Finetuning shift (activation-shift metric) — generalizes the single snapshot logic of → Base vs. RLHF-chat model representation comparison

CAA's base-vs-chat analysis was a single, retrospective snapshot: compare steering vectors independently extracted from the frozen Llama-2-base and Llama-2-Chat checkpoints by cosine similarity, layer by layer, to ask what one specific, already-completed RLHF run left alone in mid-network representations (contrastive-activation-addition, §"8.3 Comparing representations between base and chat models", p. 7). It answers a fixed historical question about one pair of models and cannot be rerun mid-training or pointed at an arbitrary finetuning job. Finetuning shift generalizes exactly this move: instead of comparing two independently derived vectors, it compares the same model's own last-prompt-token activations before and after any finetuning run, projected onto a persona direction rather than compared vector-to-vector (persona-vectors, §"4.2 Activation shift along persona vector predicts trait expression", p. 6). That reformulation turns a one-off comparison into a repeatable, general-purpose gauge: it can be computed for any checkpoint pair, at any point during or after training, against any of dozens of trait directions, yielding a single interpretable number rather than a layer-wise similarity curve requiring visual inspection. Read backward, base-chat-representation-comparison looks like the corpus's earliest instance of exactly the question finetuning shift is built to answer generally -- how much does this training process move a model's internal representation along a meaningful direction -- but 2023's version had no vocabulary for "a meaningful direction" beyond an entire steering vector, and no way to ask the question of any process besides RLHF specifically.