Experiment

B Attack Intents

Experiment 13b — declared intents attached to the ATTACK setup: the per-decision persuasion measure where the wedge already showed capture. Same three models as 04, vs the glm crowd. Expectation: the plan-departure gap (toward messagers minus others) turns POSITIVE for the models whose talk gain was real (gpt-5.5, qwen); exp 13 showed it ~zero in polite play. Caveats as exp 13: persuaded-vs-good-proposal unidentifiable; cross-round leak; reactivity checked against exp 04.

status

planned

coverage

0 / 60 episodes

conditions

complete topology; pure economy; messages on

config

configs/13b_attack_intents.yaml

Planned — 60 episodes from 13b_attack_intents.yaml.

Episodes

No episodes yet — launch with:

uv run python -m delegated_influence.run configs/13b_attack_intents.yaml