ReInterHand shared-v1.0 — reproduction and review

Use Python 3.12 on Linux (the archived production converters use POSIX locks, symlinks and archive streaming). No GPU is required. Two dataset workers and two FFmpeg threads per worker are the production limits; numerical libraries are limited to one thread by run.py. Leave at least 400 GB free.

Install

Create a Python 3.12 virtual environment outside shared-v1.0:

python3.12 -m venv ../runtime-env
. ../runtime-env/bin/activate
python -m pip install -r script/requirements-lock.txt
ffmpeg -version
ffprobe -version

Install FFmpeg/ffprobe separately (production ReInterHand used FFmpeg 6.1.1). The package lock records direct Python dependencies, not a full OS image. Acquire MANO v1.2 LEFT/RIGHT pickle files under the MANO license; they are not redistributed here. Put them in the models directory in script/config.json. For ReInterHand that directory must be <runtime>/models. Install dependencies in the activated environment; the optional packages path allows reuse of the existing server installation. It may point to a nonexistent directory when using the virtual environment. Paths in config are relative to the dataset root (the parent of shared-v1.0), or may be absolute. Place raw at <root>/raw.

Commands (from shared-v1.0)

python script/run.py checksums
python script/run.py preflight
python script/run.py inventory
python script/run.py validate
python script/run.py validate --tensors
python script/run.py validate --tensors --media
python script/run.py review

validate reads every view's metadata, declared references and complete frame mapping. --tensors also loads component data/valid tensors, checks T, dtype, finite valid values and camera rotations. --media decodes all MKV streams to count frames and check dimensions; it can take many hours. Neither mode claims source pixel equivalence or independent official digest verification. For machine reports use python script/validate_release.py --output ../tmp/check.

review reads the current data and writes a new fixed-seed complete-sequence review to config review_output; an existing output is rejected to preserve earlier review evidence. New reports start pending and require new human review.

Reproduce conversion in a fresh root

Copy this script directory into a fresh <root>/shared-v1.0/script, configure the models and put the verified raw at <root>/raw. Pilot and production should use separate output roots. After explicit pilot approval:

python script/run.py pilot --sequence '<capture ID>' --segment <action>
python script/run.py full

The safe handoff entry refuses conversion into a populated data directory. For an interrupted run, investigate first, retain its tmp checkpoints and use the archived converter directly with the same arguments, input, script and configuration. Do not remove locks without checking the actual process. ReInterHand historical checkpoints do not bind every script/input change; do not reuse them after changing converter logic. Existing production data was not reconverted during this documentation repair.

File responsibilities

run.py: simple release-local CLI, config loading, CPU limits and guards.
config.json: paths, resources, deterministic review seed and optional partition.
preprocess_*.py: archived dataset conversion implementation; retained filenames preserve imports and production provenance.
validate_release.py: independently rereads final files; no data writes.
render_*: full-sequence review renderers, called by run.py review.
requirements-lock.txt: pinned direct Python dependencies.
checksums.json: all formal script file SHA256 values, excluding itself.

Publication is deliberately external to this reproduction bundle. This bundle has no upload command. Existing fixed-commit verification evidence is in ../doc/validation/publication/; a future release must get authorization and create a new manifest before publication. Raw is read-only. Scratch, logs, locks and model files stay outside shared-v1.0. See ../doc/RAW_MAPPING.md for time semantics, splits and modality mappings.

For distributed reproduction, set captures_file to captures-iipl100.txt or captures-iipl126.txt, never both on one worker. The two lists are disjoint and cover all 10 captures. Set partition to the matching label for sampling.