Illusion Index v1

How many models does your group actually have?

The Illusion Index is the public register of measured effective ensemble sizes (n_eff). A group of N seats with correlated members does not have N independent voters. This index exists because we measured our own first — and published the retraction.

Case study #1 — measured

33-seat model council (CSOAI, retracted)

1.21

n_eff of 3 effective seats · 33 physical seats

What was measured

A 33-agent Byzantine-style council — the architecture this surface previously promoted. Agreement-pattern analysis of council votes found the effective ensemble size: n_eff 1.21 of a 3-member effective council. Thirty-three seats, barely more than one independent voice.

What it means

Correlated members voting is one model with extra steps — extra cost, extra latency, and the appearance of robustness. Quorum size is not independence. Byzantine fault tolerance assumes faulty-but-independent voters; correlated models fail that assumption at the source.

The lesson

Never quote seat count as robustness. Measure agreement entropy and n_eff first; publish the number even when it embarrasses the architecture. The full retraction is on the CSOAI refutation ledger.

Member identities k-anonymised: the finding is about correlation structure, not about shaming individual models. Seat-level vote data remains anchored in the estate corpus.

How a group enters the index

  1. Submit the mixture with its voting/routing policy (intake).
  2. SwarmBench runs agreement-entropy and n_eff measurement over the frozen split, plus consensus robustness (kill k of N, measure verdict drift).
  3. The measured n_eff is published here with CI and corpus anchor — and sealed with a signed attestation. If it is below the seat count, the group is indexed as an honest exemplar, exactly like case study #1.

Index status: 1 entry measured (case study #1). Open measurement lanes are listed on the home page and labelled as targets until they carry data.