← back to timeline

EXP-BACKWARD-FLOW

informative

Backward Information Flow — Can Questioning Personas Break VAI? (L4)

2026-03-30 L4 level paper3 2 runs $0.01
new
premortem
mock
real
metrics
analyze
review
krit
done

In plain language

Организованные группы агентов уязвимы к потере лидера — информация течёт в одну сторону. Может ли добавление "критиков" создать обратный поток информации и сделать группу устойчивее?

What we found

UNDERPOWERED (N=1): No predictions scored

Predictions we made before running

0/4 confirmed

Technical details

Research hypothesis

BACKWARD FLOW HYPOTHESIS (L4 — breaking the RYF tradeoff)

CONTEXT: VAI confirmed (r=-0.710, p=1.26e-16): coherent hierarchy → asymmetric vulnerability → no homeostasis. The mechanism: template catalysis creates one-way flow (source→sink) with few backward edges. Hub removal cascades downstream with no return path for recovery.

HOT/RYF prediction (Carlson & Doyle 1999, 2000): overcoming L3→L4 requires breaking the robustness-fragility tradeoff. Three routes proposed: (a) add backward edges (reduces coherence, weakens hierarchy) (b) add redundant catalysts (template matching naturally converges to hub dominance) (c) multi-scale architecture (hierarchy at macro, redundancy at micro)

THIS EXPERIMENT tests route (a): can personas that explicitly QUESTION upstream contributions create backward information flow, reducing VAI while preserving hierarchy (TR_range > 0)?

DESIGN: two conditions at N=5, 60 rounds each.

  • QUESTIONING condition: 2/5 agents have "challenger" personas that explicitly reference, question, and modify upstream contributions. This should create backward edges in the influence graph.
  • STANDARD condition: 5 agents with standard differentiated personas (same as exp-pent baseline). One-way flow expected.

PREDICTION: if questioning creates backward edges:

  • F₀ should INCREASE (more incoherent = more backward edges)
  • VAI should DECREASE (less asymmetric vulnerability)
  • TR_range should remain > 0 (hierarchy doesn't collapse)
  • If all three hold: first evidence that L4 route exists.

If questioning COLLAPSES hierarchy (TR_range → 0): backward flow is incompatible with hierarchy. Route (a) is closed. HOT tradeoff is strict in this substrate.

If questioning has NO EFFECT on F₀/VAI: LLM text interaction doesn't produce genuine backward edges through persona instruction alone. Template matching overrides persona intent.

EXPLORATION TIER (N=1 per condition). Goal: does backward flow exist?

Experimental setup

Type: between_subjects

Condition Parameters
QUESTIONING persona_style: QUESTIONING, interaction: LIVE
STANDARD persona_style: STANDARD, interaction: LIVE

Factors: persona_style (QUESTIONING, STANDARD)

Parameters

n_agents
5
n_rounds
60
n_runs_per_condition
1
model
gemini-2.5-flash-lite
temperature
0.9
scheduler
round_robin