Issue 01
Protocol Paper · 2026
The Mecellem Protocol: a new grammar of reasoning
How structured ontology turns scattered legal records into contestable, governed reasoning — the theory behind newmind's semantic layer.
Read the report →newmind AI · AI Report
The AI Report.
Research from inside the operating system.
A periodic research publication from newmind AI — field studies, protocol papers, and the structural ideas behind governed, agentic legal intelligence.
Current Issues
Issue 02
Annual Report · 2026
State of Agentic Legal AI — 2026
What changed when reasoning moved from chatbots to governed operating systems — adoption, architecture, and the year ahead.
Read the report →Issue 03
Domain Models · 2026
Mecellem Models: Turkish Models for the Legal Domain
ModernBERT encoders pre-trained from scratch on 112.7B Turkish-dominant tokens and continually adapted — reaching top-3 on the Turkish retrieval leaderboard at a fraction of the parameters.
Read the report →Issue 04
Embeddings · 2025
TurkEmbed: a Turkish Embedding Model for Inference & Similarity
A Turkish embedding model trained on native — not machine-translated — data with matryoshka representation learning, setting new marks on natural-language inference and semantic textual similarity.
Read the report →Issue 05
Retrieval · 2025
TurkEmbed4Retrieval: Turkish Embeddings Tuned for Retrieval
A retrieval-tuned TurkEmbed variant fine-tuned on MS-MARCO-TR that outperforms TurkColBERT on Scifact-TR by 19–26%, raising the bar for Turkish information retrieval.
Read the report →Issue 06
Benchmark · 2025
TurkColBERT: Benchmarking Late-Interaction Retrieval in Turkish
The first comprehensive benchmark of dense bi-encoders against late-interaction ColBERT-style retrievers for morphologically rich Turkish, across five BEIR-TR domains.
Read the report →Issue 07
Hallucination · 2025
Turk-LettuceDetect: Hallucination Detection for Turkish RAG
The first suite of hallucination-detection models for Turkish RAG, framing the problem as token-level classification across question answering, data-to-text, and summarization.
Read the report →Issue 08
RAG Systems · 2025
Guided Decoding and its Role in Retrieval-Augmented Generation
A systematic study of guided decoding — Outlines, XGrammar, and LM Format Enforcer — and how multi-turn prompting shapes structured, hallucination-resistant RAG output.
Read the report →Issue 09
Robustness · 2025
PARROT: a Sycophancy Robustness Benchmark for LLMs
A robustness benchmark that measures how far language models bend the truth under authority and social pressure, with an eight-state behavioral taxonomy evaluated across 22 models.
Read the report →Issue 10
Fine-Tuning · 2026
RDP LoRA: Geometry-Driven Layer Selection for Fine-Tuning
A training-free method that uses the Ramer–Douglas–Peucker algorithm to pick which layers to adapt — beating full LoRA fine-tuning while touching far fewer layers.
Read the report →Issue 11
Semantics · 2026
Beyond Cosine Similarity: Taming a 15M-Node Synonym Graph
A 15-million-node Turkish synonym graph with a three-way relation discriminator that finally separates synonyms from antonyms and tames semantic drift in large-scale clustering.
Read the report →Issue 12
Datasets · 2026
A Hybrid Protocol for Turkish Semantic Relations
A hybrid pipeline — FastText clustering, Gemini classification, and curated dictionaries — that builds an 843K-pair Turkish semantic-relations corpus, a 10× scale-up for about $65.
Read the report →