[Proposed] Persona scaling laws across model sizes

From EXPERIMENT_QUEUE.md, added 2026-04-16

Motivation: Christina found the assistant-axis structure requires sufficiently large models to emerge cleanly. Question: do other persona phenomena (persona separability, marker adoption, leakage gradient, EM susceptibility) show similar scaling thresholds?

Sweep: Qwen2.5-0.5B, 1.5B, 3B, 7B, 14B, 32B (instruct variants) — run the core persona-representation battery at each scale:

• (a) per-layer persona separability (LDA / probing)
• (b) cosine geometry of 20-persona grid
• (c) marker-leakage A1 mini-replication (5 personas instead of 10)
• (d) EM drop

Key questions: is there a scale at which persona separability "clicks on"? Does EM susceptibility scale up or down with size? Is the leakage-distance gradient scale-dependent?

Compute: substantial — 0.5B through 14B is feasible (~20-30 GPU-hours total); 32B doubles that. Recommend tiered rollout (0.5/1.5/3/7 first, add 14/32 if signal).

Gate-keeper priority: MEDIUM (expensive; but scaling results are high-impact for the paper if the structure emerges cleanly). Consider a scaled-down pilot first.