docs/EXECUTION_PLAN.md

Full-release 96-cell execution plan

Objective

Publish and verify one Docker image for every frozen task-tier cell in manifests/task_matrix.csv: 48 tasks times lite and standard, for 96 images total.

The Lite release is the packaging reference: downloadable Docker archives are grouped by track. The Full release preserves its stricter isolation and data policy: task images contain code, prompts, contracts, and provenance only. Dataset bytes are mounted at runtime and are never embedded in an uploaded image.

Artifact layout

• Build one immutable task image per matrix row, tagged under the automedbench-full/ namespace and locked to one task and one tier.
• Retain the common agent runtime and seven track evaluator runtimes.
• Group exports into the same seven per-track paths and archive names used by AutoMedBench-Lite. Each archive contains all task-tier images for its track; the largest contains 28 images. The already-published runtime archive supplies the shared agent and seven evaluator images.
• Process upload archives serially. Keep each local archive until the private Hugging Face remote reports the same size and SHA-256, then remove only that verified archive and its uniquely namespaced task tags.
• Never run a global Docker prune or remove unrelated images, containers, volumes, or caches.

Dataset resolution

For every task, record one of these outcomes with source revision, license, selected IDs, layout checksum, and evidence:

1. staged_verified: authorized source data were deterministically staged and scanned.
2. acquisition_ready: the source and staging recipe are frozen, but an authorized user must supply gated data.
3. externally_blocked: access approval, a DUA, or upstream clarification is required and cannot be substituted by a private upload.

Restricted or unclear-license data may be used locally only when access is already authorized. They must not be uploaded. Synthetic inputs may verify a container contract, but never change a task's dataset readiness status.

Verification gates

Each of the 96 cells must pass:

1. task schema, internal checksums, path safety, and credential scan;
2. Docker build with a digest-pinned base and immutable task/tier labels;
3. non-root container self-test and task/tier identity check;
4. prompt and output-contract rendering against a staged or synthetic public item;
5. a bounded ChatGPT API workflow smoke run covering S1 and S2 only;
6. termination before S3 validation starts;
7. exported-archive OCI graph, layer digest, tag inventory, and secret scan;
8. private remote size and SHA-256 equality after upload.

S1-S2 smoke results are launch/readiness evidence only. They are not benchmark scores and must not be presented as task completion.

Execution order

1. Audit local/remote access, disk, Docker, GPU visibility, and credentials without printing sensitive values.
2. Resolve public datasets first; freeze acquisition recipes for gated tasks; update readiness records from evidence.
3. Implement and test the task-image builder, S1-S2 verification mode, archive exporter, and remote verifier.
4. Build and verify all 96 images.
5. Run S1-S2 smoke verification in bounded batches, stopping before S3.
6. Export and upload one track at a time, verifying the remote before cleanup.
7. Audit the final 96-image inventory and update release documentation.

Known environmental constraints at plan freeze

• LLM verification uses the configured remote ChatGPT-compatible API only; host GPUs are intentionally not used.
• Docker is available through passwordless sudo; the user is not in the docker group.
• The Docker daemon does not currently have NVIDIA Container Toolkit support. This is not a release blocker because the requested S1-S2 verification is CPU-safe and uses the remote API; verification launches omit --gpus.
• The configured ChatGPT-compatible endpoint and private Hugging Face token both authenticate successfully.
• Gated/DUA datasets remain external dependencies even when the build and upload pipeline is otherwise complete.