# Dynamical Systems **Making Science Programmable** ## The experience layer for physical science Dynamical Systems builds infrastructure for existing materials labs to become agentic. We compile the historical record of a lab, the process records, instrument traces, calibration artifacts, source documents, and outcomes a campaign leaves behind, into source-located traces, deterministic verifiers, value signals, and replay environments so every experiment reduces uncertainty, improves the next decision, and compounds into better models, protocols, and materials. - Thesis: https://dynamicalsystems.ai/thesis - Research: https://dynamicalsystems.ai/#research ## Current Focus We turn repeated experimental workflows into auditable traces and replay environments. We start with one repeated materials workflow, passive trace capture, no instrument control, no replacement of existing lab systems, and lab-owned data. The first public proof surface is a replay benchmark that measures whether self-driving-lab agents can tell when the evidence in hand is enough to act on, before any model goes near a live instrument. ## Thesis The future is thousands of existing labs becoming programmable, verifiable, and compounding, thousands of labs learning from reality, and from themselves. Turning intent into execution is being built. Turning execution into experience is the missing layer. - URL: https://dynamicalsystems.ai/thesis ## Research ### Can a Self-Driving-Lab Agent Tell When the Evidence Is Enough? An environment compiler turns three NIST AM Bench workflows into 104 source-located replay tasks that sit on the evidence boundary, the point where the evidence in hand is enough to act on, defective enough to reject, or short of the one record that would change the call. Across six frontier models and 1,872 trajectories, agents reach a valid, evidence-grounded decision on 90 percent of runs and reproduce the reference-equivalent path on 72 percent. No model clears the benchmark, and where they fail they fail on judgment rather than formatting or retrieval. - Venue: Technical report - Published: June 19, 2026 - Updated: June 22, 2026 - URL: https://dynamicalsystems.ai/blog/benchmarking-self-driving-lab-agents ### Scaling Test-Time Verification for Novel Materials Crystal diffusion models encode property signals in their hidden states that they never use during sampling. Probe-gradient guidance steers unconditional generation toward target properties at test time, comparable to conditional models at over 50x the speed. - Venue: Blog - Date: April 2026 - URL: https://dynamicalsystems.ai/blog/scaling-test-time-verification ### The Missing Layer in Autonomous Science Multi-turn RL in verified campaign environments lifts hypothesis accuracy from 55.2% to 79.3% on held-out campaigns, surpassing GPT-5.4 with 3B active parameters. - Venue: Blog - Date: April 2026 - URL: https://dynamicalsystems.ai/blog/training-scientific-judgment ## Metadata - Canonical URL: https://dynamicalsystems.ai/ - Sitemap: https://dynamicalsystems.ai/sitemap.xml - LLMs text: https://dynamicalsystems.ai/llms.txt - LLMs full text: https://dynamicalsystems.ai/llms-full.txt - Agent skills: https://dynamicalsystems.ai/.well-known/agent-skills/index.json - Content policy (robots.txt Content-Signal): ai-train=yes, search=yes, ai-input=yes