Skip to content

XTRAIN-24: Prove HOMEGOLF score gain from XTRAIN or keep XTRAIN off the critical path #639

@AtlantisPleb

Description

@AtlantisPleb

Summary

Prove whether XTRAIN materially improves the HOMEGOLF or PGOLF score path or formally keep it off the critical path.

Related tracking issue: #633

Why

Current retained XTRAIN truth is real but still bounded:

  • XTRAIN completes train-to-infer closure
  • XTRAIN emits actual text
  • direct and served inference still agree
  • current retained XTRAIN BPB is still 4.202998425869111
  • there is still no retained evidence that XTRAIN improves the HOMEGOLF score path

The current risk is spending iteration time integrating XTRAIN more deeply into HOMEGOLF without any proof that it helps score.

Scope

  • define the exact experiment needed to attribute a HOMEGOLF score gain to XTRAIN
  • either produce that retained gain on the same competitive PGOLF family
  • or codify the “XTRAIN stays sidecar or non-gating” rule in docs and operator flow
  • keep the issue focused on score relevance, not generic XTRAIN correctness

Acceptance Criteria

  • either a retained HOMEGOLF score receipt improves because of XTRAIN involvement
  • or the repo explicitly marks XTRAIN as non-critical to the HOMEGOLF competitive loop until such evidence exists
  • the HOMEGOLF path docs and issue stack reflect the verdict clearly

References

  • docs/2026-03-28-psionic-clean-main-xtrain-run-audit.md
  • docs/2026-03-28-homegolf-xtrain-pgolf-tailrun-loop-and-xtrain-refresh-audit.md
  • docs/2026-03-28-parameter-golf-winner-gap-and-psionic-path-audit.md

Metadata

Metadata

Assignees

No one assigned

    Labels

    backendBackend workqaQuality and validation workroadmapRoadmap worktype:featureFeature request

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions