Skip to content

fix(config): preserve zero primary model sampling weights - #533

Merged
codelion merged 1 commit into
algorithmicsuperintelligence:mainfrom
justavibedev:fix/preserve-zero-primary-model-weight
Oct 22, 2026
Merged

codelion merged 1 commit into
algorithmicsuperintelligence:mainfrom
justavibedev:fix/preserve-zero-primary-model-weight

Conversation

@justavibedev

Copy link
Copy Markdown
Contributor

Legacy configuration loading and rebuild_models() treated primary_model_weight: 0 as a missing value and replaced it with 1.0. Preserve zero so weighted generation selects the positive-weight secondary model instead. Omitted weights still default to 1.0, and positive weights remain unchanged. Evaluator configurations that inherit these models also retain the explicit zero weight.

Regression tests exercise constructor, YAML loading and rebuilding through the real seeded ensemble with synthetic recording clients. All three zero-weight paths select the wrong client before the fix. Validation: 36 focused tests passed; full Python 3.10 unit suite passed (792 passed, 2 skipped); Black and git diff --check passed.

Model initialization, all-model judge generation and all-zero-weight validation are unchanged. Implemented and tested with Codex assistance and independently reviewed by another Codex agent.

@codelion
codelion merged commit 831f716 into algorithmicsuperintelligence:main Oct 22, 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.

2 participants