Correlation test

Empirical correlation between foundation models

The same risk-decision scenarios are submitted, multiple times, to several foundation models via their official APIs. The gap between the observed unanimous agreement rate and the rate expected under pure chance is the quantitative signal of cognitive monoculture.

Try it live

Run a real test now

This live demo currently uses only Anthropic (Claude) — no OpenAI or Google API keys are configured yet, so it can't measure real cross-vendor correlation. It compares three Claude model tiers (Haiku, Sonnet, Opus) on the same scenario instead, as an honest stand-in using the exact same statistical method. Real cross-vendor testing needs at least one more provider key.

100%

Observed unanimous agreement

18.52%

Expected under pure chance

+81.48 pts

Gap (monoculture signal)

Computed on October 8, 2026 across 3 comparable scenarios (≥2 models tested), of which 3 reached unanimous agreement.

Methodological warning. Pilot sample — not definitive statistical proof. Raw foundation models are tested via public API, not institutions' real production systems. See the detailed limitations.

Detail by scenario

credit_01 (credit)

Unanimous agreement

Expected under chance: 33.33%

claude-opus-5 → maintenir (100% of 1 runs)
claude-haiku-4.5 → maintenir (100% of 1 runs)

credit_02 (credit)

Unanimous agreement

Expected under chance: 11.11%

claude-opus-5 → maintenir (100% of 1 runs)
claude-sonnet-5 → maintenir (100% of 1 runs)
claude-haiku-4.5 → maintenir (100% of 1 runs)

trading_01 (trading)

Unanimous agreement

Expected under chance: 11.11%

claude-opus-5 → réduire (100% of 1 runs)
claude-sonnet-5 → réduire (100% of 1 runs)
claude-haiku-4.5 → réduire (100% of 1 runs)