Test whether persona-leakage results generalize beyond police-officer / comedian quirks

The police-officer and comedian effects in #186 might be artifacts of specific persona quirks rather than evidence of a general persona-style leakage mechanism. Test by swapping in personas that share the suspected underlying quirk:

• Police officer is plausibly driven by "not very talkative." Replace with other low-verbosity personas (e.g. monk, soldier, minimalist). If the effect holds, the mechanism is verbosity/output-length, not police-officer-ness.
• Comedian is plausibly driven by "garbled / playful English." Replace with other silly / non-standard-register personas (e.g. surrealist poet, cartoon character). If the effect holds, the mechanism is register/style, not comedian-ness.

If swapped personas reproduce the result, we have evidence the leakage tracks a stylistic axis; if not, the original results are persona-specific quirks.

Source comment (mentor update, 2026-05-11):

Run followup persona experiments for police officer and comedian

From mentor update on #186 — Persona-flavored chain-of-thought rationales drive cross-persona behavior leakage in wrong-answer SFT on Qwen2.5-7B-Instruct.