Elo Score (helpfulness/harmlessness) — generalizes → Win rate (preference against baseline)

Both metrics compress the same raw material — pairwise human preferences — but aggregate it differently, and the difference tracks what each paper needed. InstructGPT compares a handful of models against one fixed anchor, so a per-model win rate against the 175B SFT baseline suffices. Constitutional AI must compare 24 model snapshots spanning multiple training runs, with 10,274 helpfulness and 8,135 harmlessness comparisons collected across them (constitutional-ai, §"3.3 Main Results", p. 8), and no single anchor is natural when the objects of interest are whole training trajectories. Elo solves this by placing every snapshot on one shared interval scale, so harmlessness can be plotted as a curve against RL training sequences (constitutional-ai, §"4.3 Main Results", p. 12). The generalization has costs a win rate avoids: the zero point is arbitrary — the SL-CAI initial snapshot is simply pinned to zero — and the scale silently moves when elicitation changes. The paper attributes its unusually small helpful-versus-HH harmlessness gap to new instructions telling crowdworkers to prefer thoughtfully harmless over evasively harmless responses (constitutional-ai, §"4.4 Harmlessness vs. Evasiveness", p. 14).