1 oracle.pipelines.surface_correction_replay

oracle.pipelines.surface_correction_replay

Source-bound correction replay and portable observation readback.

The fixed eight-state panel pairs twelve candidates, two strides, two seeds, and two sample counts. Corrections change the point-to-triangle metric while holding prepared geometry and samples fixed; neither completion nor a historical sample-count gate establishes evaluator admission. The existing surface panel owns measurement and metric semantics.

Portable readback recomputes contrasts from retained JSON observations. An independently retained manifest SHA-256 anchors the bundle; archived paths and Git revisions remain provenance and are never accessed. Prepared tensor hashes are recorded provenance, not proof of a fresh tensor replay.

1.1 Functions

Name Description
compare_case Compare the same candidate treatments without treating v1 as ground truth.
freeze_assignment Freeze a clean implementation and every historical input before GPU work.
main Expose freeze, bounded replay, portable sealing, and cold readback.
read_result Verify a sealed relocated result using an independently retained hash.
run_assignment Replay exact frozen source in opaque subprocesses under an outer watchdog.
seal_result Copy retained JSON into a new exclusive portable bundle and return its seal.

1.1.1 compare_case

oracle.pipelines.surface_correction_replay.compare_case(old, new)

Compare the same candidate treatments without treating v1 as ground truth.

1.1.2 freeze_assignment

oracle.pipelines.surface_correction_replay.freeze_assignment(
    repo,
    output,
    old_assignment_path,
    old_run,
    deadline,
)

Freeze a clean implementation and every historical input before GPU work.

The historical v1 receipt is accepted only as a fully hash-bound source; it is never relabeled as a current v3 panel receipt. The deadline is an aware absolute time checked before and during execution.

1.1.3 main

oracle.pipelines.surface_correction_replay.main()

Expose freeze, bounded replay, portable sealing, and cold readback.

1.1.4 read_result

oracle.pipelines.surface_correction_replay.read_result(
    manifest_path,
    *,
    expected_manifest_sha256,
)

Verify a sealed relocated result using an independently retained hash.

Returns the historical receipt after verifying all retained observations and recomputing every reached case and the entire completed-panel aggregate. No historical stores, prepared tensors, Git checkout, or GPU are required. The seal proves consistency of captured observations; it cannot establish the correctness of the historical acquisition or reproduce CUDA numerics.

1.1.5 run_assignment

oracle.pipelines.surface_correction_replay.run_assignment(
    assignment_path,
    output_dir,
    *,
    case_limit=None,
)

Replay exact frozen source in opaque subprocesses under an outer watchdog.

Each case is limited by the lesser of its frozen 900 second bound and the remaining absolute deadline, reserving 15 seconds for the existing runner’s TERM/KILL cleanup. A blocked CUDA call cannot prevent the parent watchdog from terminating the child’s process group. A killed case retains failure evidence and unreached support; it never publishes a completed comparison.

1.1.6 seal_result

oracle.pipelines.surface_correction_replay.seal_result(
    assignment_path,
    output_dir,
    *,
    historical_assignment_path,
    historical_run,
    destination,
)

Copy retained JSON into a new exclusive portable bundle and return its seal.

Explicit historical locations permit relocated evidence without consulting absolute paths in archived JSON. All original bytes and source hashes are preserved. Keep the returned manifest hash independently of the bundle.

Back to top