Does the cosine/JS leakage predictor's accuracy change with model size (Qwen-2.5 1.5B→32B)?

Goal

Run the #468 narrow→broad cosine/JS leakage-predictor recipe at ≥3 Qwen-2.5 model sizes (e.g. 1.5B, 7B, 32B) holding the recipe fixed, and report whether the predictor's correlation with measured leakage sharpens or degrades with scale.

Motivation

The entire predictor line (cosine/JS → leakage) is on Qwen-2.5-7B. #413 and #9 propose scaling load-bearing findings generally, but none tests whether the predictor itself sharpens or degrades with model size. If persona geometry cleans up with scale (the "assistant axis needs scale" observation), the predictor should get more accurate at larger sizes — and if it does not, that bounds how far the cheap-predictor program generalizes.

What exists to reuse

• The predictor scripts are model-size-agnostic (#404 / #458 / #468).
• #413 / #9 enumerate the load-bearing battery to port. Truthification already has a 7B → 32B scale point (open-q 6.3).

Design sketch (for /adversarial-planner)

Fix the cleanest predictor cell — the #468 narrow → emergent-misalignment recipe (in-context-example narrow persona, read at the newline-after-assistant token, L25). Run it at ≥ 3 Qwen-2.5 sizes (e.g. 1.5B, 7B, 32B): for each size, train the narrow-behavior datasets, measure the post-SFT broad rate, compute base-model cosine/JS, and report ρ vs size.

Hypothesis

The cosine → leakage ρ trends upward with model size (geometry sharpens with scale).

Caveats

• Heavy: 32B training is ft-70b-class pods. Consider running this as one explicit cell of #413 rather than standalone if compute is tight.
• Hold the recipe fixed across sizes — the single manipulated variable is model size.

Lineage / open questions

Advances q:leak-predictor (3.1) / q:beh-b-to-bprime (3.6) along the scaling axis. Parent #413; predictor recipe #468; scaling siblings #9 / #436.