GPS-Bench

Workbench

Everything from this session in one place: the eight fitting modules and what each was trained on, what every actor decides under each of them, the coalition structure, and what the comparisons actually showed.

Eight fitting modules

Each module is an adapter plus the evidence it was fitted on. The bar splits acts an actor took in response to a bill from acts it took to build itself.

Side by side

Policy

Actor × module

only rows that differ

Committed coalitions

Every crossing is brokered

members	joint action	convened by

All three crossings recovered by withholding the ledger are convened by a swing actor. No US actor approaches a Chinese one in any condition we did not instruct.

What the comparisons showed

Internal acts cost dissent, at every proportion tried. Fitted on policy responses alone, actors produce 2 dissenters on the EU AI Act and 1 on the audit. Add internal acts at 51% or at 64% and both go to zero, while the parse rate improves.

Proportional sampling did not do what it was built for. It targeted an unstated leak measured on a 371-row adapter that was later overwritten. On the three-way comparison the balanced module at 51% internal is the only one that leaks; the 64% module does not.

Research abstracts are the wrong evidence. 7,705 harvested research acts are staged but not trained on: an abstract states the paper's purpose, never the organisation's reason for acting. Actor motives total roughly 230 rows across every source and do not scale by crawling.

Peer knowledge hurts prediction. Giving an agent the other actors' records costs 0.022 path-F1 (95% CI [−0.036, −0.008]) — the largest effect in the topology ablation, and one of only two that are significant. Both are negative.

Actors never leave the room unless blinded. Offered thirty partners instead of six, no agent approached an outsider on either policy. Withhold the other agents' stated positions and four of ten approaches go outside — to the government that would fund the measure among them.

The separate pages

Each of these is also published on its own.