medimage-model-research

NC-track scaffold for a planned research brain-MR multimodal model. No weights, no training code, and no dataset manifests exist here yet.

Maintenance status (2026-09): passive. This repository is kept available as a reference implementation; CI runs on pushes and pull requests only, Dependabot security alerts remain enabled, and no scheduled jobs or hosted services consume ongoing resources. No active development is planned.

The NC-constrained sibling of medimage-model. The plan: same architecture and the same medimage-eval substrate, but ingesting CC BY-NC-SA training corpora (MR-RATE, CT-RATE). Weights eventually trained here would be research-only — commercial use requires separate contracts with the data providers (Forithmus and co-licensors).

Status: pre-v0.1 scaffold. See docs/LICENSE_MAP.md for the constraint chain and STATUS.md for the dated state.

Why a separate repo

Maintaining a clean license boundary. The commercial-OK track (medimage-model) must never see CC BY-NC-SA data. Putting MR-RATE-derived work here makes that boundary visible at the repo level, not just at the manifest level.

Implemented

• Two-gate MR-RATE access preflight (src/medimage_model_research/data/mr_rate.py) — refuses to proceed unless (1) a per-user NC acknowledgement is on file and its recorded terms hash matches the current terms, and (2) HF_TOKEN is present (checked length-only, never echoed). No download code exists yet; this is the gate the future loader must pass through.
• NC acknowledgement opt-in (scripts/nc_acknowledge.py, make nc-ack) — interactive "I AGREE" prompt that records a timestamped receipt with the SHA-256 of the terms text (canonical text lives in medimage_model_research.nc_terms, shared by the recorder and the enforcing gate).
• Append-only access log — JSONL writer for per-shard provenance records (repo id, revision, sha256, license), written under gitignored data/.
• Attestation skeleton (receipts/attestation.py) — canonical-JSON leaf hash plus a fail-closed nc_obligation flag: normalised license-text matching (license_implies_nc), an explicit commercial_use input, and "no license recorded" treated as NC. Downstream consumers must check this field.
• NC boundary CI check (scripts/nc_boundary_check.py) — refuses imaging/weight blobs in the tree.

Planned (not yet built)

• MR-RATE / CT-RATE dataset manifests and an actual gated fetch path (huggingface_hub, pinned revisions, per-shard hashes into the access log)
• Model, training (SFT + GRPO), serving, and eval wire-up to medimage-eval (declared as a pinned optional extra; not yet consumed by any code)
• Receipts ledger integration for attestations

Not for clinical use

Nothing here is a medical device. Any future model outputs are research artifacts and must not be used for diagnosis or treatment decisions.

License

Code: Apache License 2.0 (full text in LICENSE). Trained weights (future): research-only per upstream CC BY-NC-SA constraints — see NOTICE and docs/LICENSE_MAP.md.