docs/RELEASE_STATUS.md

Full-release status

The release target is frozen to the task selector published by the AutoMedBench website at commit 1e6471746c94adfd53f552a7bc0b81386f5a5f78:

• 48 unique tasks across seven tracks;
• exactly two included tiers, lite and standard;
• 96 task-tier combinations;
• the segmentation-only pro UI option is excluded because its 14 result arrays are empty and the website disables it.

The exact ordered rows are in manifests/task_matrix.csv. Run python scripts/verify_matrix.py to verify the counts and invariants.

Source readiness

The locked source branch tips contain canonical task configuration directories for 33 of the 48 published tasks. Fifteen website tasks are not present in those refs:

• enhancement: 5;
• VQA: 4;
• report generation: 3;
• classification: 3.

The missing harnesses are:

• enhancement: lidc-idri-denoising-task, deeplesion-denoising-task, brats-t1c-sr-task, ixi-t1-sr-task, and nih-cxr-sr-task;
• VQA: vqa-kvasir-task, vqa-omnimedvqa-task, vqa-pmc-vqa-task, and vqa-mmmu-medical-task;
• report generation: iu-xray-report-task, pathology-caption-100-task, and pathology-caption-500-task;
• classification: crc-histology-cls-task, patchcamelyon-cls-task, and chest-xray-pneumonia-cls-task.

An exhaustive search covered all local refs, 131 reachable commits, 6,729 Git objects, path history, blob content, and unreachable objects. It found no recoverable complete harness for these tasks. Historical results preserve the three classification IDs, and an older report commit contains only incomplete PathCap staging code; neither is a runnable task definition.

The release now supplies those 15 definitions as release_owned_v1 reconstructions, not as purported canonical source. They are tied to website commit 1e6471746c94adfd53f552a7bc0b81386f5a5f78, recorded one-for-one in manifests/release_owned_tasks.csv, and keep their unresolved data splits, permissions, baselines, and metric assets explicit. The frozen matrix therefore contains 33 committed tasks and 15 release_owned tasks, with both Lite and Standard rows for all 48.

Every task now has a concrete record in manifests/dataset_resolution.csv: 23 task rows are staged_verified, 16 are acquisition_ready, and nine are externally_blocked. The staged rows occupy 22 unique local dataset directories because the PathCap 100/500 tasks share a pinned 500-case source with distinct ID lists. Local staging contains 28,916 files and about 10.09 GB of logical data. It includes pinned MSD, VQA, detection, MMMU, PMC-VQA, and authorized PathCap additions beyond the original seven Lite snapshots.

Staged trees pass regular-file, symlink, sensitive-filename, path, selection, and layout-hash checks. Restricted bytes remain local and Git-ignored. The matrix no longer contains not_found; acquisition-ready and externally blocked states remain visible rather than being replaced with synthetic data.

All 48 task definitions build reproducibly into 48 data-free task archives (1,451,034 bytes total). Each archive contains both Lite and Standard scaffolding, an internal member checksum manifest, and no dataset directory. A second clean build produced byte-identical archives and catalog. All 1,568 archive members pass exact-secret, token-pattern, host-path, path-safety, and internal-checksum verification. A separate runtime gate renders every package in both tiers and exactly matches all 96 matrix cells. The package manifest records zero missing definitions; readiness gates remain visible in the release-owned task configs.

Runtime and task-tier image architecture

The release has exactly eight logical shared runtime images, frozen in manifests/runtime_images.csv:

• one common agent runtime covering all 48 tasks and both tiers;
• seven track-specific evaluator runtimes covering 14, 7, 9, 5, 4, 5, and 4 tasks respectively.

The requested execution layer adds 96 thin task-tier images, one per frozen matrix row. The builder disables non-deterministic provenance attestations and normalizes context timestamps; two clean builds produce the same 96 image IDs. Every image passes a non-root self-test. A final CPU-only remote-API smoke run covers S1 and S2 for all 96 IDs, records stopped_before_stage=s3, and never requests a host GPU. Build and verification identities are frozen in manifests/cell_images.csv and manifests/cell_verification.csv.

All eight are data-free, built, self-tested, privately uploaded, and recorded as uploaded_verified. The combined 4,297,282,868-byte archive has SHA-256 b9b7bea8e0e3cb0a03673f00d5b98ba6dfcc3c8329f37f26cf440cec980d38e8. The archive audit covered 44 unique layers, 143,878 layer members, 88,216 regular files, and 13,749,492,925 bytes without an exact match to any of the four locally loaded sensitive values. The eight shared images remain the common base beneath the 96 small task-tier layers rather than being copied as separate full runtimes.

The private Hugging Face repository reports the same LFS size and SHA-256 at revision cbeb96a85325da6dbe039a5a0cf6053293aa7ba6. Authenticated metadata confirms private=true, and an anonymous API request remains denied. All 48 task archives were likewise verified after five commits of 10 + 10 + 10 + 10 + 8 files.

The seven task-tier archives contain all 96 image tags and total 26,464,155,973 bytes. Each remote LFS object matches the size and SHA-256 in manifests/cell_archives.csv; the independent post-publication audit passed at cell-artifact revision e23ec6423db3833b0725ba021f222e8277d19a03 with the repository still private and anonymous access denied.

Only after those checks, the exact local runtime archive and sidecar plus the eight automedbench-full/*:0.1.0 image tags were removed. No shared or global Docker prune was run. The local task packages are retained because their total compressed size is only 1,451,034 bytes and they support repeatable audits.

The shared archive passed tag inventory, OCI graph, blob digest, diff-ID, path, historical-layer, and exact-secret checks. Its two upstream bases are digest-pinned and their nine ordered diff IDs are locked; heuristic token/path checks apply to every release-owned layer. The same fail-closed checks apply to every task-tier track archive. Definition coverage is complete, but task readiness is separate: reconstructed tasks remain blocked where their manifests record unresolved data or metric assets, and every runnable task receives independently acquired data through a runtime mount rather than through an image layer.

Packaging rule

Task-tier exports align with AutoMedBench-Lite: seven per-track paths named docker/<track>/automedbench-<track>-docker-image-v0.1.0.tar.gz, with synthesis used for the local synthetic track. The archives contain 28 + 14 + 18 + 10 + 8 + 10 + 8 = 96 image tags across segmentation, enhancement, VQA, report, detection, classification, and synthesis. Each is retained locally until the private remote reports the same size and LFS SHA-256; only then are that track's archive, sidecar, and uniquely namespaced tags removed. Shared Docker state is never globally pruned. Exact remote identities are recorded in manifests/cell_archives.csv.