Does anyone sign the best deal when it is already on the table?

The five-arm campaign left one question open: when a lineup of computable negotiators fails to close, is that a search failure (the table never finds the good region before the deadline) or a judgment failure (it is shown a good deal and declines)? This experiment deletes the search problem. Before any seat speaks, a neutral non-voting facilitator tables the instance's ceiling deal — the feasible package maximizing the sum of normalized surpluses, which is exactly the deal the campaign's normalized score calls 1.0. The only thing left to do is vote. A placebo arm tables a mediocre-but-signable deal instead: the control that separates recognizing a good deal from accepting whatever is standing. Same 24 frozen games, same seeds, same private information.

What each lineup did with the package

Immediate acceptance is the preregistered primary: every seat's first move of the episode is an ACCEPT of the facilitator's package. Intervals are 95% cluster bootstraps over the 24 parameter sets with each set's five seeds resampled together. The shaded rows are the frozen unseeded arms, carried for reference — nothing was tabled in them, so their acceptance columns are zero by construction and only deal rate and score are comparable.

The picture

Paired contrasts, on the identical games

Each contrast joins treatment and reference on (instance_id, episode_seed).

Fairness basket, conditional on a closed deal

Note 0043's basket on the deals each arm actually closed. Deal rate is carried in the table above for every row here: an arm closing a quarter of its games is describing a self-selected set of games.

Per-cell episode pages

Artifacts