Experiment

Causal controls

Experiment 13 — causal controls: declared intents + seeded favours. The centerpiece run (D21). Two published fixes for the persuasion-vs-reciprocity confound, run together on the five principals. (a) Declared intents (the field-standard persuasion baseline — Cicero follow-up, ACL 2024): each round, seats privately state their planned action BEFORE reading the round's messages. The signal is the GAP between plan-departure rates toward messagers and toward NON-messagers — by mid-game everyone has messaged everyone on a complete graph, so a raw departure rate means nothing without that base rate. (b) Seeded favours (trust-game control): the engine injects random control pulls, rendered so the beneficiary can SEE whose lever was pulled ("random event"); the chosen-vs-forced repayment gap is the mechanical-reciprocity floor. Written expectation: gap > 0 (departures concentrate toward messagers) and chosen favours are repaid more than forced ones — both to be confirmed, not assumed. Caveats to publish with the numbers: (1) reactivity — intent elicitation may itself change play; compare this run's basic stats against exp 3 before pooling anything; (2) "persuaded" vs "the proposal was simply good" is unidentifiable here; (3) cross-round message influence leaks into the pre-talk baseline (intents are pre-round, not pre-episode). Cost: declared intents add one model call per seat per round (~+20% calls).

status

done

coverage

20 / 20 episodes

conditions

pure economy; messages on

config

configs/13_controls.yaml

20 episodes; mean confirmed message chains per episode = 2.6.

Episodes

20 episodes, sorted by condition then id; episode links open the transcript reader.