded every unit, including units that can only lose money, so efficiency sat at the trade-everything floor. The sealed-bid column scores the allocation above a random one (0 = random, 100 = efficient); winners' profit is per auction against the equilibrium bidders' result on the same draws.