Skip to content

fix(consensus): disable unsafe LPIPS early stopping - #13

Merged
vex0209-bt merged 1 commit into
mainfrom
fix/disable-unsafe-early-stop
Jul 21, 2026
Merged

fix(consensus): disable unsafe LPIPS early stopping#13
vex0209-bt merged 1 commit into
mainfrom
fix/disable-unsafe-early-stop

Conversation

@vex0209-bt

Copy link
Copy Markdown
Contributor

Summary

  • add an explicit consensus-pinned early-stop enablement flag
  • ship early stopping disabled so every duel scores the complete corpus
  • expose the setting in validator dashboard metadata
  • add an adversarial regression for a legitimate late comeback

Why

The existing 20x delta-threshold heuristic is not a proven global upper bound on per-clip LPIPS advantage. A challenger can fall behind on early clips, exceed that assumed bound later, and be rejected even though the completed duel would crown it.

Consensus impact

The enablement flag participates in the consensus digest. Validators and eval servers must deploy the same chain configuration.

Testing

  • venv/bin/pytest -q: 615 passed
  • focused consensus, duel-runner, bootstrap, and dashboard tests: 80 passed
  • git diff --check

The existing 20x delta-threshold heuristic is not a proven upper bound on per-clip LPIPS advantage. A challenger can therefore fall behind early, exceed that assumed bound on later clips, and be rejected even though the complete duel would crown it.

Add an explicit consensus-pinned enablement flag that ships disabled, pass no bound to the evaluator while disabled, and expose the setting in dashboard metadata. Add an adversarial late-comeback regression and cover the flag's required-field and consensus-digest behavior.
@vex0209-bt
vex0209-bt merged commit ca25c09 into main Jul 21, 2026
3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant