The Adimension bridges
Human Ingenuity &
Machine Intelligence
Open code, published measurements. Behind every figure on this site: the evaluation data, the training records, and the check that could have falsified it.
ehjimp. The workstation, the literature platform, and its journal edition Hugging Face Two adapters,
two datasets, one Space 10.57967/hf/9541 · 10.57967/hf/9540 Source Eight public
repositories Apache-2.0 for the code · CC-BY-4.0 for the traces Kaggle The ATLAS
notebook 8,596 traces generated at runtime, before fine-tuning
The imaging work also has a technical thread on the Google AI Developers Forum. REDACTS was built against the INFINITERED campaign documented by Google Cloud threat intelligence.
The DEITY
Principles Framework
Five principles proposed in European Heart Journal — Imaging Methods and Practice. A letter argues a position; it does not measure one. Each line below is what that principle is made to cost in the codebase.
Built for the ground
that shifts
Clinical AI is usually built to one fixed dimension — one dataset, one regulator, one vendor, one release — and then the ground shifts underneath it.
The Adimension builds for the opposite case. Across imaging, literature, security and reasoning, one choice repeats: structured, inspectable, revisable output rather than a confident opaque answer.
A mask can only be accepted or discarded. Ordered coordinates can be dragged, measured against an expert tracing, and exported — the correction becomes the next training example.
The active
bridges
Different domains, one discipline. Each states its own scope, publishes its own numbers, and names the check that could have failed it.
Every repository behind these five, with its real metadata, is in Solutions. Four of them carry an open problem someone else could take.
One discipline,
four different domains
Imaging, literature, security and reasoning have little in common on the surface. Underneath, the same four practices carry all of them — and each is an executable assertion rather than a promise in a README.
What is open,
and what is gated
The clinical adapters are gated on Hugging Face: a research prototype should not slide into clinical use.
Everything else is open — the evaluation, the training records, the pipeline, and a review tier that runs with no weights and no GPU.
Every model here is a research prototype. None is a cleared or approved medical device, and every clinical output needs verification by a qualified specialist.
Scope, measured accuracy and known limits are stated in full on each project page.