submite_60

Private backup of the CoWork task directories, split into independent, uncompressed POSIX tar archives.

• 250 task directories, named CoWork-Tasks-ByteDance-<n> (n = 1..250), grouped ten-per-split under CoWork-Tasks-ByteDance-<lo>-<hi>/.
• 48 archives. Each holds six consecutive tasks, except submite_60_part_12_tasks_67-70.tar, submite_60_part_14_tasks_77-80.tar, submite_60_part_16_tasks_87-90.tar, submite_60_part_18_tasks_97-100.tar, submite_60_part_20_tasks_107-110.tar, submite_60_part_22_tasks_117-120.tar, submite_60_part_24_tasks_127-130.tar, submite_60_part_26_tasks_137-140.tar, submite_60_part_28_tasks_147-150.tar, submite_60_part_30_tasks_157-160.tar, submite_60_part_32_tasks_167-170.tar, submite_60_part_34_tasks_177-180.tar, submite_60_part_36_tasks_187-190.tar, submite_60_part_38_tasks_197-200.tar, submite_60_part_40_tasks_207-210.tar, submite_60_part_42_tasks_217-220.tar, submite_60_part_44_tasks_227-230.tar, submite_60_part_46_tasks_237-240.tar and submite_60_part_48_tasks_247-250.tar, which hold four.
• ROSTER.csv is available both at the repository root and inside submite_60_part_01_tasks_01-06.tar. Columns: split,dir,seq,orig_name,tier,F2P,P2P,gaps,size.

Reading the roster

F2P and P2P are taken from the task's own evidence/build/SINGLE_TASK_VALIDATION.json (native_summary.gold), falling back to evidence/build/NATIVE_VALIDATION.json (tasks[0].gold) for the few tasks whose roll-up file was written before native validation finished. Every task listed satisfies all three of: the empty baseline passes zero F2P tests, the gold patch passes every F2P test, and the gold patch passes every P2P test.

Tasks 111 and up additionally satisfy two contract-thickness floors that were introduced after the first hundred were authored: f2p_total >= 3 and f2p_total + p2p_total >= 12.

tier is derived from the data, not assigned by hand:

• A — P2P > 0. The task has regression coverage: there are tests that must keep passing, so a solution that breaks unrelated behaviour is caught.
• B — P2P = 0. Fail-to-pass coverage is real and measured, but no pass-to-pass set was frozen for this task, so regressions are not detected. Filling p2p_node_ids would require re-running node discovery, which is authoring work rather than a configuration change.

All 250 tasks in this snapshot are tier A.

gaps is the number of entries in that task's ASSEMBLY_RECORD.json gaps list — evidence the authoring pipeline could not produce, recorded rather than synthesized. A non-zero value is a provenance note, not a defect.

Container images are intentionally not shipped

Every images/*.tar.gz blob is excluded to keep the archives small. The Dockerfiles needed to rebuild them are present, so run the package's own scripts/00_build_images.sh. Each task ships two manifests:

• MANIFEST.sha256 — covers exactly the files present in the archive.
• MANIFEST.sha256.with-images — the original manifest, image rows included, kept for provenance. MANIFEST.sha256 carries its checksum.

scripts/40_verify_package.sh checks MANIFEST.sha256, so it passes as shipped. All 250 tasks were run through that gate on this snapshot and returned 结论: PASS. That gate covers package structure, manifest integrity, credential scanning, and the 0700 permissions on gold/, authoring/, and task/tests/; it does not re-run the tasks themselves.

Verify and extract

sha256sum -c SHA256SUMS

mkdir -p restored
for archive in submite_60_part_*.tar; do
  tar -xf "$archive" -C restored
done

The archives preserve numeric ownership, permissions, ACLs, extended attributes, symlinks, and sparse files where supported. Extract as root only if ownership and privileged metadata must be restored exactly — the confidential surfaces (gold/, authoring/, task/tests/) rely on mode 0700 and must never be mounted into a solver environment.