| # Reproducibility: R3 |
|
|
| ## Public release, 8 September 2026 |
|
|
| Code, editable sources, manuscript, procedural images/masks, and exact score records are public at https://huggingface.co/datasets/Ethosoft/RefSeg-CA. Data tag `r3-data-v1` resolves to commit `d5ce7eaa159c8b0a1e81f439b651c0125c0b31e0`. An unauthenticated download into a fresh directory reproduced all 30 primary table rows and 22 sensitivity settings; maximum absolute error was 3.33e-15 in that Python runtime. See protocol/public_download_verification.json. |
|
|
| The public download script is `code/download_release_data.py`; add `--synthetic` to reconstruct all procedural image folders. The public natural index preserves IDs, while original natural-expression text and images are obtained from their upstream sources. The fuller author supplement below retains the original natural manifest. Neither package claims completed human validation. The browsable code host is https://huggingface.co/datasets/Ethosoft/RefSeg-CA/tree/main/source, under the Ethosoft organization. |
|
|
| ## No-cluster metric replay |
|
|
| Run `python code/verify_portable_results.py` from this package. It reads six gzip CSV tables under results/r3/portable and both compressed reviewer-replay traces. Thirty primary table rows reproduce within 2.22e-16; all 22 sensitivity means and harm counts pass. The recorded verification is results/r3/portable_verification.json. |
| |
| These are exact derived score records, not score-map caches. They suffice to independently recompute the main metrics but do not independently verify model inference pixels. Full pixel-level reruns require the upstream data/models and inference code below. |
| |
| ## Inputs and software |
| |
| - Model/data revisions and file hashes: protocol/asset_manifest.json. |
| - ReLA checkpoint identity and preparation: protocol/rela_assets.json, code/prepare_rela.py, code/prepare_d2.py, code/rela_engine.py. |
| - Executed package versions: protocol/runtime_versions.json. Hardware-specific Torch wheels must be installed appropriately for a new machine. |
| - All 49,492 natural records: results/r3/portable/natural_full_manifest.jsonl.gz. |
| - Complete rich query/scene manifests: data/rich_synthetic.jsonl and data/rich_scenes.jsonl. |
| - Historical preparation/inference source is bundled under legacy_project/code. Local paths resolve through code/project_paths.py; REFSEG_LEGACY_ROOT can explicitly point to another prepared copy. |
| |
| ## Reconstruct images and inference |
| |
| 1. Fetch pinned assets with `python legacy_project/code/prepare_assets.py` (checksums verified). |
| 2. Prepare the historical fit/calibration data with `python legacy_project/code/prepare_data.py` when reproducing that phase. |
| 3. Prepare full natural inputs and official masks with `python code/prepare_full_natural.py` and `python code/prepare_official_masks.py`. |
| 4. Generate all 1,200 rich scenes with `python code/generate_rich.py`. The shipped ten example scenes are a preview; the complete generation is deterministic from five pinned seeds. |
| 5. Use code/infer_r2.py for CLIPSeg/Grounded SAM or native ReLA. The Slurm examples preserve executed settings; adapt only account, partition, paths, and obsolete dependency job IDs for another cluster. Install the pinned models/dependencies before inference. |
| 6. Run code/r3_replay.py on caches for each of 16 shards per model; run code/r3_summarize.py after all shards complete. A default replay must exactly match the archived primary IoUs before any sensitivity result is accepted. |
| |
| ## Executed reviewer jobs |
| |
| - Array 6334005: 32 completed CPU tasks, two interfaces × 16 shards. |
| - Summary/export 6334038: completed. |
| - Primary accepted records: 1,557 per interface; maximum primary replay IoU discrepancy: zero. |
| - Four scopes and eleven operating points are post-review exploratory diagnostics; the R2 primary rule remains unchanged. |
| |
| The raw TRUBA archive path is retained only as provenance: /arf/scratch/egitimg16u1/mcp-workspace/refseg_ca_r2. Portable primary metrics do not need that path. |
| |
| ## Archive integrity |
| |
| - portable_primary_results.tar.gz SHA-256: 4dccb64342faa3c96fd50cce8f7bfed6237d196310ad4947b4d88a14c095708a |
| - review_replay_records.tar.gz SHA-256: 1df51ae12a48f74f8015b14f637dccd31c103971299a4b4b976e5cb72e4051de |
| |
| A release-wide SHA256SUMS file accompanies the final package. Public repository deposit is complete; a DOI remains unassigned. The public DATA_MANIFEST.json additionally records hashes for each distributed data shard and result archive. |
|
|
