- Keep the native method under
EXIST/3D/orEXIST/4D/, preserving its license and internal imports. Record its source revision and modification provenance. - Create an isolated environment profile. Do not add Torch/MMCV/model packages
to the lightweight
occstresspackage dependencies. - Add the formal dataset configs and entrypoints to
configs/evaluation_contract.json. State input offsets, modality, available tracks, control policy and required temporal-alignment switches. Direct-camera models must not be advertised as accepting manual occupancy inputs. - Extend the thin command adapter only when the native interface differs. Keep model internals and model registries in their own subprocess. Dataset input must obey the canonical H4/current/F6 contract and class/coordinate convention.
- Export six-horizon raw confusion matrices or sufficient counts, record the actual evaluated anchor IDs and all checkpoint components. Add a count reader only if a native format cannot already be read. Do not normalize rounded mIoU.
- Add CPU tests for config selection, input window, raw-count conversion and resume identity. Refresh the derived method inventory and reviewed source lock.
- In the native environment, pass checkpoint loading, clean/corruption smoke, original-loader versus clean-adapter equivalence and paper-reproduction gates. CPU tests are not substitutes for those gates. Publish the gate evidence with the reference result, including the precise score policy and anchor coverage.
See docs/CORE_WORKFLOW.md for output identity and docs/FORMAL_EVALUATION.md
for the authoritative contract. Adding a supported interface does not imply
that its entire dataset/source/protocol product has been evaluated.