EXP-QUAD2 revealed: integrative persona = hierarchy suppressor. N=4 with narrative/synthesis Delta → flat (0.059). N=4 with challenger Delta → hierarchy (0.509).
OPEN QUESTION: Can a SINGLE integrative persona destroy hierarchy in an N=5 system?
EXP-PENT baseline: N=5, all non-integrative → TR range 0.528 (robust hierarchy). This experiment: replace ONE agent (Epsilon gap-tracker → Epsilon synthesizer/weaver). All other agents identical. Same seeds, same parameters.
Three competing predictions: H1 (ECOSYSTEM ENGINEER): One synthesizer is sufficient to flatten hierarchy. TR range drops to <0.15 (approaching flat). Mechanism: synthesizer redistributes vocabulary bidirectionally, destroying the directed flow gradient. H2 (CRITICAL MASS): One synthesizer attenuates but doesn't destroy hierarchy. TR range 0.15-0.35 (reduced but present). Integration requires >1 agent or >20% of the population. H3 (RESILIENT HIERARCHY): One synthesizer has negligible effect. TR range >0.35 (close to baseline 0.528). N=5 topology + 4 non-integrative agents overwhelm single integrator's redistribution capacity.
Confound audit (per EXP-QUAD2 process improvement):
- Only ONE variable changed: Epsilon persona (gap-tracker → synthesizer)
- Same N, same seeds, same model, same other 4 personas
- Position confound: Epsilon is always position 4 (last in rotation). Last position may amplify synthesis effect (sees most accumulated text). This is noted but NOT controlled — first test is existence, not mechanism.