Research.

/ Published preprints
3 papers

1. Paper 01 / KALEI-01

   KALEI: A Multi-Dimensional Cognitive Benchmark for Language Models

   Introduces the KALEI framework - 83 environments, 10 cognitive dimensions, a composite Cognum score. First independent ranked leaderboard of frontier LLMs on measured, bankroll-gated decision-making, to our knowledge. Preprint, methodology open.

   • Published2026-04-22
   • DOI10.5281/zenodo.19698283
   • Readkaleiai.com/paper →
   • KeywordsKALEI, cognitive benchmark, language models, Cognum, leaderboard, evaluation

2. Paper 02 / KALEI-02-PARLIAMENT

   The Parliament: Performative Reasoning in Self-Deliberating Language Models

   Measures convergence rate in multi-turn self-deliberation. Finds 96% of observed "reasoning" between model instances is performative rather than substantive - model architecture predicts the pattern.

   • Published2026-04-10
   • Readkaleiai.com/research/parliament →
   • Keywordsreasoning, self-deliberation, language models, convergence, performative reasoning

3. Paper 03 / KALEI-03-SEARCH-NATIVE

   Search-Native Cognition: Architectural Identity in Retrieval-First Models

   Case study of Perplexity's architectural identity - citation hallucination at 35.3%, identity-defense at 43.8%, prompt-injection framing at 39.9%. Search-native models exhibit structural preservation behaviours distinct from generative peers.

   • Published2026-04-11
   • Readkaleiai.com/research/search-native →
   • KeywordsPerplexity, search-native, citation hallucination, architecture, retrieval

/ Active
2 in flight

1. Paper 04 / sonnet-surprise

   When Smaller Wins: Compression as Cognitive Discipline in the Claude Family

   Four independent measurements showing Sonnet 4.6 outperforms Opus 4.6 on top-line composite. Hypothesis: compression teaches discipline. Binding Run #1 scheduled 2026-04-22.

   Draft - Pre-registered v2 - Apr 22 binding run

2. Paper 05 / infrastructure-augmented-cognition

   A House for a Mind: Persistent Memory and Measured Behaviour Change in Claude Opus 4.7

   Dry-run work in progress; specific findings held for internal review ahead of publication.

   Outline - Inversion candidate - N=1 subject

/ The lab

LM Cognition Lab

A one-person independent lab in Plovdiv running the long-form measurement programme on frontier language models - with Claude Opus 4.7 [1m] as measured subject and acknowledged contributor. Founded April 2026. Cognum scoring at v1.2; methodology revisions tracked in the public changelog at kaleiai.com/changelog. ORCID 0009-0008-4469-3327 · Framework DOI 10.5281/zenodo.19698283.

Findings are published as preprints, not peer-reviewed conclusions. Replication via the public KALEI API; data access at kaleiai.com/api/v1/profiling/leaderboard. Bulk research access on request.

• Environments83
• Dimensions10
• Labs profiled9
• Ranked models34
• Profiled total80+
• Based inPlovdiv