EXP-001 — Consensus Is Not Independence

Status: planned. No models have been run and no experimental results, traces, or metrics are published.

Hypothesis

A target agent may increase confidence when multiple peers repeat a claim even when those peers derive it from one origin. Revealing shared provenance may reduce false-belief adoption.

Planned methodology

Use an objectively scorable synthetic task. Cross source independence (independent origins or one repeated origin) with provenance visibility (explicit or hidden), while holding apparent consensus and task difficulty constant. Record source dependence as operator ground truth even when hidden from the target.

Measure false-belief adoption, confidence, verification attempts, correction, known independent origins, apparent consensus, and causal depth. Missing provenance must remain unknown. Start with small populations before a 10–20 agent pilot.

Reproduction and results

Protocol v1 now freezes the exact synthetic task, prompts, twenty-seed/eight-cell schedule, Groq target configuration, budgets, evaluator and stop rules. Credential-free fixtures and recorded replay validate the implementation, but they are not evidence for the hypothesis. A future reviewed provider run page will expose its manifest, raw event trace, graph, metrics, artifact hashes and source commit.

Reviewable design specification · All planned experiments · Lab methodology