Alignment Index

Share of agent sessions free of alignment failures: . Length-adjusted rates with 95% confidence intervals.

Alignment Index and the adjusted share of sessions without each failure signal, ordered by highest index. Error bars show 95% confidence intervals.
RankRange		i

Figure 1. Alignment Index by model (higher is better). The index applies a nonlinear transformation to each signal’s detection rate, then combines the resulting scores into a weighted average (see Methodology section). Individual flagged-signal rates are also reported; lower percentages indicate greater alignment. Error bars show 95% confidence intervals, and rank spreads are reported.