Research
Research at Eschatia Labs
Our work in five areas. Each area names the question, the method, one measured result and what is public: preprints here, models on Hugging Face, code on GitHub.
Decision models
A decision model reads a situation, picks one of the given answers and gives each answer a probability. Deem-4B works in English, Karar-4B in Turkish and Wähler-4B in German. All three were trained from Intern-Decision-4B.On jevalt-bench, Deem-4B reaches 94.7% accuracy, versus 84.7% for Kev-4B.Papers, code, models and data
The three models were trained from Intern-Decision-4B, first on all three languages together, then on each language alone. The models, the training code and the test data are public, and the released files run on a laptop with about 3 GB of memory. Emberwick shows them at work: a village game in which they decide how the villagers react.
Publications
Models
Datasets
On test questions kept apart from training, Deem-4B reaches 94.7% accuracy, versus 84.7% for Kev-4B, an open decision model based on Qwen3.5-4B. Karar-4B reaches 96.8%, versus 87.1% for Kev-4B, and Wähler-4B reaches 92.0%, versus 81.1% for Kev-4B.
Open benchmarks
A benchmark is a fixed set of test questions with a scoring rule. jevalt-bench has 3,116 written situations in three languages. JevOss checks whether an answer changes when the question is rewritten a little. Tholos-Bench has 160 tasks in a shared workspace of tables, notes and a task board.3 suites · 3,116 situations · 160 workspace tasksPapers, code, models and data
Test questions are set aside before the final training run, so a score says something about new questions. Each benchmark publishes its data or code, its scoring rule, its results and its known limits. jevalt-bench and Tholos-Bench are datasets on Hugging Face; JevOss is a Python package on GitHub.
Open benchmarks
In 200 JevOss tests with an instruction hidden in the text, Deem-4B changed its answer in 14.0% of cases, versus 36.0% for Kev-4B. After about 600 words of unrelated text were added, Deem-4B lost 17.4 accuracy points, versus 5.4 for Kev-4B.
Explainable deepfake detection
A deepfake is a video in which AI has changed or replaced a face. xdfdet checks each video frame with EfficientNet-B4, an image recognition model. Grad-CAM heat maps show which parts of the face each decision relied on. AUC scores a detector from 0.5 (guessing) to 1 (perfect).On FaceForensics++, EfficientNet-B4 reaches 0.897 AUC with varied and partly blacked-out training images, versus 0.868 without them.Papers, code, models and data
We compared nine ways to train the detector, including augmentation (varying the training images) and cutout (blacking out parts of them). Eight trained models, the code, the UBMK 2026 paper and the MSc thesis are public.
Publications
On FaceForensics++, a public collection of real and altered face videos, the image model EfficientNet-B4 reaches 0.897 AUC when trained with varied and partly blacked-out images, versus 0.868 for EfficientNet-B4 trained without them.
Local task assistants
A local task assistant does work on your own computer under rules you set. Tholos assistants share tables, notes and a task board. Each action runs, waits for your approval or is blocked.On Tholos-Bench, on a Kaggle T4 graphics card, Tholos-2B passes 137 of 160 tasks, versus 112 for MiniCPM5-2B.Papers, code, models and data
Tholos-2B is the small model behind the assistants, trained from MiniCPM5-2B on recordings of completed tasks. The model, the benchmark and the training data are public. reevesagents, the tool we use to run coding assistants side by side, belongs here too.
Publications
Datasets
On the same graphics card (a Kaggle T4), Tholos-2B passes 137 of the 160 Tholos-Bench tasks, versus 112 for MiniCPM5-2B, the model it was trained from. On the same ordinary processor (CPU), Tholos-2B passes 134, versus 117 for MiniCPM5-2B.
Language preservation
Eldalambë helps people read and write Quenya and Sindarin, the Elvish languages Tolkien invented, and their script, Tengwar. A small language model checks its answers against a dictionary and grammar rules. The paper is public; the code, data and model files are private.Public preprint · code, data, models and test set privatePapers, code, models and data
The paper describes a small model that works with tools: a dictionary lookup, grammar rules and a Tengwar renderer. The paper is public; the code, data, model files and test set are private.
Publications
Projects
In the paper, the adopted version, Pro-Preserve, writes the expected sentence for 40 of 100 test requests, versus 21 for the earlier version, Pro2508, and identifies Tolkien’s words in 277 of 300 cases, versus 259 for Pro2508.
Publications
Eschatia Labs preprints, papers and technical reports
Augmentation and Cutout in Deepfake Detection: A Comparative Study of Accuracy, Calibration, and Attention
Compares nine ways to train a deepfake detector: how accurate it is, how well its confidence matches its results, and which parts of the face it relies on. Presented at UBMK 2026 in Istanbul and published in IEEE Xplore.
Language Preservation Using Agentic AI Architectures: A Tool-Grounded Small Language Model on a Constructed-Language Testbed
A small language model for Quenya and Sindarin, the Elvish languages Tolkien invented, with tools that check every answer against a dictionary and grammar rules.
JevAlt: Technical report
How Deem-4B, Karar-4B and Wähler-4B were trained and tested, with results, limits and download links.
Tholos-2B: Technical report
How Tholos-2B was trained from MiniCPM5-2B, and how it scores on the 160 Tholos-Bench tasks.
Explainable deepfake detection using frame level CNN models: A comparative study of augmentation and cutout techniques
The thesis behind xdfdet: EfficientNet-B4 detectors tested frame by frame on FaceForensics++, with Grad-CAM heat maps. TED University, 2025. DOI 10.5281/zenodo.18998566.
Approach
One model, one job, one defined test.

- IEach model is trained for one job and tested on that job.
- IITest questions stay out of the final training run.
- IIIResults, known failures and the files to repeat them are published together.
- IVThe released JevAlt files run on a laptop processor with about 3 GB of memory.




