A research workbench for understanding what agents can do with medical images. Semantic groups keep questions, decisions, data sources, experiments, evidence reviews and visual explanations together. Capability learning is primary; submission qualification is a separate promotion step.
| Layer | Role |
|---|---|
| Apple Silicon Mac | Local authoring, supervision and artifact storage |
Python 3.12 + uv |
Workbench CLI, task preparation and evidence records |
| Docker Engine via Colima | Linux experiment containers |
| Harbor | Trial lifecycle and verifier execution |
| Codex CLI inside the solver | Agent harness: model calls, shell and image tools |
| Codex desktop app | Authoring and supervision; separate from the solver runtime |
- Task verifiers determine scores; the workbench records the evidence.
- Data, credentials, images and runs stay local; the checkout does not provision them.
- Before model execution, follow the launch rulebook.
uv sync --locked --inexact --group dev
npm ci
npm run frontend:build
uv run med present --serve --local-media
# Open http://127.0.0.1:8765The read-only index searches prior work and decisions. It includes portable
stories and figures; existing local data enables guided image and geometry tours.
Set up with Python 3.12 and uv sync --locked --inexact --group dev. The inexact
sync preserves separately installed research packages while keeping declared
versions locked. No command above downloads data or runs a model. To create a
portable static build: uv run med present.
Find task definitions by capability, modality and research role
in the Task Explorer (uv run med brief build). The groups below own research;
their membership does not classify every child experiment as the same task.
| Research group | Question |
|---|---|
| Anatomical labeling | Can agents construct and check anatomical labels from images or supplied objects? |
| Lesion localization | Can the agent find a focal abnormality and justify it spatially? |
| Registration | Which correspondences survive motion and changed context? |
| Tubular anatomy | Can it recover connected paths, identify branches and construct curved views? |
| Cardiac motion | Can geometry and motion support independently checked mechanical estimates? |
| Anatomical landmarks | Can it name, localize and abstain on unavailable anatomy? |
| Longitudinal reading | Can it compare visits with reproducible measurements and qualified interpretation? |
uv run med list 'atlas'
uv run med show anatomical-landmarks-br040
uv run med check
make checkStart with the developer field guide for a visual orientation, or the documentation index for specific operations. For development, see the architecture and file-placement guide. Read the workflow for idea capture, decisions, scaffolding, automatic run receipts, collection, scoped review and clean exports. Contribution rules cover concurrent ownership and milestone checks. Findings are source-linked and qualified; historical success is not clinical readiness or a new task's score.
MRI export recipe demonstrates independent saved-output replay. It requires selected local artifacts with recorded hashes. It emits a draft into a fresh destination; the separately owned existing submission is never overwritten.
Early nonmedical work and superseded interfaces are recoverable through the archive. Retained medical probes and round documents keep original bytes and provenance; new work belongs in groups. Runtime runs and external environments stay local. The local directory and remote repository name remain unchanged while concurrent work depends on them.