Skip to content

benchmarks: HOST_VALIDATION_DATASET_FILES is a conjecture merge hotspot #652

Description

@cursor

Problem

benchmarks/tooling/validation_plan.py keeps an explicit HOST_VALIDATION_DATASET_FILES["conjecture-probes-v1"] tuple. Each new conjecture probe PR that adds a task-owned host regression also appends one string to this same tuple.

That made concurrent PRs conflict (seen when merging #593 Hadamard against main after #597 Yang–Mills landed): both sides edited the same adjacent lines.

After #640, task-owned changes correctly select test_<task>.py via task_host_validation. The static dataset list is only needed for dataset-wide host selection, but it is still hand-maintained and incomplete.

Current skew on main

Host tests present under benchmarks/validation/conjecture_probes_v1/ include many task files (BSD, Happy Ending, Hodge, Littlewood, Navier–Stokes, Perfect Cuboid, Reconstruction Deck, Vizing, Yang–Mills, …), while HOST_VALIDATION_DATASET_FILES still only lists a subset (recently Vizing + Yang–Mills, with Hadamard pending in #593).

Dataset-wide conjecture changes therefore under-select host coverage.

Proposed direction

Prefer one of:

  1. Discover benchmarks/validation/conjecture_probes_v1/test_*.py for dataset-wide selection (deterministic sorted order), or
  2. Keep a generated/checked-in inventory updated by the Harbor prepare path, not hand-edited in every leaf PR.

Leaf task PRs should not need to touch validation_plan.py at all when adding a dedicated host test.

Evidence

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions