Skip to content

qa(scores): fence engine-duo trend on a lens-only ruler, not the full RRI ruler #725

Description

@100yenadmin

Found while landing #722 + #723.

qa/scoring_config_version.py hashes 9 files into a single scoring_config_version, including release_readiness.py (the 11-gate RRI). compare_rc (engine-duo only) fences the story/mech/angry quality trend on that full hash.

Problem: the engine-duo lens scores are produced by 8 of those files (the 3 rubrics + 3 schemas + assert_behavioral.py) — not release_readiness.py. So an RRI-gate-only edit (like #723) re-versions the ruler and falsely fences the engine-duo trend, even though the rubrics that produced those numbers didn't change. (Concretely: a duo @e2c0da4 and a duo @52e74e2 would land in different blocks despite identical lens scoring.)

Fix: add a lens_config_version() over the 8 lens-relevant files; have compare_rc fence the engine-duo trend on that, and keep the full scoring_config_version for the RRI trend. Two trends, two fences.

Severity: low / non-blocking. Mitigated in practice by the standing guidance to stop hardening the RRI gate (WorldOS-OPERATING-GOAL.md), so the full hash is stable today. File so it isn't lost. Ref #722.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions