# Historical A/B Assignment Record Notice

## Authoritative record

The stored `answer_a_was` and `answer_b_was` fields in the 480 distributed verdict records identify the assignments actually used for the completed blind evaluations. Those stored assignments are authoritative.

## Bundled helper status

A post-run review found that the deterministic A/B assignment implementation preserved in the historical reference judge helper does not reproduce every stored assignment exactly. The helper remains available for audit, but it must not be used to recalculate, replace, or reinterpret the completed assignment record.

## Effect on the benchmark

This is a documentation and reproduction-package discrepancy only. No question, key, model answer, reasoning trace, judge decision, stored assignment, verdict, interruption treatment, score, or result was changed.

The original DeepSeek versus Claude record remains 8 wins, 3 losses, and 69 ties for DeepSeek, with 77 of 80 matters matched or beat. Release CV-LRB-2026.08.15-R10 adds Qwen 3.8 and 240 new stored pairings without altering those prior records. Qwen 3.8's expanded record is 30 wins, 4 losses, and 206 ties across 240 judgments, equal or better in 236 of 240.

Q59, Q61, and Q75 remain scored as delivered. The disclosed Series 1 Q12 rerun history is unchanged.

Release CV-LRB-2026.08.15-R11 changes presentation and reproducibility documentation only. The stored record and results remain unchanged.