Skip to content

Latest commit

 

History

History
90 lines (72 loc) · 4.52 KB

public-runtime-packets.md

File metadata and controls

90 lines (72 loc) · 4.52 KB

Optional public runtime packets

Only tasks declaring environment/public-input-packet.json use this bridge. Ordinary image-contained inputs, licensed project trees and required empty projects retain their original paths and need no new declaration or wrapper.

Use CPython 3.12 or 3.14, Docker with Buildx, Git/Git LFS, the declared host compiler tools and the task's original resources. Native qualification needs the original eight CPUs and 16 GiB; a configuration or controller unit test is not native admission.

Prepare the original public data

For an input-free image whose original visible root was inputs/files:

{"schema":"optimbench-public-input-packet-v1","mounts":["inputs/files"]}

Mounts are workload-relative and bind read-only at the same original /app/workload/ path. Commit the declaration as protected image context and rebuild/rebind the image through the canonical recipe. Existing images carrying the old recipe label are not relabeled as the new recipe.

python3 -m judge.public_input_preflight --task TASK --fetch
python3 -m tools.public_runtime_packet --task TASK --image ORIGINAL_IMAGE \
  --cache /absolute/trusted-cache --out /absolute/new-agent-packet
cat /absolute/new-agent-packet/agent-command.txt

The preflight authenticates the existing public input commitments. A corrupt source cache fails instead of silently reacquiring a replacement. Acquisition archives are not implicitly runtime inputs.

For a protected original derivation, declare its original output byte counts and SHA-256 values under derived_inputs, plus the original digest-pinned preparation image, argv and exact output parent directories under prepare. Preparation uses no network, a read-only root and environment, and writable binds only for those declared output directories. Every derived output must authenticate before its runtime root is exposed. A task therefore exposes only its declared data roots and authenticated derived outputs, never raw acquisition inputs. The original protected recipe/helper/guard must remain unchanged.

The helper emits packet.json and exact read-only Docker mount arguments; use those mounts when you start the agent container (see agent-runner.md). For compatibility it also writes a local task wrapper for the legacy Harbor harness, which is no longer used. The wrapper changes only the image reference; original task resources, networking, artifacts, source contract and prompt are preserved. It adds no profiling or security privileges. The generated Harbor command explicitly disables its verifier and is agent-only, not a correctness, performance, cold-network or qualification claim.

Generated task inputs are contained, with an exact copy of the original prompt. The task's original tests, solution, data and other root files are not copied or linked into the wrapper. Its small contained verifier always exits 2, awards no reward, and directs users to the separate canonical judge if verification is accidentally enabled. Harbor's path-containment checks remain enabled. The generic agent command is still an unvalidated configuration template: passing task validation does not establish provider setup or agent startup.

Replay the task's existing source contract

Export only the task's admitted production source roots using its existing source contract. The canonical artifact descriptor supports paths relative to the descriptor, including wrapped multi-root maps. Scoring and both dedicated/direct verifier CLIs resolve these paths independently of the caller's working directory.

On the separate judging host, retain the original policy:

python3 -m judge.heldout_check --task tasks/TASK --image ORIGINAL_IMAGE \
  --cpuset ORIGINAL_EIGHT_CPU_SET --artifacts /absolute/submission.json \
  --output /absolute/heldout-result.json
python3 -m judge.score --task TASK --image ORIGINAL_IMAGE --separate-sandbox \
  --artifacts /absolute/submission.json --out /absolute/new-score

Candidate upstream rejection is incorrect and retains its rejection receipt. Pristine baseline qualification and infrastructure failures remain invalid; neither status proves that the corpus comparator ran. A failing upstream negative control is not promoted to a successful corpus control.

Host registration is separate from provisional numeric interpretation. An unregistered host is not officially claimable. No component test, prepared packet or correct-source no-op proves a resolved speedup or passing A/A quality.

This component does not publish or certify Git/source-export tooling or a clean-bootstrap proof.