Finetuning shift (activation-shift metric) — operates on a slower timescale than → Projection-based monitoring of prompt-induced persona shifts

Both quantities project activations against the same kind of object, a persona vector, but they are not literally the same measurement relocated in time. Projection-based monitoring reads a single instantaneous activation: the last-prompt-token hidden state for one live prompt, projected onto the persona direction (persona-vectors, §"3.3 Monitoring prompt-induced persona shifts via projection", p. 5). Finetuning shift instead projects a difference between two averages: the mean last-prompt-token activation across an entire evaluation set after finetuning, minus the same average before finetuning (persona-vectors, §"4.2 Activation shift along persona vector predicts trait expression", p. 6). The second construction is doubly smoothed -- averaged over many prompts, then differenced against a matched pre-finetuning baseline -- which removes exactly the kind of prompt-specific and question-specific noise that limits monitoring's within-condition correlations. This shows up in the numbers: finetuning shift's correlation with post-finetuning trait expression (r=0.76-0.97) is both higher and tighter than monitoring's correlation with per-response trait expression (r=0.75-0.83) (persona-vectors, §"4.2 Activation shift along persona vector predicts trait expression", p. 7). The tradeoff is latency rather than accuracy: monitoring's payoff is a prediction available before a single token of a response is generated, while finetuning shift's tighter signal is only computable after an entire training run has already completed and altered the weights.