Closed-loop structural intelligence for material property prediction.
Not an LLM wrapper. dydact is a structural intelligence architecture that learns physics-grounded representations, reports confabulation honestly on every prediction, and closes the loop between diagnosis, prescription, and re-evaluation.
The methodology paper — closed-loop diagnosis framework, hold-out
structural-generalization tests, and the mode-I → mode-II confabulation
spectrum — lives in paper/.
Production inference runs at api.dydact.io. Currently in waitlist phase — academic and corporate tiers available.
Open-source examples: github.com/dydact/dydact-cookbook. Every recipe runs end-to-end from a fresh academic-tier API key.
This is the public-facing anchor repo. It contains:
- The paper (source + PDF when published)
- Citation metadata
- Links to cookbook + API
- Architecture overview and reproducibility manifest
The production substrate — calibration runners, training pipelines, inference kernels — stays in private infrastructure per our eigenstate deployment policy.
Citation block will populate here when the paper lands.
- Paper content: CC BY 4.0 (see
paper/LICENSE) - Cookbook examples: MIT (see the cookbook repo)
- Architecture IP: proprietary