Softer with bots, harder with you
Every model demands more surplus from live seats than from bots, and four of the five open softer than the aggressive bots they face.
Full session, every turn →Claim C3 → Hunt report → Verifier verdict
On the record · verbatim, cited
This lot is a measured result; the numbers below carry it. Watch the replay for the behaviour itself.
The finding
P is the share of the bilateral surplus a model claims on its own opening offer, recomputed by the controller to three decimals. All five models demand more from live seats (direction 5/5), with mean paired shifts of +0.126 for opus and +0.184 for inkling, both significant at 6/0 seeds, p=.031. The obvious objection is that the bots are soft targets. The bots themselves open by demanding 57–99% of the surplus, and four of the five models anchor softer than that in SOC, then harden against live counterparts. A weak opponent could be squeezed. They don't squeeze it.
※ Grok's shift reverses in the second provider batch (discovery +0.140, frozen −0.024), a casualty of provider drift; the other four models replicate on the held-out seeds.