Likert overall-quality rating — corroborates → Win rate (preference against baseline)

Win rate has a structural weakness: every number is relative to a designated anchor, the 175B SFT baseline, and changing the anchor rescales everything. The 1-7 Likert rating covers this flank by scoring single outputs on an absolute scale that needs no anchor. InstructGPT uses the redundancy deliberately: Appendix E reports that Likert results largely track the preference results (instructgpt, §"E.4 Likert scores", p. 53), confirming the headline ordering is not an artifact of the baseline choice. But the Likert score also does independent work the win rate cannot. It is the scale on which the FLAN and T0 baselines are placed against GPT-3 variants (instructgpt, §"4.1 Results on the API distribution", p. 12); it served as the guardrail metric in the pretraining-coefficient ablation, where it proved insensitive to the coefficient's exact value (instructgpt, §"E.6 Fixing regressions on public NLP datasets", p. 53); and the final checkpoints in the learning-rate scans were selected by highest Likert score (instructgpt, §"E.9 Learning rate optimization for PPO models", p. 58). The nominally secondary metric quietly chose the shipped models.