Kynovar

Overview

Kynovar is a research system for studying dynamics learning and scientific reasoning in controlled simulated universes.

Problem

Studying whether models can infer and revise explanations of unfamiliar physical systems requires controllable worlds and observable evidence.

Architecture

1. 01

   Seeded simulator / controlled universe

2. 02

   Laboratory / observable trajectories

3. 03

   Dataset generation, splits, normalization

4. 04

   Linear, MLP, GRU, GNN model families

5. 05

   Rollout benchmarks / OOD evaluation

This flow follows the repository's simulator, laboratory, data, model factory/training, and evaluation modules.

What I Built

• Controlled synthetic universes and a laboratory interface for generating observations.
• Dataset generation, normalization, split, and OOD utilities.
• Trainable linear, MLP, GRU, GNN, and interaction-GNN dynamics models.
• Simulation, training, benchmark, and OOD scripts with recorded run metadata.

Technical Challenges

Model comparisons depend on stable data splits and uncontended, reproducible runs; the latest repository notes call out unstable regimes and pending acceptance reruns.

Stack

Validation / Testing

• A pytest suite covers simulator, data, models, evaluation, and reproducibility paths.
• The latest repository status says no milestone has passed its sequential, uncontended acceptance run; benchmark results remain preliminary.

Current Status

GitHub / Demo