96% of AI reasoning
is performative.

A cognitive measurement framework for language models. Ten dimensions, eighty-three environments, 35 models profiled, measured the way we measure people, not benchmarks.

Leaderboard

Section 01 · Leaderboard
Cognum v1.2 · top of the table
→ FULL TABLE
#01
Claude Sonnet 4.6anthropic
58.10
#02
Claude Sonnet 4anthropic
57.84
#03
Claude Opus 4.6anthropic
55.72
#04
Claude Opus 4.7 (STACK)anthropic
54.83
#05
mistralai/mistral-large-3-675b-instruct-2512mistral
54.02
Rank
Model
Cognum

Papers

Section 02 · Papers
Paper 1

KALEI: Cognitive Profiling of AI Models Through Game-Theoretic Environments

19 models, 10 labs, human baseline (n=14), Cognum v1.2 scoring methodology, Sonnet Surprise. 13 pages.

Videnov, V. (2026). KALEI Research. Preprint. ORCID: 0009-0008-4469-3327

DOI: 10.5281/zenodo.19698283
Paper 2

The Parliament Inside: Detecting Internal Argumentative Voices in AI Reasoning Models

96% performative reasoning, 6 voice archetypes, cross-lab comparison across 6 labs.

Videnov, V. (2026). KALEI Research. Preprint. ORCID: 0009-0008-4469-3327

DOI: 10.5281/zenodo.19698941
Paper 3

Search-Native Reasoning: How Perplexity Defends Its Architectural Identity

35.3% citation hallucination, 43.8% identity defense, 39.9% prompt injection framing. Architectural behavior study.

Videnov, V. (2026). KALEI Research. Preprint. ORCID: 0009-0008-4469-3327

DOI: 10.5281/zenodo.19699272

KALEI · LM Cognition Lab · Plovdivv1.2 · 35 models · live